Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+PixelGen-XL-16-256/demo.png filter=lfs diff=lfs merge=lfs -text
+PixelGen-XXL-16-512-t2i/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text

PixelGen-XL-16-256/model_index.json

@@ -0,0 +1,1017 @@
+{
+  "_class_name": [
+    "pipeline",
+    "PixelGenC2IPipeline"
+  ],
+  "_diffusers_version": "0.38.0",
+  "scheduler": [
+    "scheduling_pixelgen",
+    "PixelGenScheduler"
+  ],
+  "transformer": [
+    "transformer_jit",
+    "JiTTransformer2DModel"
+  ],
+  "id2label": {
+    "0": "tench, Tinca tinca",
+    "1": "goldfish, Carassius auratus",
+    "2": "great white shark, white shark, man-eater, man-eating shark, Carcharodon carcharias",
+    "3": "tiger shark, Galeocerdo cuvieri",
+    "4": "hammerhead, hammerhead shark",
+    "5": "electric ray, crampfish, numbfish, torpedo",
+    "6": "stingray",
+    "7": "cock",
+    "8": "hen",
+    "9": "ostrich, Struthio camelus",
+    "10": "brambling, Fringilla montifringilla",
+    "11": "goldfinch, Carduelis carduelis",
+    "12": "house finch, linnet, Carpodacus mexicanus",
+    "13": "junco, snowbird",
+    "14": "indigo bunting, indigo finch, indigo bird, Passerina cyanea",
+    "15": "robin, American robin, Turdus migratorius",
+    "16": "bulbul",
+    "17": "jay",
+    "18": "magpie",
+    "19": "chickadee",
+    "20": "water ouzel, dipper",
+    "21": "kite",
+    "22": "bald eagle, American eagle, Haliaeetus leucocephalus",
+    "23": "vulture",
+    "24": "great grey owl, great gray owl, Strix nebulosa",
+    "25": "European fire salamander, Salamandra salamandra",
+    "26": "common newt, Triturus vulgaris",
+    "27": "eft",
+    "28": "spotted salamander, Ambystoma maculatum",
+    "29": "axolotl, mud puppy, Ambystoma mexicanum",
+    "30": "bullfrog, Rana catesbeiana",
+    "31": "tree frog, tree-frog",
+    "32": "tailed frog, bell toad, ribbed toad, tailed toad, Ascaphus trui",
+    "33": "loggerhead, loggerhead turtle, Caretta caretta",
+    "34": "leatherback turtle, leatherback, leathery turtle, Dermochelys coriacea",
+    "35": "mud turtle",
+    "36": "terrapin",
+    "37": "box turtle, box tortoise",
+    "38": "banded gecko",
+    "39": "common iguana, iguana, Iguana iguana",
+    "40": "American chameleon, anole, Anolis carolinensis",
+    "41": "whiptail, whiptail lizard",
+    "42": "agama",
+    "43": "frilled lizard, Chlamydosaurus kingi",
+    "44": "alligator lizard",
+    "45": "Gila monster, Heloderma suspectum",
+    "46": "green lizard, Lacerta viridis",
+    "47": "African chameleon, Chamaeleo chamaeleon",
+    "48": "Komodo dragon, Komodo lizard, dragon lizard, giant lizard, Varanus komodoensis",
+    "49": "African crocodile, Nile crocodile, Crocodylus niloticus",
+    "50": "American alligator, Alligator mississipiensis",
+    "51": "triceratops",
+    "52": "thunder snake, worm snake, Carphophis amoenus",
+    "53": "ringneck snake, ring-necked snake, ring snake",
+    "54": "hognose snake, puff adder, sand viper",
+    "55": "green snake, grass snake",
+    "56": "king snake, kingsnake",
+    "57": "garter snake, grass snake",
+    "58": "water snake",
+    "59": "vine snake",
+    "60": "night snake, Hypsiglena torquata",
+    "61": "boa constrictor, Constrictor constrictor",
+    "62": "rock python, rock snake, Python sebae",
+    "63": "Indian cobra, Naja naja",
+    "64": "green mamba",
+    "65": "sea snake",
+    "66": "horned viper, cerastes, sand viper, horned asp, Cerastes cornutus",
+    "67": "diamondback, diamondback rattlesnake, Crotalus adamanteus",
+    "68": "sidewinder, horned rattlesnake, Crotalus cerastes",
+    "69": "trilobite",
+    "70": "harvestman, daddy longlegs, Phalangium opilio",
+    "71": "scorpion",
+    "72": "black and gold garden spider, Argiope aurantia",
+    "73": "barn spider, Araneus cavaticus",
+    "74": "garden spider, Aranea diademata",
+    "75": "black widow, Latrodectus mactans",
+    "76": "tarantula",
+    "77": "wolf spider, hunting spider",
+    "78": "tick",
+    "79": "centipede",
+    "80": "black grouse",
+    "81": "ptarmigan",
+    "82": "ruffed grouse, partridge, Bonasa umbellus",
+    "83": "prairie chicken, prairie grouse, prairie fowl",
+    "84": "peacock",
+    "85": "quail",
+    "86": "partridge",
+    "87": "African grey, African gray, Psittacus erithacus",
+    "88": "macaw",
+    "89": "sulphur-crested cockatoo, Kakatoe galerita, Cacatua galerita",
+    "90": "lorikeet",
+    "91": "coucal",
+    "92": "bee eater",
+    "93": "hornbill",
+    "94": "hummingbird",
+    "95": "jacamar",
+    "96": "toucan",
+    "97": "drake",
+    "98": "red-breasted merganser, Mergus serrator",
+    "99": "goose",
+    "100": "black swan, Cygnus atratus",
+    "101": "tusker",
+    "102": "echidna, spiny anteater, anteater",
+    "103": "platypus, duckbill, duckbilled platypus, duck-billed platypus, Ornithorhynchus anatinus",
+    "104": "wallaby, brush kangaroo",
+    "105": "koala, koala bear, kangaroo bear, native bear, Phascolarctos cinereus",
+    "106": "wombat",
+    "107": "jellyfish",
+    "108": "sea anemone, anemone",
+    "109": "brain coral",
+    "110": "flatworm, platyhelminth",
+    "111": "nematode, nematode worm, roundworm",
+    "112": "conch",
+    "113": "snail",
+    "114": "slug",
+    "115": "sea slug, nudibranch",
+    "116": "chiton, coat-of-mail shell, sea cradle, polyplacophore",
+    "117": "chambered nautilus, pearly nautilus, nautilus",
+    "118": "Dungeness crab, Cancer magister",
+    "119": "rock crab, Cancer irroratus",
+    "120": "fiddler crab",
+    "121": "king crab, Alaska crab, Alaskan king crab, Alaska king crab, Paralithodes camtschatica",
+    "122": "American lobster, Northern lobster, Maine lobster, Homarus americanus",
+    "123": "spiny lobster, langouste, rock lobster, crawfish, crayfish, sea crawfish",
+    "124": "crayfish, crawfish, crawdad, crawdaddy",
+    "125": "hermit crab",
+    "126": "isopod",
+    "127": "white stork, Ciconia ciconia",
+    "128": "black stork, Ciconia nigra",
+    "129": "spoonbill",
+    "130": "flamingo",
+    "131": "little blue heron, Egretta caerulea",
+    "132": "American egret, great white heron, Egretta albus",
+    "133": "bittern",
+    "134": "crane",
+    "135": "limpkin, Aramus pictus",
+    "136": "European gallinule, Porphyrio porphyrio",
+    "137": "American coot, marsh hen, mud hen, water hen, Fulica americana",
+    "138": "bustard",
+    "139": "ruddy turnstone, Arenaria interpres",
+    "140": "red-backed sandpiper, dunlin, Erolia alpina",
+    "141": "redshank, Tringa totanus",
+    "142": "dowitcher",
+    "143": "oystercatcher, oyster catcher",
+    "144": "pelican",
+    "145": "king penguin, Aptenodytes patagonica",
+    "146": "albatross, mollymawk",
+    "147": "grey whale, gray whale, devilfish, Eschrichtius gibbosus, Eschrichtius robustus",
+    "148": "killer whale, killer, orca, grampus, sea wolf, Orcinus orca",
+    "149": "dugong, Dugong dugon",
+    "150": "sea lion",
+    "151": "Chihuahua",
+    "152": "Japanese spaniel",
+    "153": "Maltese dog, Maltese terrier, Maltese",
+    "154": "Pekinese, Pekingese, Peke",
+    "155": "Shih-Tzu",
+    "156": "Blenheim spaniel",
+    "157": "papillon",
+    "158": "toy terrier",
+    "159": "Rhodesian ridgeback",
+    "160": "Afghan hound, Afghan",
+    "161": "basset, basset hound",
+    "162": "beagle",
+    "163": "bloodhound, sleuthhound",
+    "164": "bluetick",
+    "165": "black-and-tan coonhound",
+    "166": "Walker hound, Walker foxhound",
+    "167": "English foxhound",
+    "168": "redbone",
+    "169": "borzoi, Russian wolfhound",
+    "170": "Irish wolfhound",
+    "171": "Italian greyhound",
+    "172": "whippet",
+    "173": "Ibizan hound, Ibizan Podenco",
+    "174": "Norwegian elkhound, elkhound",
+    "175": "otterhound, otter hound",
+    "176": "Saluki, gazelle hound",
+    "177": "Scottish deerhound, deerhound",
+    "178": "Weimaraner",
+    "179": "Staffordshire bullterrier, Staffordshire bull terrier",
+    "180": "American Staffordshire terrier, Staffordshire terrier, American pit bull terrier, pit bull terrier",
+    "181": "Bedlington terrier",
+    "182": "Border terrier",
+    "183": "Kerry blue terrier",
+    "184": "Irish terrier",
+    "185": "Norfolk terrier",
+    "186": "Norwich terrier",
+    "187": "Yorkshire terrier",
+    "188": "wire-haired fox terrier",
+    "189": "Lakeland terrier",
+    "190": "Sealyham terrier, Sealyham",
+    "191": "Airedale, Airedale terrier",
+    "192": "cairn, cairn terrier",
+    "193": "Australian terrier",
+    "194": "Dandie Dinmont, Dandie Dinmont terrier",
+    "195": "Boston bull, Boston terrier",
+    "196": "miniature schnauzer",
+    "197": "giant schnauzer",
+    "198": "standard schnauzer",
+    "199": "Scotch terrier, Scottish terrier, Scottie",
+    "200": "Tibetan terrier, chrysanthemum dog",
+    "201": "silky terrier, Sydney silky",
+    "202": "soft-coated wheaten terrier",
+    "203": "West Highland white terrier",
+    "204": "Lhasa, Lhasa apso",
+    "205": "flat-coated retriever",
+    "206": "curly-coated retriever",
+    "207": "golden retriever",
+    "208": "Labrador retriever",
+    "209": "Chesapeake Bay retriever",
+    "210": "German short-haired pointer",
+    "211": "vizsla, Hungarian pointer",
+    "212": "English setter",
+    "213": "Irish setter, red setter",
+    "214": "Gordon setter",
+    "215": "Brittany spaniel",
+    "216": "clumber, clumber spaniel",
+    "217": "English springer, English springer spaniel",
+    "218": "Welsh springer spaniel",
+    "219": "cocker spaniel, English cocker spaniel, cocker",
+    "220": "Sussex spaniel",
+    "221": "Irish water spaniel",
+    "222": "kuvasz",
+    "223": "schipperke",
+    "224": "groenendael",
+    "225": "malinois",
+    "226": "briard",
+    "227": "kelpie",
+    "228": "komondor",
+    "229": "Old English sheepdog, bobtail",
+    "230": "Shetland sheepdog, Shetland sheep dog, Shetland",
+    "231": "collie",
+    "232": "Border collie",
+    "233": "Bouvier des Flandres, Bouviers des Flandres",
+    "234": "Rottweiler",
+    "235": "German shepherd, German shepherd dog, German police dog, alsatian",
+    "236": "Doberman, Doberman pinscher",
+    "237": "miniature pinscher",
+    "238": "Greater Swiss Mountain dog",
+    "239": "Bernese mountain dog",
+    "240": "Appenzeller",
+    "241": "EntleBucher",
+    "242": "boxer",
+    "243": "bull mastiff",
+    "244": "Tibetan mastiff",
+    "245": "French bulldog",
+    "246": "Great Dane",
+    "247": "Saint Bernard, St Bernard",
+    "248": "Eskimo dog, husky",
+    "249": "malamute, malemute, Alaskan malamute",
+    "250": "Siberian husky",
+    "251": "dalmatian, coach dog, carriage dog",
+    "252": "affenpinscher, monkey pinscher, monkey dog",
+    "253": "basenji",
+    "254": "pug, pug-dog",
+    "255": "Leonberg",
+    "256": "Newfoundland, Newfoundland dog",
+    "257": "Great Pyrenees",
+    "258": "Samoyed, Samoyede",
+    "259": "Pomeranian",
+    "260": "chow, chow chow",
+    "261": "keeshond",
+    "262": "Brabancon griffon",
+    "263": "Pembroke, Pembroke Welsh corgi",
+    "264": "Cardigan, Cardigan Welsh corgi",
+    "265": "toy poodle",
+    "266": "miniature poodle",
+    "267": "standard poodle",
+    "268": "Mexican hairless",
+    "269": "timber wolf, grey wolf, gray wolf, Canis lupus",
+    "270": "white wolf, Arctic wolf, Canis lupus tundrarum",
+    "271": "red wolf, maned wolf, Canis rufus, Canis niger",
+    "272": "coyote, prairie wolf, brush wolf, Canis latrans",
+    "273": "dingo, warrigal, warragal, Canis dingo",
+    "274": "dhole, Cuon alpinus",
+    "275": "African hunting dog, hyena dog, Cape hunting dog, Lycaon pictus",
+    "276": "hyena, hyaena",
+    "277": "red fox, Vulpes vulpes",
+    "278": "kit fox, Vulpes macrotis",
+    "279": "Arctic fox, white fox, Alopex lagopus",
+    "280": "grey fox, gray fox, Urocyon cinereoargenteus",
+    "281": "tabby, tabby cat",
+    "282": "tiger cat",
+    "283": "Persian cat",
+    "284": "Siamese cat, Siamese",
+    "285": "Egyptian cat",
+    "286": "cougar, puma, catamount, mountain lion, painter, panther, Felis concolor",
+    "287": "lynx, catamount",
+    "288": "leopard, Panthera pardus",
+    "289": "snow leopard, ounce, Panthera uncia",
+    "290": "jaguar, panther, Panthera onca, Felis onca",
+    "291": "lion, king of beasts, Panthera leo",
+    "292": "tiger, Panthera tigris",
+    "293": "cheetah, chetah, Acinonyx jubatus",
+    "294": "brown bear, bruin, Ursus arctos",
+    "295": "American black bear, black bear, Ursus americanus, Euarctos americanus",
+    "296": "ice bear, polar bear, Ursus Maritimus, Thalarctos maritimus",
+    "297": "sloth bear, Melursus ursinus, Ursus ursinus",
+    "298": "mongoose",
+    "299": "meerkat, mierkat",
+    "300": "tiger beetle",
+    "301": "ladybug, ladybeetle, lady beetle, ladybird, ladybird beetle",
+    "302": "ground beetle, carabid beetle",
+    "303": "long-horned beetle, longicorn, longicorn beetle",
+    "304": "leaf beetle, chrysomelid",
+    "305": "dung beetle",
+    "306": "rhinoceros beetle",
+    "307": "weevil",
+    "308": "fly",
+    "309": "bee",
+    "310": "ant, emmet, pismire",
+    "311": "grasshopper, hopper",
+    "312": "cricket",
+    "313": "walking stick, walkingstick, stick insect",
+    "314": "cockroach, roach",
+    "315": "mantis, mantid",
+    "316": "cicada, cicala",
+    "317": "leafhopper",
+    "318": "lacewing, lacewing fly",
+    "319": "dragonfly, darning needle, devils darning needle, sewing needle, snake feeder, snake doctor, mosquito hawk, skeeter hawk",
+    "320": "damselfly",
+    "321": "admiral",
+    "322": "ringlet, ringlet butterfly",
+    "323": "monarch, monarch butterfly, milkweed butterfly, Danaus plexippus",
+    "324": "cabbage butterfly",
+    "325": "sulphur butterfly, sulfur butterfly",
+    "326": "lycaenid, lycaenid butterfly",
+    "327": "starfish, sea star",
+    "328": "sea urchin",
+    "329": "sea cucumber, holothurian",
+    "330": "wood rabbit, cottontail, cottontail rabbit",
+    "331": "hare",
+    "332": "Angora, Angora rabbit",
+    "333": "hamster",
+    "334": "porcupine, hedgehog",
+    "335": "fox squirrel, eastern fox squirrel, Sciurus niger",
+    "336": "marmot",
+    "337": "beaver",
+    "338": "guinea pig, Cavia cobaya",
+    "339": "sorrel",
+    "340": "zebra",
+    "341": "hog, pig, grunter, squealer, Sus scrofa",
+    "342": "wild boar, boar, Sus scrofa",
+    "343": "warthog",
+    "344": "hippopotamus, hippo, river horse, Hippopotamus amphibius",
+    "345": "ox",
+    "346": "water buffalo, water ox, Asiatic buffalo, Bubalus bubalis",
+    "347": "bison",
+    "348": "ram, tup",
+    "349": "bighorn, bighorn sheep, cimarron, Rocky Mountain bighorn, Rocky Mountain sheep, Ovis canadensis",
+    "350": "ibex, Capra ibex",
+    "351": "hartebeest",
+    "352": "impala, Aepyceros melampus",
+    "353": "gazelle",
+    "354": "Arabian camel, dromedary, Camelus dromedarius",
+    "355": "llama",
+    "356": "weasel",
+    "357": "mink",
+    "358": "polecat, fitch, foulmart, foumart, Mustela putorius",
+    "359": "black-footed ferret, ferret, Mustela nigripes",
+    "360": "otter",
+    "361": "skunk, polecat, wood pussy",
+    "362": "badger",
+    "363": "armadillo",
+    "364": "three-toed sloth, ai, Bradypus tridactylus",
+    "365": "orangutan, orang, orangutang, Pongo pygmaeus",
+    "366": "gorilla, Gorilla gorilla",
+    "367": "chimpanzee, chimp, Pan troglodytes",
+    "368": "gibbon, Hylobates lar",
+    "369": "siamang, Hylobates syndactylus, Symphalangus syndactylus",
+    "370": "guenon, guenon monkey",
+    "371": "patas, hussar monkey, Erythrocebus patas",
+    "372": "baboon",
+    "373": "macaque",
+    "374": "langur",
+    "375": "colobus, colobus monkey",
+    "376": "proboscis monkey, Nasalis larvatus",
+    "377": "marmoset",
+    "378": "capuchin, ringtail, Cebus capucinus",
+    "379": "howler monkey, howler",
+    "380": "titi, titi monkey",
+    "381": "spider monkey, Ateles geoffroyi",
+    "382": "squirrel monkey, Saimiri sciureus",
+    "383": "Madagascar cat, ring-tailed lemur, Lemur catta",
+    "384": "indri, indris, Indri indri, Indri brevicaudatus",
+    "385": "Indian elephant, Elephas maximus",
+    "386": "African elephant, Loxodonta africana",
+    "387": "lesser panda, red panda, panda, bear cat, cat bear, Ailurus fulgens",
+    "388": "giant panda, panda, panda bear, coon bear, Ailuropoda melanoleuca",
+    "389": "barracouta, snoek",
+    "390": "eel",
+    "391": "coho, cohoe, coho salmon, blue jack, silver salmon, Oncorhynchus kisutch",
+    "392": "rock beauty, Holocanthus tricolor",
+    "393": "anemone fish",
+    "394": "sturgeon",
+    "395": "gar, garfish, garpike, billfish, Lepisosteus osseus",
+    "396": "lionfish",
+    "397": "puffer, pufferfish, blowfish, globefish",
+    "398": "abacus",
+    "399": "abaya",
+    "400": "academic gown, academic robe, judge robe",
+    "401": "accordion, piano accordion, squeeze box",
+    "402": "acoustic guitar",
+    "403": "aircraft carrier, carrier, flattop, attack aircraft carrier",
+    "404": "airliner",
+    "405": "airship, dirigible",
+    "406": "altar",
+    "407": "ambulance",
+    "408": "amphibian, amphibious vehicle",
+    "409": "analog clock",
+    "410": "apiary, bee house",
+    "411": "apron",
+    "412": "ashcan, trash can, garbage can, wastebin, ash bin, ash-bin, ashbin, dustbin, trash barrel, trash bin",
+    "413": "assault rifle, assault gun",
+    "414": "backpack, back pack, knapsack, packsack, rucksack, haversack",
+    "415": "bakery, bakeshop, bakehouse",
+    "416": "balance beam, beam",
+    "417": "balloon",
+    "418": "ballpoint, ballpoint pen, ballpen, Biro",
+    "419": "Band Aid",
+    "420": "banjo",
+    "421": "bannister, banister, balustrade, balusters, handrail",
+    "422": "barbell",
+    "423": "barber chair",
+    "424": "barbershop",
+    "425": "barn",
+    "426": "barometer",
+    "427": "barrel, cask",
+    "428": "barrow, garden cart, lawn cart, wheelbarrow",
+    "429": "baseball",
+    "430": "basketball",
+    "431": "bassinet",
+    "432": "bassoon",
+    "433": "bathing cap, swimming cap",
+    "434": "bath towel",
+    "435": "bathtub, bathing tub, bath, tub",
+    "436": "beach wagon, station wagon, wagon, estate car, beach waggon, station waggon, waggon",
+    "437": "beacon, lighthouse, beacon light, pharos",
+    "438": "beaker",
+    "439": "bearskin, busby, shako",
+    "440": "beer bottle",
+    "441": "beer glass",
+    "442": "bell cote, bell cot",
+    "443": "bib",
+    "444": "bicycle-built-for-two, tandem bicycle, tandem",
+    "445": "bikini, two-piece",
+    "446": "binder, ring-binder",
+    "447": "binoculars, field glasses, opera glasses",
+    "448": "birdhouse",
+    "449": "boathouse",
+    "450": "bobsled, bobsleigh, bob",
+    "451": "bolo tie, bolo, bola tie, bola",
+    "452": "bonnet, poke bonnet",
+    "453": "bookcase",
+    "454": "bookshop, bookstore, bookstall",
+    "455": "bottlecap",
+    "456": "bow",
+    "457": "bow tie, bow-tie, bowtie",
+    "458": "brass, memorial tablet, plaque",
+    "459": "brassiere, bra, bandeau",
+    "460": "breakwater, groin, groyne, mole, bulwark, seawall, jetty",
+    "461": "breastplate, aegis, egis",
+    "462": "broom",
+    "463": "bucket, pail",
+    "464": "buckle",
+    "465": "bulletproof vest",
+    "466": "bullet train, bullet",
+    "467": "butcher shop, meat market",
+    "468": "cab, hack, taxi, taxicab",
+    "469": "caldron, cauldron",
+    "470": "candle, taper, wax light",
+    "471": "cannon",
+    "472": "canoe",
+    "473": "can opener, tin opener",
+    "474": "cardigan",
+    "475": "car mirror",
+    "476": "carousel, carrousel, merry-go-round, roundabout, whirligig",
+    "477": "carpenters kit, tool kit",
+    "478": "carton",
+    "479": "car wheel",
+    "480": "cash machine, cash dispenser, automated teller machine, automatic teller machine, automated teller, automatic teller, ATM",
+    "481": "cassette",
+    "482": "cassette player",
+    "483": "castle",
+    "484": "catamaran",
+    "485": "CD player",
+    "486": "cello, violoncello",
+    "487": "cellular telephone, cellular phone, cellphone, cell, mobile phone",
+    "488": "chain",
+    "489": "chainlink fence",
+    "490": "chain mail, ring mail, mail, chain armor, chain armour, ring armor, ring armour",
+    "491": "chain saw, chainsaw",
+    "492": "chest",
+    "493": "chiffonier, commode",
+    "494": "chime, bell, gong",
+    "495": "china cabinet, china closet",
+    "496": "Christmas stocking",
+    "497": "church, church building",
+    "498": "cinema, movie theater, movie theatre, movie house, picture palace",
+    "499": "cleaver, meat cleaver, chopper",
+    "500": "cliff dwelling",
+    "501": "cloak",
+    "502": "clog, geta, patten, sabot",
+    "503": "cocktail shaker",
+    "504": "coffee mug",
+    "505": "coffeepot",
+    "506": "coil, spiral, volute, whorl, helix",
+    "507": "combination lock",
+    "508": "computer keyboard, keypad",
+    "509": "confectionery, confectionary, candy store",
+    "510": "container ship, containership, container vessel",
+    "511": "convertible",
+    "512": "corkscrew, bottle screw",
+    "513": "cornet, horn, trumpet, trump",
+    "514": "cowboy boot",
+    "515": "cowboy hat, ten-gallon hat",
+    "516": "cradle",
+    "517": "crane",
+    "518": "crash helmet",
+    "519": "crate",
+    "520": "crib, cot",
+    "521": "Crock Pot",
+    "522": "croquet ball",
+    "523": "crutch",
+    "524": "cuirass",
+    "525": "dam, dike, dyke",
+    "526": "desk",
+    "527": "desktop computer",
+    "528": "dial telephone, dial phone",
+    "529": "diaper, nappy, napkin",
+    "530": "digital clock",
+    "531": "digital watch",
+    "532": "dining table, board",
+    "533": "dishrag, dishcloth",
+    "534": "dishwasher, dish washer, dishwashing machine",
+    "535": "disk brake, disc brake",
+    "536": "dock, dockage, docking facility",
+    "537": "dogsled, dog sled, dog sleigh",
+    "538": "dome",
+    "539": "doormat, welcome mat",
+    "540": "drilling platform, offshore rig",
+    "541": "drum, membranophone, tympan",
+    "542": "drumstick",
+    "543": "dumbbell",
+    "544": "Dutch oven",
+    "545": "electric fan, blower",
+    "546": "electric guitar",
+    "547": "electric locomotive",
+    "548": "entertainment center",
+    "549": "envelope",
+    "550": "espresso maker",
+    "551": "face powder",
+    "552": "feather boa, boa",
+    "553": "file, file cabinet, filing cabinet",
+    "554": "fireboat",
+    "555": "fire engine, fire truck",
+    "556": "fire screen, fireguard",
+    "557": "flagpole, flagstaff",
+    "558": "flute, transverse flute",
+    "559": "folding chair",
+    "560": "football helmet",
+    "561": "forklift",
+    "562": "fountain",
+    "563": "fountain pen",
+    "564": "four-poster",
+    "565": "freight car",
+    "566": "French horn, horn",
+    "567": "frying pan, frypan, skillet",
+    "568": "fur coat",
+    "569": "garbage truck, dustcart",
+    "570": "gasmask, respirator, gas helmet",
+    "571": "gas pump, gasoline pump, petrol pump, island dispenser",
+    "572": "goblet",
+    "573": "go-kart",
+    "574": "golf ball",
+    "575": "golfcart, golf cart",
+    "576": "gondola",
+    "577": "gong, tam-tam",
+    "578": "gown",
+    "579": "grand piano, grand",
+    "580": "greenhouse, nursery, glasshouse",
+    "581": "grille, radiator grille",
+    "582": "grocery store, grocery, food market, market",
+    "583": "guillotine",
+    "584": "hair slide",
+    "585": "hair spray",
+    "586": "half track",
+    "587": "hammer",
+    "588": "hamper",
+    "589": "hand blower, blow dryer, blow drier, hair dryer, hair drier",
+    "590": "hand-held computer, hand-held microcomputer",
+    "591": "handkerchief, hankie, hanky, hankey",
+    "592": "hard disc, hard disk, fixed disk",
+    "593": "harmonica, mouth organ, harp, mouth harp",
+    "594": "harp",
+    "595": "harvester, reaper",
+    "596": "hatchet",
+    "597": "holster",
+    "598": "home theater, home theatre",
+    "599": "honeycomb",
+    "600": "hook, claw",
+    "601": "hoopskirt, crinoline",
+    "602": "horizontal bar, high bar",
+    "603": "horse cart, horse-cart",
+    "604": "hourglass",
+    "605": "iPod",
+    "606": "iron, smoothing iron",
+    "607": "jack-o-lantern",
+    "608": "jean, blue jean, denim",
+    "609": "jeep, landrover",
+    "610": "jersey, T-shirt, tee shirt",
+    "611": "jigsaw puzzle",
+    "612": "jinrikisha, ricksha, rickshaw",
+    "613": "joystick",
+    "614": "kimono",
+    "615": "knee pad",
+    "616": "knot",
+    "617": "lab coat, laboratory coat",
+    "618": "ladle",
+    "619": "lampshade, lamp shade",
+    "620": "laptop, laptop computer",
+    "621": "lawn mower, mower",
+    "622": "lens cap, lens cover",
+    "623": "letter opener, paper knife, paperknife",
+    "624": "library",
+    "625": "lifeboat",
+    "626": "lighter, light, igniter, ignitor",
+    "627": "limousine, limo",
+    "628": "liner, ocean liner",
+    "629": "lipstick, lip rouge",
+    "630": "Loafer",
+    "631": "lotion",
+    "632": "loudspeaker, speaker, speaker unit, loudspeaker system, speaker system",
+    "633": "loupe, jewelers loupe",
+    "634": "lumbermill, sawmill",
+    "635": "magnetic compass",
+    "636": "mailbag, postbag",
+    "637": "mailbox, letter box",
+    "638": "maillot",
+    "639": "maillot, tank suit",
+    "640": "manhole cover",
+    "641": "maraca",
+    "642": "marimba, xylophone",
+    "643": "mask",
+    "644": "matchstick",
+    "645": "maypole",
+    "646": "maze, labyrinth",
+    "647": "measuring cup",
+    "648": "medicine chest, medicine cabinet",
+    "649": "megalith, megalithic structure",
+    "650": "microphone, mike",
+    "651": "microwave, microwave oven",
+    "652": "military uniform",
+    "653": "milk can",
+    "654": "minibus",
+    "655": "miniskirt, mini",
+    "656": "minivan",
+    "657": "missile",
+    "658": "mitten",
+    "659": "mixing bowl",
+    "660": "mobile home, manufactured home",
+    "661": "Model T",
+    "662": "modem",
+    "663": "monastery",
+    "664": "monitor",
+    "665": "moped",
+    "666": "mortar",
+    "667": "mortarboard",
+    "668": "mosque",
+    "669": "mosquito net",
+    "670": "motor scooter, scooter",
+    "671": "mountain bike, all-terrain bike, off-roader",
+    "672": "mountain tent",
+    "673": "mouse, computer mouse",
+    "674": "mousetrap",
+    "675": "moving van",
+    "676": "muzzle",
+    "677": "nail",
+    "678": "neck brace",
+    "679": "necklace",
+    "680": "nipple",
+    "681": "notebook, notebook computer",
+    "682": "obelisk",
+    "683": "oboe, hautboy, hautbois",
+    "684": "ocarina, sweet potato",
+    "685": "odometer, hodometer, mileometer, milometer",
+    "686": "oil filter",
+    "687": "organ, pipe organ",
+    "688": "oscilloscope, scope, cathode-ray oscilloscope, CRO",
+    "689": "overskirt",
+    "690": "oxcart",
+    "691": "oxygen mask",
+    "692": "packet",
+    "693": "paddle, boat paddle",
+    "694": "paddlewheel, paddle wheel",
+    "695": "padlock",
+    "696": "paintbrush",
+    "697": "pajama, pyjama, pjs, jammies",
+    "698": "palace",
+    "699": "panpipe, pandean pipe, syrinx",
+    "700": "paper towel",
+    "701": "parachute, chute",
+    "702": "parallel bars, bars",
+    "703": "park bench",
+    "704": "parking meter",
+    "705": "passenger car, coach, carriage",
+    "706": "patio, terrace",
+    "707": "pay-phone, pay-station",
+    "708": "pedestal, plinth, footstall",
+    "709": "pencil box, pencil case",
+    "710": "pencil sharpener",
+    "711": "perfume, essence",
+    "712": "Petri dish",
+    "713": "photocopier",
+    "714": "pick, plectrum, plectron",
+    "715": "pickelhaube",
+    "716": "picket fence, paling",
+    "717": "pickup, pickup truck",
+    "718": "pier",
+    "719": "piggy bank, penny bank",
+    "720": "pill bottle",
+    "721": "pillow",
+    "722": "ping-pong ball",
+    "723": "pinwheel",
+    "724": "pirate, pirate ship",
+    "725": "pitcher, ewer",
+    "726": "plane, carpenters plane, woodworking plane",
+    "727": "planetarium",
+    "728": "plastic bag",
+    "729": "plate rack",
+    "730": "plow, plough",
+    "731": "plunger, plumbers helper",
+    "732": "Polaroid camera, Polaroid Land camera",
+    "733": "pole",
+    "734": "police van, police wagon, paddy wagon, patrol wagon, wagon, black Maria",
+    "735": "poncho",
+    "736": "pool table, billiard table, snooker table",
+    "737": "pop bottle, soda bottle",
+    "738": "pot, flowerpot",
+    "739": "potters wheel",
+    "740": "power drill",
+    "741": "prayer rug, prayer mat",
+    "742": "printer",
+    "743": "prison, prison house",
+    "744": "projectile, missile",
+    "745": "projector",
+    "746": "puck, hockey puck",
+    "747": "punching bag, punch bag, punching ball, punchball",
+    "748": "purse",
+    "749": "quill, quill pen",
+    "750": "quilt, comforter, comfort, puff",
+    "751": "racer, race car, racing car",
+    "752": "racket, racquet",
+    "753": "radiator",
+    "754": "radio, wireless",
+    "755": "radio telescope, radio reflector",
+    "756": "rain barrel",
+    "757": "recreational vehicle, RV, R.V.",
+    "758": "reel",
+    "759": "reflex camera",
+    "760": "refrigerator, icebox",
+    "761": "remote control, remote",
+    "762": "restaurant, eating house, eating place, eatery",
+    "763": "revolver, six-gun, six-shooter",
+    "764": "rifle",
+    "765": "rocking chair, rocker",
+    "766": "rotisserie",
+    "767": "rubber eraser, rubber, pencil eraser",
+    "768": "rugby ball",
+    "769": "rule, ruler",
+    "770": "running shoe",
+    "771": "safe",
+    "772": "safety pin",
+    "773": "saltshaker, salt shaker",
+    "774": "sandal",
+    "775": "sarong",
+    "776": "sax, saxophone",
+    "777": "scabbard",
+    "778": "scale, weighing machine",
+    "779": "school bus",
+    "780": "schooner",
+    "781": "scoreboard",
+    "782": "screen, CRT screen",
+    "783": "screw",
+    "784": "screwdriver",
+    "785": "seat belt, seatbelt",
+    "786": "sewing machine",
+    "787": "shield, buckler",
+    "788": "shoe shop, shoe-shop, shoe store",
+    "789": "shoji",
+    "790": "shopping basket",
+    "791": "shopping cart",
+    "792": "shovel",
+    "793": "shower cap",
+    "794": "shower curtain",
+    "795": "ski",
+    "796": "ski mask",
+    "797": "sleeping bag",
+    "798": "slide rule, slipstick",
+    "799": "sliding door",
+    "800": "slot, one-armed bandit",
+    "801": "snorkel",
+    "802": "snowmobile",
+    "803": "snowplow, snowplough",
+    "804": "soap dispenser",
+    "805": "soccer ball",
+    "806": "sock",
+    "807": "solar dish, solar collector, solar furnace",
+    "808": "sombrero",
+    "809": "soup bowl",
+    "810": "space bar",
+    "811": "space heater",
+    "812": "space shuttle",
+    "813": "spatula",
+    "814": "speedboat",
+    "815": "spider web, spiders web",
+    "816": "spindle",
+    "817": "sports car, sport car",
+    "818": "spotlight, spot",
+    "819": "stage",
+    "820": "steam locomotive",
+    "821": "steel arch bridge",
+    "822": "steel drum",
+    "823": "stethoscope",
+    "824": "stole",
+    "825": "stone wall",
+    "826": "stopwatch, stop watch",
+    "827": "stove",
+    "828": "strainer",
+    "829": "streetcar, tram, tramcar, trolley, trolley car",
+    "830": "stretcher",
+    "831": "studio couch, day bed",
+    "832": "stupa, tope",
+    "833": "submarine, pigboat, sub, U-boat",
+    "834": "suit, suit of clothes",
+    "835": "sundial",
+    "836": "sunglass",
+    "837": "sunglasses, dark glasses, shades",
+    "838": "sunscreen, sunblock, sun blocker",
+    "839": "suspension bridge",
+    "840": "swab, swob, mop",
+    "841": "sweatshirt",
+    "842": "swimming trunks, bathing trunks",
+    "843": "swing",
+    "844": "switch, electric switch, electrical switch",
+    "845": "syringe",
+    "846": "table lamp",
+    "847": "tank, army tank, armored combat vehicle, armoured combat vehicle",
+    "848": "tape player",
+    "849": "teapot",
+    "850": "teddy, teddy bear",
+    "851": "television, television system",
+    "852": "tennis ball",
+    "853": "thatch, thatched roof",
+    "854": "theater curtain, theatre curtain",
+    "855": "thimble",
+    "856": "thresher, thrasher, threshing machine",
+    "857": "throne",
+    "858": "tile roof",
+    "859": "toaster",
+    "860": "tobacco shop, tobacconist shop, tobacconist",
+    "861": "toilet seat",
+    "862": "torch",
+    "863": "totem pole",
+    "864": "tow truck, tow car, wrecker",
+    "865": "toyshop",
+    "866": "tractor",
+    "867": "trailer truck, tractor trailer, trucking rig, rig, articulated lorry, semi",
+    "868": "tray",
+    "869": "trench coat",
+    "870": "tricycle, trike, velocipede",
+    "871": "trimaran",
+    "872": "tripod",
+    "873": "triumphal arch",
+    "874": "trolleybus, trolley coach, trackless trolley",
+    "875": "trombone",
+    "876": "tub, vat",
+    "877": "turnstile",
+    "878": "typewriter keyboard",
+    "879": "umbrella",
+    "880": "unicycle, monocycle",
+    "881": "upright, upright piano",
+    "882": "vacuum, vacuum cleaner",
+    "883": "vase",
+    "884": "vault",
+    "885": "velvet",
+    "886": "vending machine",
+    "887": "vestment",
+    "888": "viaduct",
+    "889": "violin, fiddle",
+    "890": "volleyball",
+    "891": "waffle iron",
+    "892": "wall clock",
+    "893": "wallet, billfold, notecase, pocketbook",
+    "894": "wardrobe, closet, press",
+    "895": "warplane, military plane",
+    "896": "washbasin, handbasin, washbowl, lavabo, wash-hand basin",
+    "897": "washer, automatic washer, washing machine",
+    "898": "water bottle",
+    "899": "water jug",
+    "900": "water tower",
+    "901": "whiskey jug",
+    "902": "whistle",
+    "903": "wig",
+    "904": "window screen",
+    "905": "window shade",
+    "906": "Windsor tie",
+    "907": "wine bottle",
+    "908": "wing",
+    "909": "wok",
+    "910": "wooden spoon",
+    "911": "wool, woolen, woollen",
+    "912": "worm fence, snake fence, snake-rail fence, Virginia fence",
+    "913": "wreck",
+    "914": "yawl",
+    "915": "yurt",
+    "916": "web site, website, internet site, site",
+    "917": "comic book",
+    "918": "crossword puzzle, crossword",
+    "919": "street sign",
+    "920": "traffic light, traffic signal, stoplight",
+    "921": "book jacket, dust cover, dust jacket, dust wrapper",
+    "922": "menu",
+    "923": "plate",
+    "924": "guacamole",
+    "925": "consomme",
+    "926": "hot pot, hotpot",
+    "927": "trifle",
+    "928": "ice cream, icecream",
+    "929": "ice lolly, lolly, lollipop, popsicle",
+    "930": "French loaf",
+    "931": "bagel, beigel",
+    "932": "pretzel",
+    "933": "cheeseburger",
+    "934": "hotdog, hot dog, red hot",
+    "935": "mashed potato",
+    "936": "head cabbage",
+    "937": "broccoli",
+    "938": "cauliflower",
+    "939": "zucchini, courgette",
+    "940": "spaghetti squash",
+    "941": "acorn squash",
+    "942": "butternut squash",
+    "943": "cucumber, cuke",
+    "944": "artichoke, globe artichoke",
+    "945": "bell pepper",
+    "946": "cardoon",
+    "947": "mushroom",
+    "948": "Granny Smith",
+    "949": "strawberry",
+    "950": "orange",
+    "951": "lemon",
+    "952": "fig",
+    "953": "pineapple, ananas",
+    "954": "banana",
+    "955": "jackfruit, jak, jack",
+    "956": "custard apple",
+    "957": "pomegranate",
+    "958": "hay",
+    "959": "carbonara",
+    "960": "chocolate sauce, chocolate syrup",
+    "961": "dough",
+    "962": "meat loaf, meatloaf",
+    "963": "pizza, pizza pie",
+    "964": "potpie",
+    "965": "burrito",
+    "966": "red wine",
+    "967": "espresso",
+    "968": "cup",
+    "969": "eggnog",
+    "970": "alp",
+    "971": "bubble",
+    "972": "cliff, drop, drop-off",
+    "973": "coral reef",
+    "974": "geyser",
+    "975": "lakeside, lakeshore",
+    "976": "promontory, headland, head, foreland",
+    "977": "sandbar, sand bar",
+    "978": "seashore, coast, seacoast, sea-coast",
+    "979": "valley, vale",
+    "980": "volcano",
+    "981": "ballplayer, baseball player",
+    "982": "groom, bridegroom",
+    "983": "scuba diver",
+    "984": "rapeseed",
+    "985": "daisy",
+    "986": "yellow ladys slipper, yellow lady-slipper, Cypripedium calceolus, Cypripedium parviflorum",
+    "987": "corn",
+    "988": "acorn",
+    "989": "hip, rose hip, rosehip",
+    "990": "buckeye, horse chestnut, conker",
+    "991": "coral fungus",
+    "992": "agaric",
+    "993": "gyromitra",
+    "994": "stinkhorn, carrion fungus",
+    "995": "earthstar",
+    "996": "hen-of-the-woods, hen of the woods, Polyporus frondosus, Grifola frondosa",
+    "997": "bolete",
+    "998": "ear, spike, capitulum",
+    "999": "toilet tissue, toilet paper, bathroom tissue"
+  }
+}

PixelGen-XL-16-256/pipeline.py

@@ -0,0 +1,333 @@
+"""Hub custom pipeline: PixelGenC2IPipeline.
+Load with native Hugging Face diffusers and trust_remote_code=True.
+"""
+
+from __future__ import annotations
+
+import importlib
+import json
+import sys
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from diffusers.utils.torch_utils import randn_tensor
+
+
+RECOMMENDED_NOISE_BY_SIZE = {
+    256: 1.0,
+    512: 2.0,
+}
+
+
+class PixelGenC2IPipeline(DiffusionPipeline):
+    model_cpu_offload_seq = "transformer"
+
+    def __init__(
+        self,
+        transformer,
+        scheduler,
+        id2label: Optional[Dict[Union[int, str], str]] = None,
+    ):
+        if isinstance(transformer, (list, tuple)):
+            transformer = self._load_transformer(Path(__file__).resolve().parent)
+        if isinstance(scheduler, (list, tuple)):
+            scheduler = self._load_scheduler(Path(__file__).resolve().parent)
+
+        super().__init__()
+        self.register_modules(transformer=transformer, scheduler=scheduler)
+        self._id2label = self._normalize_id2label(id2label)
+        self.labels = self._build_label2id(self._id2label)
+        self._labels_loaded_from_model_index = bool(self._id2label)
+
+    @staticmethod
+    def _resolve_variant_dir(pretrained_model_name_or_path: Optional[str]) -> Path:
+        repo_root = Path(__file__).resolve().parent
+        if pretrained_model_name_or_path in (None, "", "."):
+            return repo_root
+        variant = Path(pretrained_model_name_or_path)
+        if not variant.is_absolute():
+            candidate = (Path.cwd() / variant).resolve()
+            variant = candidate if candidate.exists() else (repo_root / variant).resolve()
+        return variant
+
+    @classmethod
+    def _load_transformer(cls, variant: Path, **model_kwargs):
+        comp_dir = variant / "transformer"
+        module_path = comp_dir / "transformer_jit.py"
+        if not module_path.exists():
+            raise FileNotFoundError(f"Missing transformer module at {module_path}")
+        comp_path = str(comp_dir)
+        if comp_path not in sys.path:
+            sys.path.insert(0, comp_path)
+        module = importlib.import_module("transformer_jit")
+        transformer_cls = getattr(module, "JiTTransformer2DModel")
+        return transformer_cls.from_pretrained(str(comp_dir), **model_kwargs)
+
+    @classmethod
+    def _load_scheduler(cls, variant: Path):
+        scheduler_dir = variant / "scheduler"
+        config_path = scheduler_dir / "scheduler_config.json"
+        if not config_path.exists():
+            raise FileNotFoundError(f"Missing scheduler config at {config_path}")
+        scheduler_path = str(scheduler_dir)
+        if scheduler_path not in sys.path:
+            sys.path.insert(0, scheduler_path)
+        module = importlib.import_module("scheduling_pixelgen")
+        scheduler_cls = getattr(module, "PixelGenScheduler")
+        return scheduler_cls.from_pretrained(str(scheduler_dir))
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
+        variant = cls._resolve_variant_dir(pretrained_model_name_or_path)
+        model_kwargs = dict(kwargs)
+        transformer = cls._load_transformer(variant, **model_kwargs)
+        scheduler = cls._load_scheduler(variant)
+        id2label = cls._read_id2label_from_model_index(str(variant))
+        return cls(transformer=transformer, scheduler=scheduler, id2label=id2label)
+
+    @staticmethod
+    def _normalize_id2label(id2label: Optional[Dict[Union[int, str], str]]) -> Dict[int, str]:
+        if not id2label:
+            return {}
+        return {int(key): value for key, value in id2label.items()}
+
+    @staticmethod
+    def _read_id2label_from_model_index(variant_path: Optional[str]) -> Dict[int, str]:
+        if not variant_path:
+            return {}
+        model_index_path = Path(variant_path).resolve() / "model_index.json"
+        if not model_index_path.exists():
+            return {}
+        raw = json.loads(model_index_path.read_text(encoding="utf-8"))
+        id2label = raw.get("id2label")
+        if not isinstance(id2label, dict):
+            return {}
+        return {int(key): value for key, value in id2label.items()}
+
+    @staticmethod
+    def _build_label2id(id2label: Dict[int, str]) -> Dict[str, int]:
+        label2id: Dict[str, int] = {}
+        for class_id, value in id2label.items():
+            for synonym in value.split(","):
+                synonym = synonym.strip()
+                if synonym:
+                    label2id[synonym] = int(class_id)
+        return dict(sorted(label2id.items()))
+
+    def _ensure_labels_loaded(self) -> None:
+        if self._labels_loaded_from_model_index:
+            return
+        loaded = self._read_id2label_from_model_index(getattr(self.config, "_name_or_path", None))
+        if loaded:
+            self._id2label = loaded
+            self.labels = self._build_label2id(self._id2label)
+        self._labels_loaded_from_model_index = True
+
+    @property
+    def id2label(self) -> Dict[int, str]:
+        self._ensure_labels_loaded()
+        return self._id2label
+
+    def get_label_ids(self, label: Union[str, List[str]]) -> List[int]:
+        self._ensure_labels_loaded()
+        if not self.labels:
+            raise ValueError("No English labels loaded. Ensure `id2label` exists in model_index.json.")
+        if isinstance(label, str):
+            label = [label]
+        missing = [item for item in label if item not in self.labels]
+        if missing:
+            preview = ", ".join(list(self.labels.keys())[:8])
+            raise ValueError(f"Unknown English label(s): {missing}. Example valid labels: {preview}, ...")
+        return [self.labels[item] for item in label]
+
+    def _normalize_class_labels(self, class_labels: Union[int, str, List[Union[int, str]]]) -> List[int]:
+        if isinstance(class_labels, int):
+            return [class_labels]
+        if isinstance(class_labels, str):
+            if class_labels in self.labels:
+                return [self.labels[class_labels]]
+            return [int(class_labels)]
+        if class_labels and isinstance(class_labels[0], str):
+            return self.get_label_ids(class_labels)
+        return list(class_labels)
+
+    def _predict_velocity(
+        self,
+        z_value: torch.Tensor,
+        t: torch.Tensor,
+        class_labels: torch.Tensor,
+        class_null: torch.Tensor,
+        do_classifier_free_guidance: bool,
+        guidance_scale: float,
+        guidance_interval_min: float,
+        guidance_interval_max: float,
+    ) -> torch.Tensor:
+        if do_classifier_free_guidance:
+            z_in = torch.cat([z_value, z_value], dim=0)
+            labels = torch.cat([class_labels, class_null], dim=0)
+        else:
+            z_in = z_value
+            labels = class_labels
+
+        t_batch = t.flatten().expand(z_in.shape[0])
+        x_pred = self.transformer(z_in, timestep=t_batch, class_labels=labels).sample
+        v = self.scheduler.velocity_from_prediction(z_in, x_pred, t)
+
+        if not do_classifier_free_guidance:
+            return v
+
+        v_cond, v_uncond = v.chunk(2, dim=0)
+        interval_mask = t < guidance_interval_max
+        if guidance_interval_min != 0.0:
+            interval_mask = interval_mask & (t > guidance_interval_min)
+        scale = torch.where(
+            interval_mask,
+            torch.tensor(guidance_scale, device=z_value.device, dtype=z_value.dtype),
+            torch.tensor(1.0, device=z_value.device, dtype=z_value.dtype),
+        )
+        return v_uncond + scale * (v_cond - v_uncond)
+
+    def _run_sampler(
+        self,
+        latents: torch.Tensor,
+        class_labels: torch.Tensor,
+        class_null: torch.Tensor,
+        num_inference_steps: int,
+        do_classifier_free_guidance: bool,
+        guidance_scale: float,
+        guidance_interval_min: float,
+        guidance_interval_max: float,
+        sampling_method: str,
+        timeshift: float,
+    ) -> torch.Tensor:
+        device = latents.device
+        self.scheduler.set_timesteps(
+            num_inference_steps, device=device, solver=sampling_method, timeshift=timeshift
+        )
+        timesteps = self.scheduler.timesteps
+
+        for i in self.progress_bar(range(num_inference_steps - 1)):
+            t = timesteps[i]
+            t_next = timesteps[i + 1]
+            v = self._predict_velocity(
+                latents,
+                t,
+                class_labels,
+                class_null,
+                do_classifier_free_guidance,
+                guidance_scale,
+                guidance_interval_min,
+                guidance_interval_max,
+            )
+
+            if sampling_method == "heun":
+                latents_euler = latents + (t_next - t) * v
+                v_next = self._predict_velocity(
+                    latents_euler,
+                    t_next,
+                    class_labels,
+                    class_null,
+                    do_classifier_free_guidance,
+                    guidance_scale,
+                    guidance_interval_min,
+                    guidance_interval_max,
+                )
+                latents = self.scheduler.step(v, t, latents, model_output_next=v_next).prev_sample
+            else:
+                latents = self.scheduler.step(v, t, latents).prev_sample
+
+        t = timesteps[-2]
+        t_next = timesteps[-1]
+        v = self._predict_velocity(
+            latents,
+            t,
+            class_labels,
+            class_null,
+            do_classifier_free_guidance,
+            guidance_scale,
+            guidance_interval_min,
+            guidance_interval_max,
+        )
+        return latents + (t_next - t) * v
+
+    @torch.inference_mode()
+    def __call__(
+        self,
+        class_labels: Union[int, str, List[Union[int, str]]],
+        guidance_scale: Optional[float] = None,
+        guidance_interval_min: float = 0.1,
+        guidance_interval_max: float = 0.9,
+        noise_scale: Optional[float] = None,
+        t_eps: Optional[float] = None,
+        sampling_method: Optional[str] = None,
+        timeshift: Optional[float] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        num_inference_steps: int = 50,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+    ) -> Union[ImagePipelineOutput, Tuple]:
+        solver = sampling_method or self.scheduler.config.solver
+        if solver not in {"heun", "euler"}:
+            raise ValueError("sampling_method must be one of: 'heun', 'euler'.")
+        if num_inference_steps < 2:
+            raise ValueError("num_inference_steps must be >= 2.")
+        if output_type not in {"pil", "np", "pt"}:
+            raise ValueError("output_type must be one of: 'pil', 'np', 'pt'.")
+
+        if t_eps is not None:
+            self.scheduler.register_to_config(t_eps=t_eps)
+        shift = timeshift if timeshift is not None else self.scheduler.config.timeshift
+
+        class_label_ids = self._normalize_class_labels(class_labels)
+        do_classifier_free_guidance = guidance_scale is not None and guidance_scale > 1.0
+        batch_size = len(class_label_ids)
+        image_size = int(self.transformer.config.sample_size)
+        channels = int(self.transformer.config.in_channels)
+        null_class_val = int(getattr(self.transformer.config, "num_classes", 1000))
+
+        if guidance_scale is None:
+            guidance_scale = 1.0
+        if noise_scale is None:
+            noise_scale = RECOMMENDED_NOISE_BY_SIZE.get(image_size, 1.0)
+
+        latents = randn_tensor(
+            shape=(batch_size, channels, image_size, image_size),
+            generator=generator,
+            device=self._execution_device,
+            dtype=self.transformer.dtype,
+        ) * noise_scale
+
+        class_labels_t = torch.tensor(class_label_ids, device=self._execution_device, dtype=torch.long).reshape(-1)
+        class_labels_t = class_labels_t.clamp(0, null_class_val - 1)
+        class_null = torch.full_like(class_labels_t, null_class_val)
+
+        latents = self._run_sampler(
+            latents,
+            class_labels_t,
+            class_null,
+            num_inference_steps,
+            do_classifier_free_guidance,
+            guidance_scale,
+            guidance_interval_min,
+            guidance_interval_max,
+            solver,
+            shift,
+        )
+
+        images_pt = ((latents.float().clamp(-1, 1) + 1.0) / 2.0).cpu()
+        if output_type == "pt":
+            images = images_pt
+        elif output_type == "np":
+            images = images_pt.permute(0, 2, 3, 1).numpy()
+        else:
+            images = self.numpy_to_pil(images_pt.permute(0, 2, 3, 1).numpy())
+
+        self.maybe_free_model_hooks()
+        if not return_dict:
+            return (images,)
+        return ImagePipelineOutput(images=images)
+
+
+PixelGenC2IPipelineOutput = ImagePipelineOutput

PixelGen-XL-16-256/scheduler/scheduler_config.json

@@ -0,0 +1,10 @@
+{
+  "_class_name": "PixelGenScheduler",
+  "_diffusers_version": "0.38.0",
+  "adam_lm_order": 2,
+  "last_step": null,
+  "num_train_timesteps": 1000,
+  "solver": "heun",
+  "t_eps": 0.05,
+  "timeshift": 2.0
+}

PixelGen-XL-16-256/scheduler/scheduling_pixelgen.py

@@ -0,0 +1,284 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+from diffusers.utils import BaseOutput
+
+
+def lagrange_preint_o1(t1, v1, int_t_start, int_t_end):
+    int1 = int_t_end - int_t_start
+    return int1 * v1, (int1 / int1,)
+
+
+def lagrange_preint_o2(t1, t2, v1, v2, int_t_start, int_t_end):
+    int1 = 0.5 / (t1 - t2) * ((int_t_end - t2) ** 2 - (int_t_start - t2) ** 2)
+    int2 = 0.5 / (t2 - t1) * ((int_t_end - t1) ** 2 - (int_t_start - t1) ** 2)
+    int_sum = int1 + int2
+    return int1 * v1 + int2 * v2, (int1 / int_sum, int2 / int_sum)
+
+
+def lagrange_preint_o3(t1, t2, t3, v1, v2, v3, int_t_start, int_t_end):
+    int1_denom = (t1 - t2) * (t1 - t3)
+    int1_end = 1 / 3 * (int_t_end) ** 3 - 1 / 2 * (t2 + t3) * (int_t_end) ** 2 + (t2 * t3) * int_t_end
+    int1_start = 1 / 3 * (int_t_start) ** 3 - 1 / 2 * (t2 + t3) * (int_t_start) ** 2 + (t2 * t3) * int_t_start
+    int1 = (int1_end - int1_start) / int1_denom
+    int2_denom = (t2 - t1) * (t2 - t3)
+    int2_end = 1 / 3 * (int_t_end) ** 3 - 1 / 2 * (t1 + t3) * (int_t_end) ** 2 + (t1 * t3) * int_t_end
+    int2_start = 1 / 3 * (int_t_start) ** 3 - 1 / 2 * (t1 + t3) * (int_t_start) ** 2 + (t1 * t3) * int_t_start
+    int2 = (int2_end - int2_start) / int2_denom
+    int3_denom = (t3 - t1) * (t3 - t2)
+    int3_end = 1 / 3 * (int_t_end) ** 3 - 1 / 2 * (t1 + t2) * (int_t_end) ** 2 + (t1 * t2) * int_t_end
+    int3_start = 1 / 3 * (int_t_start) ** 3 - 1 / 2 * (t1 + t2) * (int_t_start) ** 2 + (t1 * t2) * int_t_start
+    int3 = (int3_end - int3_start) / int3_denom
+    int_sum = int1 + int2 + int3
+    return int1 * v1 + int2 * v2 + int3 * v3, (int1 / int_sum, int2 / int_sum, int3 / int_sum)
+
+
+def larange_preint_o4(t1, t2, t3, t4, v1, v2, v3, v4, int_t_start, int_t_end):
+    int1_denom = (t1 - t2) * (t1 - t3) * (t1 - t4)
+    int1_end = (
+        1 / 4 * (int_t_end) ** 4
+        - 1 / 3 * (t2 + t3 + t4) * (int_t_end) ** 3
+        + 1 / 2 * (t3 * t4 + t2 * t3 + t2 * t4) * int_t_end**2
+        - t2 * t3 * t4 * int_t_end
+    )
+    int1_start = (
+        1 / 4 * (int_t_start) ** 4
+        - 1 / 3 * (t2 + t3 + t4) * (int_t_start) ** 3
+        + 1 / 2 * (t3 * t4 + t2 * t3 + t2 * t4) * int_t_start**2
+        - t2 * t3 * t4 * int_t_start
+    )
+    int1 = (int1_end - int1_start) / int1_denom
+    int2_denom = (t2 - t1) * (t2 - t3) * (t2 - t4)
+    int2_end = (
+        1 / 4 * (int_t_end) ** 4
+        - 1 / 3 * (t1 + t3 + t4) * (int_t_end) ** 3
+        + 1 / 2 * (t3 * t4 + t1 * t3 + t1 * t4) * int_t_end**2
+        - t1 * t3 * t4 * int_t_end
+    )
+    int2_start = (
+        1 / 4 * (int_t_start) ** 4
+        - 1 / 3 * (t1 + t3 + t4) * (int_t_start) ** 3
+        + 1 / 2 * (t3 * t4 + t1 * t3 + t1 * t4) * int_t_start**2
+        - t1 * t3 * t4 * int_t_start
+    )
+    int2 = (int2_end - int2_start) / int2_denom
+    int3_denom = (t3 - t1) * (t3 - t2) * (t3 - t4)
+    int3_end = (
+        1 / 4 * (int_t_end) ** 4
+        - 1 / 3 * (t1 + t2 + t4) * (int_t_end) ** 3
+        + 1 / 2 * (t4 * t2 + t1 * t2 + t1 * t4) * int_t_end**2
+        - t1 * t2 * t4 * int_t_end
+    )
+    int3_start = (
+        1 / 4 * (int_t_start) ** 4
+        - 1 / 3 * (t1 + t2 + t4) * (int_t_start) ** 3
+        + 1 / 2 * (t4 * t2 + t1 * t2 + t1 * t4) * int_t_start**2
+        - t1 * t2 * t4 * int_t_start
+    )
+    int3 = (int3_end - int3_start) / int3_denom
+    int4_denom = (t4 - t1) * (t4 - t2) * (t4 - t3)
+    int4_end = (
+        1 / 4 * (int_t_end) ** 4
+        - 1 / 3 * (t1 + t2 + t3) * (int_t_end) ** 3
+        + 1 / 2 * (t3 * t2 + t1 * t2 + t1 * t3) * int_t_end**2
+        - t1 * t2 * t3 * int_t_end
+    )
+    int4_start = (
+        1 / 4 * (int_t_start) ** 4
+        - 1 / 3 * (t1 + t2 + t3) * (int_t_start) ** 3
+        + 1 / 2 * (t3 * t2 + t1 * t2 + t1 * t3) * int_t_start**2
+        - t1 * t2 * t3 * int_t_start
+    )
+    int4 = (int4_end - int4_start) / int4_denom
+    int_sum = int1 + int2 + int3 + int4
+    return int1 * v1 + int2 * v2 + int3 * v3 + int4 * v4, (
+        int1 / int_sum,
+        int2 / int_sum,
+        int3 / int_sum,
+        int4 / int_sum,
+    )
+
+
+def lagrange_preint(order, pre_vs, pre_ts, int_t_start, int_t_end):
+    order = min(order, len(pre_vs), len(pre_ts))
+    if order == 1:
+        return lagrange_preint_o1(pre_ts[-1], pre_vs[-1], int_t_start, int_t_end)
+    if order == 2:
+        return lagrange_preint_o2(pre_ts[-2], pre_ts[-1], pre_vs[-2], pre_vs[-1], int_t_start, int_t_end)
+    if order == 3:
+        return lagrange_preint_o3(
+            pre_ts[-3], pre_ts[-2], pre_ts[-1], pre_vs[-3], pre_vs[-2], pre_vs[-1], int_t_start, int_t_end
+        )
+    if order == 4:
+        return larange_preint_o4(
+            pre_ts[-4], pre_ts[-3], pre_ts[-2], pre_ts[-1],
+            pre_vs[-4], pre_vs[-3], pre_vs[-2], pre_vs[-1],
+            int_t_start, int_t_end,
+        )
+    raise ValueError("Invalid order")
+
+
+@dataclass
+class PixelGenSchedulerOutput(BaseOutput):
+    prev_sample: torch.Tensor
+
+
+def shift_respace_fn(t: torch.Tensor, shift: float = 1.0) -> torch.Tensor:
+    return t / (t + (1 - t) * shift)
+
+
+class PixelGenScheduler(SchedulerMixin, ConfigMixin):
+    """Flow-matching scheduler for PixelGen with Euler, Heun, and Adam-LM solvers."""
+
+    order = 2
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        t_eps: float = 5e-2,
+        solver: str = "heun",
+        timeshift: float = 1.0,
+        adam_lm_order: int = 2,
+        last_step: float | None = None,
+    ):
+        del num_train_timesteps
+        if solver not in {"heun", "euler", "adam_lm"}:
+            raise ValueError("solver must be one of: 'heun', 'euler', 'adam_lm'.")
+        self.timesteps: Optional[torch.Tensor] = None
+        self.timedeltas: Optional[torch.Tensor] = None
+        self.sigmas: Optional[List[float]] = None
+        self.num_inference_steps: Optional[int] = None
+        self._step_index: Optional[int] = None
+        self.solver_coeffs: Optional[list] = None
+        self._last_step = last_step
+
+    @property
+    def init_noise_sigma(self) -> float:
+        return 1.0
+
+    def set_timesteps(
+        self,
+        num_inference_steps: int,
+        device: Union[str, torch.device, None] = None,
+        solver: Optional[str] = None,
+        timeshift: Optional[float] = None,
+        adam_lm_order: Optional[int] = None,
+    ) -> None:
+        if num_inference_steps < 2:
+            raise ValueError("num_inference_steps must be >= 2.")
+
+        if solver is not None:
+            self.register_to_config(solver=solver)
+        if timeshift is not None:
+            self.register_to_config(timeshift=timeshift)
+        if adam_lm_order is not None:
+            self.register_to_config(adam_lm_order=adam_lm_order)
+
+        last_step = self._last_step
+        if last_step is None:
+            last_step = 1.0 / num_inference_steps
+
+        self.num_inference_steps = num_inference_steps
+        timesteps = torch.linspace(0.0, 1.0 - last_step, num_inference_steps, device=device, dtype=torch.float32)
+        timesteps = torch.cat([timesteps, torch.tensor([1.0], device=device, dtype=torch.float32)], dim=0)
+        self.timesteps = shift_respace_fn(timesteps, self.config.timeshift)
+        self.timedeltas = self.timesteps[1:] - self.timesteps[:-1]
+        sigma_grid = torch.linspace(0.0, 1.0, num_inference_steps, device=device, dtype=torch.float32)
+        self.sigmas = (1.0 - sigma_grid).tolist()
+        self._step_index = 0
+
+        if self.config.solver == "adam_lm":
+            self._reparameterize_coeffs()
+
+    def _reparameterize_coeffs(self) -> None:
+        solver_coeffs = [[] for _ in range(self.num_inference_steps)]
+        for i in range(self.num_inference_steps):
+            pre_vs = [1.0] * (i + 1)
+            pre_ts = self.timesteps[: i + 1]
+            int_t_start = self.timesteps[i]
+            int_t_end = self.timesteps[i + 1]
+            order = min(self.config.adam_lm_order, i + 1)
+            _, coeffs = lagrange_preint(order, pre_vs, pre_ts.tolist(), float(int_t_start), float(int_t_end))
+            solver_coeffs[i] = coeffs
+        self.solver_coeffs = solver_coeffs
+
+    def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor:
+        del timestep
+        return sample
+
+    def _resolve_step_index(self, timestep: Union[float, torch.Tensor, None]) -> int:
+        if self._step_index is not None:
+            return self._step_index
+        if self.timesteps is None:
+            raise ValueError("Call `set_timesteps` before `step`.")
+        if timestep is None:
+            return 0
+        t_value = float(timestep) if not isinstance(timestep, torch.Tensor) else float(timestep.flatten()[0])
+        matches = (self.timesteps - t_value).abs() < 1e-6
+        if matches.any():
+            return int(matches.nonzero(as_tuple=False)[0].item())
+        return 0
+
+    def step(
+        self,
+        model_output: torch.Tensor,
+        timestep: Union[float, torch.Tensor, None],
+        sample: torch.Tensor,
+        model_output_next: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[PixelGenSchedulerOutput, Tuple[torch.Tensor]]:
+        if self.timesteps is None:
+            raise ValueError("Call `set_timesteps` before `step`.")
+
+        step_index = self._resolve_step_index(timestep)
+        if step_index >= len(self.timesteps) - 1:
+            raise ValueError("Scheduler has already reached the final timestep.")
+
+        t = self.timesteps[step_index]
+        t_next = self.timesteps[step_index + 1]
+        dt = t_next - t
+
+        if self.config.solver == "heun" and model_output_next is not None:
+            prev_sample = sample + dt * 0.5 * (model_output + model_output_next)
+        else:
+            prev_sample = sample + dt * model_output
+
+        self._step_index = step_index + 1
+
+        if not return_dict:
+            return (prev_sample,)
+        return PixelGenSchedulerOutput(prev_sample=prev_sample)
+
+    def velocity_from_prediction(
+        self,
+        sample: torch.Tensor,
+        x_pred: torch.Tensor,
+        timestep: Union[float, torch.Tensor],
+    ) -> torch.Tensor:
+        t = torch.as_tensor(timestep, device=sample.device, dtype=sample.dtype)
+        while t.ndim < sample.ndim:
+            t = t.unsqueeze(-1)
+        denom = (1.0 - t).clamp_min(self.config.t_eps)
+        return (x_pred - sample) / denom
+
+    def apply_adam_lm_step(
+        self,
+        sample: torch.Tensor,
+        velocity_history: list[torch.Tensor],
+        step_index: int,
+    ) -> torch.Tensor:
+        if self.solver_coeffs is None:
+            raise ValueError("Adam-LM coefficients not initialized. Call set_timesteps with solver='adam_lm'.")
+        coeffs = self.solver_coeffs[step_index]
+        order = len(coeffs)
+        v = torch.zeros_like(sample)
+        for j, coeff in enumerate(coeffs):
+            v = v + coeff * velocity_history[-order:][j]
+        dt = self.timedeltas[step_index]
+        return sample + dt * v

PixelGen-XL-16-256/transformer/config.json

@@ -0,0 +1,21 @@
+{
+  "_class_name": "JiTTransformer2DModel",
+  "_diffusers_version": "0.38.0",
+  "attention_dropout": 0.0,
+  "bottleneck_dim": 128,
+  "dropout": 0.1,
+  "hidden_size": 1152,
+  "in_channels": 3,
+  "in_context_len": 32,
+  "in_context_start": 8,
+  "mlp_ratio": 4.0,
+  "model_type": "PixelGen-XL/16",
+  "norm_eps": 1e-06,
+  "num_attention_heads": 16,
+  "num_class_embeds": null,
+  "num_classes": 1000,
+  "num_layers": 28,
+  "patch_size": 16,
+  "sample_size": 256,
+  "use_bottleneck": false
+}

PixelGen-XL-16-256/transformer/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f4fe96addb58bafa66fa08e175cf2fb007cb93be5c6b69f5474902e605fd6ab
+size 2708016440

PixelGen-XL-16-256/transformer/transformer_jit.py

@@ -0,0 +1,684 @@
+from __future__ import annotations
+
+import argparse
+import math
+from collections.abc import Mapping
+from typing import Dict, Literal, Tuple
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import RMSNorm
+
+
+# Architecture presets aligned with the official JiT checkpoints.
+JIT_PRESET_CONFIGS: Dict[str, Dict[str, object]] = {
+    "JiT-B/16": {
+        "sample_size": 256,
+        "patch_size": 16,
+        "hidden_size": 768,
+        "num_layers": 12,
+        "num_attention_heads": 12,
+        "bottleneck_dim": 128,
+        "in_context_len": 32,
+        "in_context_start": 4,
+        "attention_dropout": 0.0,
+        "dropout": 0.0,
+    },
+    "JiT-B/32": {
+        "sample_size": 512,
+        "patch_size": 32,
+        "hidden_size": 768,
+        "num_layers": 12,
+        "num_attention_heads": 12,
+        "bottleneck_dim": 128,
+        "in_context_len": 32,
+        "in_context_start": 4,
+        "attention_dropout": 0.0,
+        "dropout": 0.0,
+    },
+    "JiT-L/16": {
+        "sample_size": 256,
+        "patch_size": 16,
+        "hidden_size": 1024,
+        "num_layers": 24,
+        "num_attention_heads": 16,
+        "bottleneck_dim": 128,
+        "in_context_len": 32,
+        "in_context_start": 8,
+        "attention_dropout": 0.0,
+        "dropout": 0.0,
+    },
+    "JiT-L/32": {
+        "sample_size": 512,
+        "patch_size": 32,
+        "hidden_size": 1024,
+        "num_layers": 24,
+        "num_attention_heads": 16,
+        "bottleneck_dim": 128,
+        "in_context_len": 32,
+        "in_context_start": 8,
+        "attention_dropout": 0.0,
+        "dropout": 0.0,
+    },
+    "JiT-H/16": {
+        "sample_size": 256,
+        "patch_size": 16,
+        "hidden_size": 1280,
+        "num_layers": 32,
+        "num_attention_heads": 16,
+        "bottleneck_dim": 256,
+        "in_context_len": 32,
+        "in_context_start": 10,
+        "attention_dropout": 0.0,
+        "dropout": 0.2,
+    },
+    "JiT-H/32": {
+        "sample_size": 512,
+        "patch_size": 32,
+        "hidden_size": 1280,
+        "num_layers": 32,
+        "num_attention_heads": 16,
+        "bottleneck_dim": 256,
+        "in_context_len": 32,
+        "in_context_start": 10,
+        "attention_dropout": 0.0,
+        "dropout": 0.2,
+    },
+    "PixelGen-XL/16": {
+        "sample_size": 256,
+        "patch_size": 16,
+        "hidden_size": 1152,
+        "num_layers": 28,
+        "num_attention_heads": 16,
+        "bottleneck_dim": 128,
+        "in_context_len": 32,
+        "in_context_start": 8,
+        "attention_dropout": 0.0,
+        "dropout": 0.1,
+        "use_bottleneck": False,
+    },
+    "PixelGen-XXL/16": {
+        "sample_size": 512,
+        "patch_size": 16,
+        "hidden_size": 1536,
+        "num_layers": 16,
+        "num_attention_heads": 24,
+        "bottleneck_dim": 256,
+        "in_context_len": 0,
+        "in_context_start": 0,
+        "attention_dropout": 0.0,
+        "dropout": 0.0,
+        "use_bottleneck": False,
+    },
+}
+
+
+def remap_legacy_state_dict(state_dict: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
+    """Map wrapper/backbone keys from legacy Hub/Lightning checkpoints to native JiTTransformer2DModel keys."""
+    remapped: Dict[str, torch.Tensor] = {}
+    for key, value in state_dict.items():
+        new_key = key
+        for prefix in ("transformer.", "net.", "denoiser.", "ema_denoiser."):
+            if new_key.startswith(prefix):
+                new_key = new_key[len(prefix) :]
+                break
+
+        new_key = new_key.replace(".adaLN_modulation.1.", ".adaLN_modulation.")
+        if new_key.startswith("final_layer."):
+            new_key = new_key.replace("final_layer.norm_final", "norm_final")
+            new_key = new_key.replace("final_layer.linear", "linear_final")
+            new_key = new_key.replace("final_layer.adaLN_modulation", "adaLN_modulation_final")
+
+        remapped[new_key] = value
+    return remapped
+
+
+def config_from_legacy(config: Dict[str, object]) -> Dict[str, object]:
+    """Build native config kwargs from a legacy config.json dict."""
+    model_type = config.get("model_type") or config.get("model_name")
+    if model_type not in JIT_PRESET_CONFIGS:
+        raise ValueError(f"Unknown JiT preset '{model_type}'. Known: {list(JIT_PRESET_CONFIGS)}")
+
+    preset = dict(JIT_PRESET_CONFIGS[model_type])
+    preset["num_classes"] = int(config.get("num_class_embeds") or config.get("num_classes") or 1000)
+
+    if config.get("attention_dropout") is not None:
+        preset["attention_dropout"] = float(config["attention_dropout"])
+    if config.get("dropout") is not None:
+        preset["dropout"] = float(config["dropout"])
+    if config.get("sample_size") is not None:
+        preset["sample_size"] = int(config["sample_size"])
+
+    return preset
+
+
+def broadcat(tensors, dim=-1):
+    num_tensors = len(tensors)
+    shape_lens = {len(t.shape) for t in tensors}
+    if len(shape_lens) != 1:
+        raise ValueError("tensors must all have the same number of dimensions")
+    shape_len = list(shape_lens)[0]
+    dim = (dim + shape_len) if dim < 0 else dim
+    dims = list(zip(*(list(t.shape) for t in tensors)))
+    expandable_dims = [(i, val) for i, val in enumerate(dims) if i != dim]
+    if not all(len(set(t[1])) <= 2 for t in expandable_dims):
+        raise ValueError("invalid dimensions for broadcastable concatenation")
+    max_dims = [(t[0], max(t[1])) for t in expandable_dims]
+    expanded_dims = [(t[0], (t[1],) * num_tensors) for t in max_dims]
+    expanded_dims.insert(dim, (dim, dims[dim]))
+    expandable_shapes = list(zip(*(t[1] for t in expanded_dims)))
+    tensors = [t[0].expand(*t[1]) for t in zip(tensors, expandable_shapes)]
+    return torch.cat(tensors, dim=dim)
+
+
+def rotate_half(x):
+    x = x.view(*x.shape[:-1], x.shape[-1] // 2, 2)
+    x1, x2 = x.unbind(dim=-1)
+    x = torch.stack((-x2, x1), dim=-1)
+    return x.view(*x.shape[:-2], -1)
+
+
+class JiTRotaryEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim,
+        pt_seq_len=16,
+        ft_seq_len=None,
+        custom_freqs=None,
+        theta=10000,
+        num_cls_token=0,
+    ):
+        super().__init__()
+        if custom_freqs is not None:
+            freqs = custom_freqs
+        else:
+            freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
+
+        if ft_seq_len is None:
+            ft_seq_len = pt_seq_len
+        t = torch.arange(ft_seq_len) / ft_seq_len * pt_seq_len
+
+        freqs = torch.einsum("..., f -> ... f", t, freqs)
+        freqs = freqs.repeat_interleave(2, dim=-1)
+        freqs = broadcat((freqs[:, None, :], freqs[None, :, :]), dim=-1)
+
+        if num_cls_token > 0:
+            freqs_flat = freqs.view(-1, freqs.shape[-1])
+            cos_img = freqs_flat.cos()
+            sin_img = freqs_flat.sin()
+            _, dim_freq = cos_img.shape
+            cos_pad = torch.ones(num_cls_token, dim_freq, dtype=cos_img.dtype)
+            sin_pad = torch.zeros(num_cls_token, dim_freq, dtype=sin_img.dtype)
+            self.register_buffer("freqs_cos", torch.cat([cos_pad, cos_img], dim=0), persistent=False)
+            self.register_buffer("freqs_sin", torch.cat([sin_pad, sin_img], dim=0), persistent=False)
+        else:
+            self.register_buffer("freqs_cos", freqs.cos().view(-1, freqs.shape[-1]), persistent=False)
+            self.register_buffer("freqs_sin", freqs.sin().view(-1, freqs.shape[-1]), persistent=False)
+
+    def forward(self, tensor):
+        seq_len = tensor.shape[1]
+        freqs_cos = self.freqs_cos[:seq_len].to(device=tensor.device, dtype=tensor.dtype)
+        freqs_sin = self.freqs_sin[:seq_len].to(device=tensor.device, dtype=tensor.dtype)
+        return tensor * freqs_cos[:, None, :] + rotate_half(tensor) * freqs_sin[:, None, :]
+
+
+def modulate(x, shift, scale):
+    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+
+
+class JiTPatchEmbed(nn.Module):
+    def __init__(
+        self,
+        img_size=224,
+        patch_size=16,
+        in_chans=3,
+        pca_dim=768,
+        embed_dim=768,
+        bias=True,
+        use_bottleneck=True,
+    ):
+        super().__init__()
+        img_size = (img_size, img_size)
+        patch_size = (patch_size, patch_size)
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
+        self.use_bottleneck = use_bottleneck
+        if use_bottleneck:
+            self.proj1 = nn.Conv2d(in_chans, pca_dim, kernel_size=patch_size, stride=patch_size, bias=False)
+            self.proj2 = nn.Conv2d(pca_dim, embed_dim, kernel_size=1, stride=1, bias=bias)
+        else:
+            self.proj1 = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias)
+            self.proj2 = None
+
+    def forward(self, x):
+        if self.use_bottleneck:
+            return self.proj2(self.proj1(x)).flatten(2).transpose(1, 2)
+        return self.proj1(x).flatten(2).transpose(1, 2)
+
+
+class JiTTimestepEmbedder(nn.Module):
+    def __init__(self, hidden_size, frequency_embedding_size=256):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
+            nn.SiLU(),
+            nn.Linear(hidden_size, hidden_size, bias=True),
+        )
+        self.frequency_embedding_size = frequency_embedding_size
+
+    @staticmethod
+    def timestep_embedding(t, dim, max_period=10000):
+        half = dim // 2
+        freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
+            device=t.device
+        )
+        args = t[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+        return embedding
+
+    def forward(self, t, dtype=None):
+        t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
+        if dtype is not None:
+            t_freq = t_freq.to(dtype=dtype)
+        return self.mlp(t_freq)
+
+
+class JiTLabelEmbedder(nn.Module):
+    def __init__(self, num_classes, hidden_size):
+        super().__init__()
+        self.embedding_table = nn.Embedding(num_classes + 1, hidden_size)
+        self.num_classes = num_classes
+
+    def forward(self, labels):
+        return self.embedding_table(labels)
+
+
+class JiTAttention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=True, qk_norm=True, attn_drop=0.0, proj_drop=0.0, eps=1e-6):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.q_norm = RMSNorm(head_dim, eps=eps) if qk_norm else nn.Identity()
+        self.k_norm = RMSNorm(head_dim, eps=eps) if qk_norm else nn.Identity()
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = attn_drop
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x, rope=None):
+        batch_size, num_tokens, channels = x.shape
+        qkv = self.qkv(x).reshape(batch_size, num_tokens, 3, self.num_heads, channels // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+        if rope is not None:
+            q = q.transpose(1, 2)
+            k = k.transpose(1, 2)
+            q = rope(q)
+            k = rope(k)
+            q = q.transpose(1, 2)
+            k = k.transpose(1, 2)
+
+        dropout_p = self.attn_drop if self.training else 0.0
+        x = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+        x = x.transpose(1, 2).reshape(batch_size, num_tokens, channels)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class JiTSwiGLUFFN(nn.Module):
+    def __init__(self, dim: int, hidden_dim: int, drop=0.0, bias=True) -> None:
+        super().__init__()
+        hidden_dim = int(hidden_dim * 2 / 3)
+        self.w12 = nn.Linear(dim, 2 * hidden_dim, bias=bias)
+        self.w3 = nn.Linear(hidden_dim, dim, bias=bias)
+        self.ffn_dropout = nn.Dropout(drop)
+
+    def forward(self, x):
+        x12 = self.w12(x)
+        x1, x2 = x12.chunk(2, dim=-1)
+        hidden = F.silu(x1) * x2
+        return self.w3(self.ffn_dropout(hidden))
+
+
+class JiTBlock(nn.Module):
+    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, attn_drop=0.0, proj_drop=0.0, eps=1e-6):
+        super().__init__()
+        self.norm1 = RMSNorm(hidden_size, eps=eps)
+        self.attn = JiTAttention(
+            hidden_size,
+            num_heads=num_heads,
+            qkv_bias=True,
+            qk_norm=True,
+            attn_drop=attn_drop,
+            proj_drop=proj_drop,
+            eps=eps,
+        )
+        self.norm2 = RMSNorm(hidden_size, eps=eps)
+        self.mlp = JiTSwiGLUFFN(hidden_size, int(hidden_size * mlp_ratio), drop=proj_drop)
+        self.act = nn.SiLU()
+        self.adaLN_modulation = nn.Linear(hidden_size, 6 * hidden_size, bias=True)
+
+    def forward(self, x, c, feat_rope=None):
+        c = self.act(c)
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=-1)
+        x = x + gate_msa.unsqueeze(1) * self.attn(modulate(self.norm1(x), shift_msa, scale_msa), rope=feat_rope)
+        x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
+        return x
+
+
+def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0):
+    grid_h = np.arange(grid_size, dtype=np.float32)
+    grid_w = np.arange(grid_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)
+    grid = np.stack(grid, axis=0)
+    grid = grid.reshape([2, 1, grid_size, grid_size])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if cls_token and extra_tokens > 0:
+        pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
+    return pos_embed
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+    if embed_dim % 2 != 0:
+        raise ValueError(f"embed_dim must be divisible by 2, but got {embed_dim}")
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])
+    return np.concatenate([emb_h, emb_w], axis=1)
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    if embed_dim % 2 != 0:
+        raise ValueError(f"embed_dim must be divisible by 2, but got {embed_dim}")
+    omega = np.arange(embed_dim // 2, dtype=np.float64)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega
+    pos = pos.reshape(-1)
+    out = np.einsum("m,d->md", pos, omega)
+    return np.concatenate([np.sin(out), np.cos(out)], axis=1)
+
+
+class JiTTransformer2DModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+    _skip_layerwise_casting_patterns = ["pos_embed", "norm"]
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: int = 256,
+        patch_size: int = 16,
+        in_channels: int = 3,
+        hidden_size: int = 768,
+        num_layers: int = 12,
+        num_attention_heads: int = 12,
+        mlp_ratio: float = 4.0,
+        attention_dropout: float = 0.0,
+        dropout: float = 0.0,
+        num_classes: int = 1000,
+        bottleneck_dim: int = 128,
+        in_context_len: int = 32,
+        in_context_start: int = 4,
+        norm_eps: float = 1e-6,
+        model_type: str | None = None,
+        num_class_embeds: int | None = None,
+        use_bottleneck: bool = True,
+    ):
+        super().__init__()
+        if num_class_embeds is not None:
+            num_classes = int(num_class_embeds)
+        if model_type in JIT_PRESET_CONFIGS:
+            preset = JIT_PRESET_CONFIGS[model_type]
+            sample_size = int(preset["sample_size"])
+            patch_size = int(preset["patch_size"])
+            hidden_size = int(preset["hidden_size"])
+            num_layers = int(preset["num_layers"])
+            num_attention_heads = int(preset["num_attention_heads"])
+            bottleneck_dim = int(preset["bottleneck_dim"])
+            in_context_len = int(preset["in_context_len"])
+            in_context_start = int(preset["in_context_start"])
+            if attention_dropout == 0.0:
+                attention_dropout = float(preset["attention_dropout"])
+            if dropout == 0.0:
+                dropout = float(preset["dropout"])
+            if "use_bottleneck" in preset:
+                use_bottleneck = bool(preset["use_bottleneck"])
+
+        self.sample_size = sample_size
+        self.patch_size = patch_size
+        self.in_channels = in_channels
+        self.out_channels = in_channels
+        self.hidden_size = hidden_size
+        self.num_layers = num_layers
+        self.num_attention_heads = num_attention_heads
+        self.in_context_len = in_context_len
+        self.in_context_start = in_context_start
+        self.norm_eps = norm_eps
+        self.gradient_checkpointing = False
+
+        self.t_embedder = JiTTimestepEmbedder(hidden_size)
+        self.y_embedder = JiTLabelEmbedder(num_classes, hidden_size)
+        self.use_bottleneck = use_bottleneck
+        self.x_embedder = JiTPatchEmbed(
+            img_size=sample_size,
+            patch_size=patch_size,
+            in_chans=in_channels,
+            pca_dim=bottleneck_dim,
+            embed_dim=hidden_size,
+            bias=True,
+            use_bottleneck=use_bottleneck,
+        )
+
+        num_patches = self.x_embedder.num_patches
+        pos_embed = get_2d_sincos_pos_embed(hidden_size, int(num_patches**0.5))
+        self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=True)
+
+        if self.in_context_len > 0:
+            self.in_context_posemb = nn.Parameter(torch.zeros(1, self.in_context_len, hidden_size))
+
+        half_head_dim = hidden_size // num_attention_heads // 2
+        hw_seq_len = sample_size // patch_size
+        self.feat_rope = JiTRotaryEmbedding(dim=half_head_dim, pt_seq_len=hw_seq_len, num_cls_token=0)
+        self.feat_rope_incontext = JiTRotaryEmbedding(
+            dim=half_head_dim,
+            pt_seq_len=hw_seq_len,
+            num_cls_token=self.in_context_len,
+        )
+
+        self.blocks = nn.ModuleList(
+            [
+                JiTBlock(
+                    hidden_size,
+                    num_attention_heads,
+                    mlp_ratio=mlp_ratio,
+                    attn_drop=attention_dropout if (num_layers // 4 * 3 > i >= num_layers // 4) else 0.0,
+                    proj_drop=dropout if (num_layers // 4 * 3 > i >= num_layers // 4) else 0.0,
+                    eps=norm_eps,
+                )
+                for i in range(num_layers)
+            ]
+        )
+
+        self.norm_final = RMSNorm(hidden_size, eps=norm_eps)
+        self.linear_final = nn.Linear(hidden_size, patch_size * patch_size * self.out_channels, bias=True)
+        self.act_final = nn.SiLU()
+        self.adaLN_modulation_final = nn.Linear(hidden_size, 2 * hidden_size, bias=True)
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        timestep: torch.LongTensor,
+        class_labels: torch.LongTensor,
+        return_dict: bool = True,
+    ):
+        timestep = torch.as_tensor(timestep, device=sample.device)
+        if timestep.ndim == 0:
+            timestep = timestep.repeat(sample.shape[0])
+        else:
+            timestep = timestep.reshape(-1)
+            if timestep.shape[0] == 1 and sample.shape[0] > 1:
+                timestep = timestep.repeat(sample.shape[0])
+
+        t_emb = self.t_embedder(timestep, dtype=sample.dtype)
+        y_emb = self.y_embedder(class_labels).to(dtype=sample.dtype)
+        c = t_emb + y_emb
+
+        x = self.x_embedder(sample)
+        x = x + self.pos_embed.to(x.dtype)
+
+        for i, block in enumerate(self.blocks):
+            if self.in_context_len > 0 and i == self.in_context_start:
+                in_context_tokens = y_emb.unsqueeze(1).repeat(1, self.in_context_len, 1)
+                in_context_tokens = in_context_tokens + self.in_context_posemb.to(in_context_tokens.dtype)
+                x = torch.cat([in_context_tokens, x], dim=1)
+
+            rope = self.feat_rope if i < self.in_context_start else self.feat_rope_incontext
+            if self.training and self.gradient_checkpointing:
+                x = torch.utils.checkpoint.checkpoint(block, x, c, rope, use_reentrant=False)
+            else:
+                x = block(x, c, feat_rope=rope)
+
+        if self.in_context_len > 0:
+            x = x[:, self.in_context_len :]
+
+        c = self.act_final(c)
+        shift, scale = self.adaLN_modulation_final(c).chunk(2, dim=1)
+        x = modulate(self.norm_final(x), shift, scale)
+        x = self.linear_final(x)
+
+        height = width = int(x.shape[1] ** 0.5)
+        x = x.reshape(shape=(x.shape[0], height, width, self.patch_size, self.patch_size, self.out_channels))
+        x = torch.einsum("nhwpqc->nchpwq", x)
+        output = x.reshape(shape=(x.shape[0], self.out_channels, height * self.patch_size, width * self.patch_size))
+
+        if not return_dict:
+            return (output,)
+        return Transformer2DModelOutput(sample=output)
+
+    @classmethod
+    def from_jit_checkpoint(
+        cls,
+        checkpoint_path: str,
+        weights: Literal["model", "ema1", "ema2"] = "ema1",
+        map_location: str = "cpu",
+        strict: bool = True,
+    ) -> Tuple["JiTTransformer2DModel", Dict[str, object]]:
+        checkpoint = torch.load(checkpoint_path, map_location=map_location, weights_only=False)
+        if "args" not in checkpoint:
+            raise ValueError("Checkpoint is missing 'args', cannot infer JiT architecture config.")
+
+        ckpt_args = checkpoint["args"]
+        if isinstance(ckpt_args, argparse.Namespace):
+            args_dict = vars(ckpt_args)
+        elif isinstance(ckpt_args, Mapping):
+            args_dict = dict(ckpt_args)
+        else:
+            raise TypeError(f"Unsupported checkpoint args type: {type(ckpt_args)}")
+
+        model_type = args_dict.get("model") or args_dict.get("model_name") or args_dict.get("model_type")
+        if model_type not in JIT_PRESET_CONFIGS:
+            raise ValueError(f"Unknown JiT preset '{model_type}'.")
+
+        config = dict(JIT_PRESET_CONFIGS[model_type])
+        config["num_classes"] = int(args_dict.get("class_num") or args_dict.get("num_classes") or 1000)
+        config["model_type"] = model_type
+        config["attention_dropout"] = float(args_dict.get("attn_dropout", args_dict.get("attention_dropout", config["attention_dropout"])))
+        config["dropout"] = float(args_dict.get("proj_dropout", args_dict.get("dropout", config["dropout"])))
+        model = cls(**config)
+
+        key = "model" if weights == "model" else f"model_{weights}"
+        if key not in checkpoint:
+            raise ValueError(f"Checkpoint key '{key}' not found. Available keys: {list(checkpoint.keys())}")
+
+        state_dict = remap_legacy_state_dict(checkpoint[key])
+        model.load_state_dict(state_dict, strict=strict)
+
+        metadata = {
+            "checkpoint_path": checkpoint_path,
+            "weights": weights,
+            "epoch": checkpoint.get("epoch"),
+            "model_type": model_type,
+            "source_args": checkpoint.get("args"),
+        }
+        return model, metadata
+
+    @classmethod
+    def from_pixelgen_checkpoint(
+        cls,
+        checkpoint_path: str,
+        weights_prefix: str = "ema_denoiser.",
+        model_type: str | None = None,
+        map_location: str = "cpu",
+        strict: bool = False,
+    ) -> Tuple["JiTTransformer2DModel", Dict[str, object]]:
+        """Load from a PixelGen/DeCo Lightning checkpoint (.ckpt)."""
+        checkpoint = torch.load(checkpoint_path, map_location=map_location, weights_only=False)
+        state_dict = checkpoint.get("state_dict", checkpoint)
+
+        filtered = {k[len(weights_prefix) :]: v for k, v in state_dict.items() if k.startswith(weights_prefix)}
+        if not filtered:
+            raise ValueError(f"No parameters found with prefix '{weights_prefix}' in checkpoint.")
+
+        inferred_type = model_type
+        if inferred_type is None:
+            hidden_size = filtered.get("t_embedder.mlp.0.weight")
+            depth = sum(1 for k in filtered if k.startswith("blocks.") and k.endswith(".norm1.weight"))
+            if hidden_size is not None and depth == 28 and hidden_size.shape[0] == 1152:
+                inferred_type = "PixelGen-XL/16"
+            elif hidden_size is not None and depth == 16 and hidden_size.shape[0] == 1536:
+                inferred_type = "PixelGen-XXL/16"
+
+        if inferred_type is None or inferred_type not in JIT_PRESET_CONFIGS:
+            raise ValueError(
+                f"Could not infer model_type from checkpoint. Pass model_type explicitly. "
+                f"Known presets: {list(JIT_PRESET_CONFIGS)}"
+            )
+
+        config = dict(JIT_PRESET_CONFIGS[inferred_type])
+        config["num_classes"] = 1000
+        config["model_type"] = inferred_type
+        model = cls(**config)
+        remapped = remap_legacy_state_dict(filtered)
+        model.load_state_dict(remapped, strict=strict)
+        metadata = {"checkpoint_path": checkpoint_path, "model_type": inferred_type, "weights_prefix": weights_prefix}
+        return model, metadata
+
+    def to_jit_checkpoint(
+        self,
+        ema_mode: Literal["none", "copy_to_both"] = "copy_to_both",
+        prefix: str = "net.",
+    ) -> Dict[str, object]:
+        base_state: Dict[str, torch.Tensor] = {}
+        for key, value in self.state_dict().items():
+            legacy_key = key
+            if legacy_key.startswith("norm_final"):
+                legacy_key = legacy_key.replace("norm_final", "final_layer.norm_final", 1)
+            if legacy_key.startswith("linear_final"):
+                legacy_key = legacy_key.replace("linear_final", "final_layer.linear", 1)
+            if legacy_key.startswith("adaLN_modulation_final"):
+                legacy_key = legacy_key.replace("adaLN_modulation_final", "final_layer.adaLN_modulation", 1)
+            legacy_key = legacy_key.replace(".adaLN_modulation.", ".adaLN_modulation.1.")
+            base_state[f"{prefix}{legacy_key}"] = value.detach().cpu()
+
+        checkpoint = {"model": base_state}
+        if ema_mode == "copy_to_both":
+            checkpoint["model_ema1"] = {k: v.clone() for k, v in base_state.items()}
+            checkpoint["model_ema2"] = {k: v.clone() for k, v in base_state.items()}
+        elif ema_mode != "none":
+            raise ValueError(f"Unsupported ema_mode='{ema_mode}'.")
+        return checkpoint
+
+    @property
+    def net(self):
+        return self
+
+
+JiTDiffusersModel = JiTTransformer2DModel

PixelGen-XXL-16-512-t2i/model_index.json

@@ -0,0 +1,23 @@
+{
+  "_class_name": [
+    "pipeline",
+    "PixelGenT2IPipeline"
+  ],
+  "_diffusers_version": "0.38.0",
+  "scheduler": [
+    "scheduling_pixelgen",
+    "PixelGenScheduler"
+  ],
+  "text_encoder": [
+    "transformers",
+    "Qwen3Model"
+  ],
+  "tokenizer": [
+    "transformers",
+    "Qwen2Tokenizer"
+  ],
+  "transformer": [
+    "transformer_jit_t2i",
+    "JiTTransformer2DT2IModel"
+  ]
+}

PixelGen-XXL-16-512-t2i/pipeline.py

@@ -0,0 +1,333 @@
+"""Hub custom pipeline: PixelGenT2IPipeline.
+Load with native Hugging Face diffusers and trust_remote_code=True.
+"""
+
+from __future__ import annotations
+
+import importlib
+import sys
+from pathlib import Path
+from typing import List, Optional, Tuple, Union
+
+import torch
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from diffusers.utils.torch_utils import randn_tensor
+from transformers import Qwen2Tokenizer, Qwen3Model
+
+
+class PixelGenT2IPipeline(DiffusionPipeline):
+    model_cpu_offload_seq = "text_encoder->transformer"
+
+    @staticmethod
+    def _infer_variant_dir(*components) -> Path:
+        for component in components:
+            if component is None:
+                continue
+            for attr in ("name_or_path",):
+                path_value = getattr(component, attr, None)
+                if path_value:
+                    path = Path(path_value)
+                    if path.is_dir() and (path / "model_index.json").exists():
+                        return path
+                    if path.parent.is_dir() and (path.parent / "model_index.json").exists():
+                        return path.parent
+            config = getattr(component, "config", None)
+            path_value = getattr(config, "_name_or_path", None) if config is not None else None
+            if path_value:
+                path = Path(path_value)
+                if path.is_dir() and (path / "model_index.json").exists():
+                    return path
+                if path.parent.is_dir() and (path.parent / "model_index.json").exists():
+                    return path.parent
+        return Path(__file__).resolve().parent
+
+    def __init__(
+        self,
+        transformer,
+        text_encoder,
+        tokenizer,
+        scheduler=None,
+    ):
+        variant = self._infer_variant_dir(transformer, text_encoder, tokenizer, scheduler)
+        if isinstance(transformer, (list, tuple)):
+            transformer = self._load_transformer(variant)
+        if isinstance(scheduler, (list, tuple)) or scheduler is None:
+            scheduler = self._load_scheduler(variant)
+
+        super().__init__()
+        self.register_modules(
+            transformer=transformer,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            scheduler=scheduler,
+        )
+        if torch.cuda.is_available():
+            torch_dtype = getattr(transformer, "dtype", None)
+            if torch_dtype is not None:
+                self.to("cuda", torch_dtype=torch_dtype)
+            else:
+                self.to("cuda")
+
+    @staticmethod
+    def _resolve_variant_dir(pretrained_model_name_or_path: Optional[str]) -> Path:
+        repo_root = Path(__file__).resolve().parent
+        if pretrained_model_name_or_path in (None, "", "."):
+            return repo_root
+        variant = Path(pretrained_model_name_or_path)
+        if not variant.is_absolute():
+            candidate = (Path.cwd() / variant).resolve()
+            variant = candidate if candidate.exists() else (repo_root / variant).resolve()
+        return variant
+
+    @classmethod
+    def _load_transformer(cls, variant: Path, **model_kwargs):
+        comp_dir = variant / "transformer"
+        module_path = comp_dir / "transformer_jit_t2i.py"
+        if not module_path.exists():
+            raise FileNotFoundError(f"Missing transformer module at {module_path}")
+        comp_path = str(comp_dir)
+        if comp_path not in sys.path:
+            sys.path.insert(0, comp_path)
+        module = importlib.import_module("transformer_jit_t2i")
+        transformer_cls = getattr(module, "JiTTransformer2DT2IModel")
+        return transformer_cls.from_pretrained(str(comp_dir), **model_kwargs)
+
+    @classmethod
+    def _load_scheduler(cls, variant: Path):
+        scheduler_dir = variant / "scheduler"
+        config_path = scheduler_dir / "scheduler_config.json"
+        if not config_path.exists():
+            raise FileNotFoundError(f"Missing scheduler config at {config_path}")
+        scheduler_path = str(scheduler_dir)
+        if scheduler_path not in sys.path:
+            sys.path.insert(0, scheduler_path)
+        module = importlib.import_module("scheduling_pixelgen")
+        scheduler_cls = getattr(module, "PixelGenScheduler")
+        return scheduler_cls.from_pretrained(str(scheduler_dir))
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
+        variant = cls._resolve_variant_dir(pretrained_model_name_or_path)
+        model_kwargs = dict(kwargs)
+        transformer = cls._load_transformer(variant, **model_kwargs)
+        scheduler = cls._load_scheduler(variant)
+
+        text_encoder_dir = variant / "text_encoder"
+        tokenizer_dir = variant / "tokenizer"
+        if text_encoder_dir.exists() and (text_encoder_dir / "config.json").exists():
+            text_encoder = Qwen3Model.from_pretrained(str(text_encoder_dir), **model_kwargs)
+            tokenizer = Qwen2Tokenizer.from_pretrained(
+                str(tokenizer_dir if tokenizer_dir.exists() else text_encoder_dir)
+            )
+        else:
+            text_encoder_name = "Qwen/Qwen3-1.7B"
+            text_encoder = Qwen3Model.from_pretrained(text_encoder_name, **model_kwargs)
+            tokenizer = Qwen2Tokenizer.from_pretrained(text_encoder_name)
+
+        pipe = cls(
+            transformer=transformer,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            scheduler=scheduler,
+        )
+        if hasattr(pipe, "register_to_config"):
+            pipe.register_to_config(_name_or_path=str(variant))
+        return pipe
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Union[str, List[str], None] = None,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        if isinstance(prompt, str):
+            prompt = [prompt]
+        if negative_prompt is None:
+            negative_prompt = [""] * len(prompt)
+        elif isinstance(negative_prompt, str):
+            negative_prompt = [negative_prompt] * len(prompt)
+
+        max_length = self.transformer.config.text_max_length
+        text_embed_dim = self.transformer.config.text_embed_dim
+
+        def _encode(texts):
+            tokenized = self.tokenizer(
+                texts, truncation=True, max_length=max_length, padding="max_length", return_tensors="pt"
+            )
+            input_ids = tokenized.input_ids.to(device)
+            attention_mask = tokenized.attention_mask.to(device)
+            hidden = self.text_encoder(input_ids=input_ids, attention_mask=attention_mask)[0]
+            if hidden.shape[2] < text_embed_dim:
+                pad = torch.zeros(
+                    hidden.shape[0],
+                    hidden.shape[1],
+                    text_embed_dim - hidden.shape[2],
+                    device=device,
+                    dtype=hidden.dtype,
+                )
+                hidden = torch.cat([hidden, pad], dim=-1)
+            elif hidden.shape[2] > text_embed_dim:
+                hidden = hidden[:, :, :text_embed_dim]
+            return hidden.to(dtype=dtype)
+
+        return _encode(prompt), _encode(negative_prompt)
+
+    def _predict_velocity(
+        self,
+        z_value: torch.Tensor,
+        t: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        uncond_hidden_states: torch.Tensor,
+        do_classifier_free_guidance: bool,
+        guidance_scale: float,
+        guidance_interval_min: float,
+        guidance_interval_max: float,
+        scale_h: float,
+        scale_w: float,
+    ) -> torch.Tensor:
+        if do_classifier_free_guidance:
+            z_in = torch.cat([z_value, z_value], dim=0)
+            hidden = torch.cat([uncond_hidden_states, encoder_hidden_states], dim=0)
+        else:
+            z_in = z_value
+            hidden = encoder_hidden_states
+
+        t = t.to(device=z_value.device, dtype=torch.float64)
+        t_batch = t.flatten().expand(z_in.shape[0]).to(dtype=z_in.dtype)
+        x_pred = self.transformer(
+            z_in, timestep=t_batch, encoder_hidden_states=hidden, scale_h=scale_h, scale_w=scale_w
+        ).sample
+        v = self.scheduler.velocity_from_prediction(z_in, x_pred, t)
+
+        if not do_classifier_free_guidance:
+            return v
+
+        v_uncond, v_cond = v.chunk(2, dim=0)
+        interval_mask = (t > guidance_interval_min) & (t < guidance_interval_max)
+        scale = torch.where(
+            interval_mask,
+            torch.tensor(guidance_scale, device=z_value.device, dtype=z_value.dtype),
+            torch.tensor(1.0, device=z_value.device, dtype=z_value.dtype),
+        )
+        return v_uncond + scale * (v_cond - v_uncond)
+
+    def _run_adam_lm_sampler(
+        self,
+        latents: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        uncond_hidden_states: torch.Tensor,
+        num_inference_steps: int,
+        do_classifier_free_guidance: bool,
+        guidance_scale: float,
+        guidance_interval_min: float,
+        guidance_interval_max: float,
+        timeshift: float,
+        adam_lm_order: int,
+        scale_h: float,
+        scale_w: float,
+    ) -> torch.Tensor:
+        device = latents.device
+        self.scheduler.set_timesteps(
+            num_inference_steps,
+            device=device,
+            solver="adam_lm",
+            timeshift=timeshift,
+            adam_lm_order=adam_lm_order,
+        )
+        batch_size = latents.shape[0]
+        t_cur = torch.zeros(batch_size, device=device, dtype=torch.float64)
+        velocity_history: list[torch.Tensor] = []
+
+        for i in self.progress_bar(range(num_inference_steps)):
+            v = self._predict_velocity(
+                latents,
+                t_cur,
+                encoder_hidden_states,
+                uncond_hidden_states,
+                do_classifier_free_guidance,
+                guidance_scale,
+                guidance_interval_min,
+                guidance_interval_max,
+                scale_h,
+                scale_w,
+            )
+            velocity_history.append(v)
+            latents = self.scheduler.apply_adam_lm_step(latents, velocity_history, i)
+            t_cur = t_cur + self.scheduler.timedeltas[i].to(dtype=torch.float64)
+
+        return latents
+
+    @torch.inference_mode()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Union[str, List[str], None] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        guidance_scale: float = 4.0,
+        guidance_interval_min: float = 0.0,
+        guidance_interval_max: float = 1.0,
+        num_inference_steps: int = 25,
+        timeshift: Optional[float] = None,
+        adam_lm_order: Optional[int] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+    ) -> Union[ImagePipelineOutput, Tuple]:
+        if isinstance(prompt, str):
+            prompt = [prompt]
+
+        height = height or self.transformer.config.sample_size
+        width = width or self.transformer.config.sample_size
+        height = height // 32 * 32
+        width = width // 32 * 32
+        scale_h = height / self.transformer.config.sample_size
+        scale_w = width / self.transformer.config.sample_size
+        shift = timeshift if timeshift is not None else self.scheduler.config.timeshift
+        order = adam_lm_order if adam_lm_order is not None else self.scheduler.config.adam_lm_order
+
+        cond, uncond = self.encode_prompt(prompt, negative_prompt)
+        do_cfg = guidance_scale > 1.0
+        batch_size = len(prompt)
+
+        latents = randn_tensor(
+            shape=(batch_size, self.transformer.config.in_channels, height, width),
+            generator=generator,
+            device=self._execution_device,
+            dtype=self.transformer.dtype,
+        )
+
+        latents = self._run_adam_lm_sampler(
+            latents,
+            cond,
+            uncond,
+            num_inference_steps,
+            do_cfg,
+            guidance_scale,
+            guidance_interval_min,
+            guidance_interval_max,
+            shift,
+            order,
+            scale_h,
+            scale_w,
+        )
+
+        images_pt = ((latents.float().clamp(-1, 1) + 1.0) / 2.0).cpu()
+        if output_type == "pt":
+            images = images_pt
+        elif output_type == "np":
+            images = images_pt.permute(0, 2, 3, 1).numpy()
+        else:
+            images = self.numpy_to_pil(images_pt.permute(0, 2, 3, 1).numpy())
+
+        self.maybe_free_model_hooks()
+        if not return_dict:
+            return (images,)
+        return ImagePipelineOutput(images=images)
+
+
+PixelGenT2IPipelineOutput = ImagePipelineOutput

PixelGen-XXL-16-512-t2i/scheduler/scheduler_config.json

@@ -0,0 +1,10 @@
+{
+  "_class_name": "PixelGenScheduler",
+  "_diffusers_version": "0.38.0",
+  "adam_lm_order": 2,
+  "last_step": null,
+  "num_train_timesteps": 1000,
+  "solver": "adam_lm",
+  "t_eps": 0.05,
+  "timeshift": 3.0
+}

PixelGen-XXL-16-512-t2i/scheduler/scheduling_pixelgen.py

@@ -0,0 +1,285 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+from diffusers.utils import BaseOutput
+
+
+def lagrange_preint_o1(t1, v1, int_t_start, int_t_end):
+    int1 = int_t_end - int_t_start
+    return int1 * v1, (int1 / int1,)
+
+
+def lagrange_preint_o2(t1, t2, v1, v2, int_t_start, int_t_end):
+    int1 = 0.5 / (t1 - t2) * ((int_t_end - t2) ** 2 - (int_t_start - t2) ** 2)
+    int2 = 0.5 / (t2 - t1) * ((int_t_end - t1) ** 2 - (int_t_start - t1) ** 2)
+    int_sum = int1 + int2
+    return int1 * v1 + int2 * v2, (int1 / int_sum, int2 / int_sum)
+
+
+def lagrange_preint_o3(t1, t2, t3, v1, v2, v3, int_t_start, int_t_end):
+    int1_denom = (t1 - t2) * (t1 - t3)
+    int1_end = 1 / 3 * (int_t_end) ** 3 - 1 / 2 * (t2 + t3) * (int_t_end) ** 2 + (t2 * t3) * int_t_end
+    int1_start = 1 / 3 * (int_t_start) ** 3 - 1 / 2 * (t2 + t3) * (int_t_start) ** 2 + (t2 * t3) * int_t_start
+    int1 = (int1_end - int1_start) / int1_denom
+    int2_denom = (t2 - t1) * (t2 - t3)
+    int2_end = 1 / 3 * (int_t_end) ** 3 - 1 / 2 * (t1 + t3) * (int_t_end) ** 2 + (t1 * t3) * int_t_end
+    int2_start = 1 / 3 * (int_t_start) ** 3 - 1 / 2 * (t1 + t3) * (int_t_start) ** 2 + (t1 * t3) * int_t_start
+    int2 = (int2_end - int2_start) / int2_denom
+    int3_denom = (t3 - t1) * (t3 - t2)
+    int3_end = 1 / 3 * (int_t_end) ** 3 - 1 / 2 * (t1 + t2) * (int_t_end) ** 2 + (t1 * t2) * int_t_end
+    int3_start = 1 / 3 * (int_t_start) ** 3 - 1 / 2 * (t1 + t2) * (int_t_start) ** 2 + (t1 * t2) * int_t_start
+    int3 = (int3_end - int3_start) / int3_denom
+    int_sum = int1 + int2 + int3
+    return int1 * v1 + int2 * v2 + int3 * v3, (int1 / int_sum, int2 / int_sum, int3 / int_sum)
+
+
+def larange_preint_o4(t1, t2, t3, t4, v1, v2, v3, v4, int_t_start, int_t_end):
+    int1_denom = (t1 - t2) * (t1 - t3) * (t1 - t4)
+    int1_end = (
+        1 / 4 * (int_t_end) ** 4
+        - 1 / 3 * (t2 + t3 + t4) * (int_t_end) ** 3
+        + 1 / 2 * (t3 * t4 + t2 * t3 + t2 * t4) * int_t_end**2
+        - t2 * t3 * t4 * int_t_end
+    )
+    int1_start = (
+        1 / 4 * (int_t_start) ** 4
+        - 1 / 3 * (t2 + t3 + t4) * (int_t_start) ** 3
+        + 1 / 2 * (t3 * t4 + t2 * t3 + t2 * t4) * int_t_start**2
+        - t2 * t3 * t4 * int_t_start
+    )
+    int1 = (int1_end - int1_start) / int1_denom
+    int2_denom = (t2 - t1) * (t2 - t3) * (t2 - t4)
+    int2_end = (
+        1 / 4 * (int_t_end) ** 4
+        - 1 / 3 * (t1 + t3 + t4) * (int_t_end) ** 3
+        + 1 / 2 * (t3 * t4 + t1 * t3 + t1 * t4) * int_t_end**2
+        - t1 * t3 * t4 * int_t_end
+    )
+    int2_start = (
+        1 / 4 * (int_t_start) ** 4
+        - 1 / 3 * (t1 + t3 + t4) * (int_t_start) ** 3
+        + 1 / 2 * (t3 * t4 + t1 * t3 + t1 * t4) * int_t_start**2
+        - t1 * t3 * t4 * int_t_start
+    )
+    int2 = (int2_end - int2_start) / int2_denom
+    int3_denom = (t3 - t1) * (t3 - t2) * (t3 - t4)
+    int3_end = (
+        1 / 4 * (int_t_end) ** 4
+        - 1 / 3 * (t1 + t2 + t4) * (int_t_end) ** 3
+        + 1 / 2 * (t4 * t2 + t1 * t2 + t1 * t4) * int_t_end**2
+        - t1 * t2 * t4 * int_t_end
+    )
+    int3_start = (
+        1 / 4 * (int_t_start) ** 4
+        - 1 / 3 * (t1 + t2 + t4) * (int_t_start) ** 3
+        + 1 / 2 * (t4 * t2 + t1 * t2 + t1 * t4) * int_t_start**2
+        - t1 * t2 * t4 * int_t_start
+    )
+    int3 = (int3_end - int3_start) / int3_denom
+    int4_denom = (t4 - t1) * (t4 - t2) * (t4 - t3)
+    int4_end = (
+        1 / 4 * (int_t_end) ** 4
+        - 1 / 3 * (t1 + t2 + t3) * (int_t_end) ** 3
+        + 1 / 2 * (t3 * t2 + t1 * t2 + t1 * t3) * int_t_end**2
+        - t1 * t2 * t3 * int_t_end
+    )
+    int4_start = (
+        1 / 4 * (int_t_start) ** 4
+        - 1 / 3 * (t1 + t2 + t3) * (int_t_start) ** 3
+        + 1 / 2 * (t3 * t2 + t1 * t2 + t1 * t3) * int_t_start**2
+        - t1 * t2 * t3 * int_t_start
+    )
+    int4 = (int4_end - int4_start) / int4_denom
+    int_sum = int1 + int2 + int3 + int4
+    return int1 * v1 + int2 * v2 + int3 * v3 + int4 * v4, (
+        int1 / int_sum,
+        int2 / int_sum,
+        int3 / int_sum,
+        int4 / int_sum,
+    )
+
+
+def lagrange_preint(order, pre_vs, pre_ts, int_t_start, int_t_end):
+    order = min(order, len(pre_vs), len(pre_ts))
+    if order == 1:
+        return lagrange_preint_o1(pre_ts[-1], pre_vs[-1], int_t_start, int_t_end)
+    if order == 2:
+        return lagrange_preint_o2(pre_ts[-2], pre_ts[-1], pre_vs[-2], pre_vs[-1], int_t_start, int_t_end)
+    if order == 3:
+        return lagrange_preint_o3(
+            pre_ts[-3], pre_ts[-2], pre_ts[-1], pre_vs[-3], pre_vs[-2], pre_vs[-1], int_t_start, int_t_end
+        )
+    if order == 4:
+        return larange_preint_o4(
+            pre_ts[-4], pre_ts[-3], pre_ts[-2], pre_ts[-1],
+            pre_vs[-4], pre_vs[-3], pre_vs[-2], pre_vs[-1],
+            int_t_start, int_t_end,
+        )
+    raise ValueError("Invalid order")
+
+
+@dataclass
+class PixelGenSchedulerOutput(BaseOutput):
+    prev_sample: torch.Tensor
+
+
+def shift_respace_fn(t: torch.Tensor, shift: float = 1.0) -> torch.Tensor:
+    return t / (t + (1 - t) * shift)
+
+
+class PixelGenScheduler(SchedulerMixin, ConfigMixin):
+    """Flow-matching scheduler for PixelGen with Euler, Heun, and Adam-LM solvers."""
+
+    order = 2
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        t_eps: float = 5e-2,
+        solver: str = "heun",
+        timeshift: float = 1.0,
+        adam_lm_order: int = 2,
+        last_step: float | None = None,
+    ):
+        del num_train_timesteps
+        if solver not in {"heun", "euler", "adam_lm"}:
+            raise ValueError("solver must be one of: 'heun', 'euler', 'adam_lm'.")
+        self.timesteps: Optional[torch.Tensor] = None
+        self.timedeltas: Optional[torch.Tensor] = None
+        self.sigmas: Optional[List[float]] = None
+        self.num_inference_steps: Optional[int] = None
+        self._step_index: Optional[int] = None
+        self.solver_coeffs: Optional[list] = None
+        self._last_step = last_step
+
+    @property
+    def init_noise_sigma(self) -> float:
+        return 1.0
+
+    def set_timesteps(
+        self,
+        num_inference_steps: int,
+        device: Union[str, torch.device, None] = None,
+        solver: Optional[str] = None,
+        timeshift: Optional[float] = None,
+        adam_lm_order: Optional[int] = None,
+    ) -> None:
+        if num_inference_steps < 2:
+            raise ValueError("num_inference_steps must be >= 2.")
+
+        if solver is not None:
+            self.register_to_config(solver=solver)
+        if timeshift is not None:
+            self.register_to_config(timeshift=timeshift)
+        if adam_lm_order is not None:
+            self.register_to_config(adam_lm_order=adam_lm_order)
+
+        last_step = self._last_step
+        if last_step is None:
+            last_step = 1.0 / num_inference_steps
+
+        self.num_inference_steps = num_inference_steps
+        timesteps = torch.linspace(0.0, 1.0 - last_step, num_inference_steps, device=device, dtype=torch.float64)
+        timesteps = torch.cat([timesteps, torch.tensor([1.0], device=device, dtype=torch.float64)], dim=0)
+        self.timesteps = shift_respace_fn(timesteps, self.config.timeshift)
+        self.timedeltas = self.timesteps[1:] - self.timesteps[:-1]
+        sigma_grid = torch.linspace(0.0, 1.0, num_inference_steps, device=device, dtype=torch.float64)
+        self.sigmas = (1.0 - sigma_grid).tolist()
+        self._step_index = 0
+
+        if self.config.solver == "adam_lm":
+            self._reparameterize_coeffs()
+
+    def _reparameterize_coeffs(self) -> None:
+        solver_coeffs = [[] for _ in range(self.num_inference_steps)]
+        for i in range(self.num_inference_steps):
+            pre_vs = [1.0] * (i + 1)
+            pre_ts = self.timesteps[: i + 1]
+            int_t_start = self.timesteps[i]
+            int_t_end = self.timesteps[i + 1]
+            order = min(self.config.adam_lm_order, i + 1)
+            _, coeffs = lagrange_preint(order, pre_vs, pre_ts.tolist(), float(int_t_start), float(int_t_end))
+            solver_coeffs[i] = coeffs
+        self.solver_coeffs = solver_coeffs
+
+    def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor:
+        del timestep
+        return sample
+
+    def _resolve_step_index(self, timestep: Union[float, torch.Tensor, None]) -> int:
+        if self._step_index is not None:
+            return self._step_index
+        if self.timesteps is None:
+            raise ValueError("Call `set_timesteps` before `step`.")
+        if timestep is None:
+            return 0
+        t_value = float(timestep) if not isinstance(timestep, torch.Tensor) else float(timestep.flatten()[0])
+        matches = (self.timesteps - t_value).abs() < 1e-6
+        if matches.any():
+            return int(matches.nonzero(as_tuple=False)[0].item())
+        return 0
+
+    def step(
+        self,
+        model_output: torch.Tensor,
+        timestep: Union[float, torch.Tensor, None],
+        sample: torch.Tensor,
+        model_output_next: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[PixelGenSchedulerOutput, Tuple[torch.Tensor]]:
+        if self.timesteps is None:
+            raise ValueError("Call `set_timesteps` before `step`.")
+
+        step_index = self._resolve_step_index(timestep)
+        if step_index >= len(self.timesteps) - 1:
+            raise ValueError("Scheduler has already reached the final timestep.")
+
+        t = self.timesteps[step_index]
+        t_next = self.timesteps[step_index + 1]
+        dt = (t_next - t).to(dtype=sample.dtype)
+
+        if self.config.solver == "heun" and model_output_next is not None:
+            prev_sample = sample + dt * 0.5 * (model_output + model_output_next)
+        else:
+            prev_sample = sample + dt * model_output
+
+        self._step_index = step_index + 1
+
+        if not return_dict:
+            return (prev_sample,)
+        return PixelGenSchedulerOutput(prev_sample=prev_sample)
+
+    def velocity_from_prediction(
+        self,
+        sample: torch.Tensor,
+        x_pred: torch.Tensor,
+        timestep: Union[float, torch.Tensor],
+    ) -> torch.Tensor:
+        t = torch.as_tensor(timestep, device=sample.device, dtype=torch.float64)
+        while t.ndim < sample.ndim:
+            t = t.unsqueeze(-1)
+        denom = (1.0 - t).clamp_min(self.config.t_eps)
+        velocity = (x_pred.to(torch.float64) - sample.to(torch.float64)) / denom
+        return velocity.to(dtype=sample.dtype)
+
+    def apply_adam_lm_step(
+        self,
+        sample: torch.Tensor,
+        velocity_history: list[torch.Tensor],
+        step_index: int,
+    ) -> torch.Tensor:
+        if self.solver_coeffs is None:
+            raise ValueError("Adam-LM coefficients not initialized. Call set_timesteps with solver='adam_lm'.")
+        coeffs = self.solver_coeffs[step_index]
+        order = len(coeffs)
+        v = torch.zeros_like(sample, dtype=torch.float64)
+        for j, coeff in enumerate(coeffs):
+            v = v + coeff * velocity_history[-order:][j].to(torch.float64)
+        dt = self.timedeltas[step_index].to(device=sample.device, dtype=torch.float64)
+        return (sample.to(torch.float64) + dt * v).to(dtype=sample.dtype)

PixelGen-XXL-16-512-t2i/text_encoder/config.json

@@ -0,0 +1,63 @@
+{
+  "architectures": [
+    "Qwen3Model"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "dtype": "bfloat16",
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 6144,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 40960,
+  "max_window_layers": 28,
+  "model_type": "qwen3",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "pad_token_id": null,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "rope_theta": 1000000,
+    "rope_type": "default"
+  },
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.0.0",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

PixelGen-XXL-16-512-t2i/text_encoder/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fbbfc1bd7cdbaa8f0759d8d3c00d8a91bf9af50093fc3c7204c1e3e2c5fe3723
+size 3441183752

PixelGen-XXL-16-512-t2i/tokenizer/chat_template.jinja

@@ -0,0 +1,89 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- endif %}
+{%- endif %}

PixelGen-XXL-16-512-t2i/tokenizer/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650

PixelGen-XXL-16-512-t2i/tokenizer/tokenizer_config.json

@@ -0,0 +1,29 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "is_local": true,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

PixelGen-XXL-16-512-t2i/transformer/config.json

@@ -0,0 +1,16 @@
+{
+  "_class_name": "JiTTransformer2DT2IModel",
+  "_diffusers_version": "0.38.0",
+  "bottleneck_dim": 256,
+  "hidden_size": 1536,
+  "in_channels": 3,
+  "mlp_ratio": 4.0,
+  "model_type": "PixelGen-XXL/16-T2I",
+  "num_attention_heads": 24,
+  "num_layers": 16,
+  "num_text_layers": 4,
+  "patch_size": 16,
+  "sample_size": 512,
+  "text_embed_dim": 2048,
+  "text_max_length": 128
+}

PixelGen-XXL-16-512-t2i/transformer/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71fe40f4f443df27b5989048c151d4746f2b6f3c0b5552e7b24e46c6d4519887
+size 4493275072

PixelGen-XXL-16-512-t2i/transformer/transformer_jit_t2i.py

@@ -0,0 +1,390 @@
+from __future__ import annotations
+
+import math
+from functools import lru_cache
+from typing import Dict, Optional, Tuple
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import RMSNorm
+
+
+PIXELGEN_T2I_PRESET_CONFIGS: Dict[str, Dict[str, object]] = {
+    "PixelGen-XXL/16-T2I": {
+        "sample_size": 512,
+        "patch_size": 16,
+        "in_channels": 3,
+        "hidden_size": 1536,
+        "num_layers": 16,
+        "num_attention_heads": 24,
+        "num_text_layers": 4,
+        "text_embed_dim": 2048,
+        "text_max_length": 128,
+        "bottleneck_dim": 256,
+    },
+}
+
+
+def remap_t2i_state_dict(state_dict: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
+    remapped: Dict[str, torch.Tensor] = {}
+    for key, value in state_dict.items():
+        new_key = key
+        for prefix in ("transformer.", "denoiser.", "ema_denoiser."):
+            if new_key.startswith(prefix):
+                new_key = new_key[len(prefix) :]
+                break
+        new_key = new_key.replace(".adaLN_modulation.1.", ".adaLN_modulation.")
+        new_key = new_key.replace(".adaLN_modulation.0.", ".adaLN_modulation.")
+        remapped[new_key] = value
+    return remapped
+
+
+def modulate(x, shift, scale, unsqueeze=True):
+    if unsqueeze:
+        return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+    return x * (1 + scale) + shift
+
+
+def get_2d_sincos_pos_embed(embed_dim, grid_size):
+    grid_h = np.arange(grid_size, dtype=np.float32)
+    grid_w = np.arange(grid_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)
+    grid = np.stack(grid, axis=0)
+    grid = grid.reshape([2, 1, grid_size, grid_size])
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])
+    return np.concatenate([emb_h, emb_w], axis=1)
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    omega = np.arange(embed_dim // 2, dtype=np.float64)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega
+    pos = pos.reshape(-1)
+    out = np.einsum("m,d->md", pos, omega)
+    return np.concatenate([np.sin(out), np.cos(out)], axis=1)
+
+
+def precompute_freqs_cis_2d(dim: int, height: int, width: int, theta: float = 10000.0, scale=1.0):
+    if isinstance(scale, float):
+        scale = (scale, scale)
+    x_pos = torch.linspace(0, height * scale[0], width)
+    y_pos = torch.linspace(0, width * scale[1], height)
+    y_pos, x_pos = torch.meshgrid(y_pos, x_pos, indexing="ij")
+    y_pos = y_pos.reshape(-1)
+    x_pos = x_pos.reshape(-1)
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
+    x_freqs = torch.outer(x_pos, freqs).float()
+    y_freqs = torch.outer(y_pos, freqs).float()
+    x_cis = torch.polar(torch.ones_like(x_freqs), x_freqs)
+    y_cis = torch.polar(torch.ones_like(y_freqs), y_freqs)
+    freqs_cis = torch.cat([x_cis.unsqueeze(dim=-1), y_cis.unsqueeze(dim=-1)], dim=-1)
+    return freqs_cis.reshape(height * width, -1)
+
+
+def apply_rotary_emb(q, k, freqs_cis):
+    freqs_cis = freqs_cis[None, None, :, :]
+    q_ = torch.view_as_complex(q.float().reshape(*q.shape[:-1], -1, 2))
+    k_ = torch.view_as_complex(k.float().reshape(*k.shape[:-1], -1, 2))
+    q_out = torch.view_as_real(q_ * freqs_cis).flatten(3)
+    k_out = torch.view_as_real(k_ * freqs_cis).flatten(3)
+    return q_out.type_as(q), k_out.type_as(k)
+
+
+class JiTT2ITimestepEmbedder(nn.Module):
+    def __init__(self, hidden_size, frequency_embedding_size=256):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
+            nn.SiLU(),
+            nn.Linear(hidden_size, hidden_size, bias=True),
+        )
+        self.frequency_embedding_size = frequency_embedding_size
+
+    @staticmethod
+    def timestep_embedding(t, dim, max_period=10000):
+        half = dim // 2
+        freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
+            device=t.device
+        )
+        args = t[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+        return embedding
+
+    def forward(self, t, dtype=None):
+        t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
+        if dtype is not None:
+            t_freq = t_freq.to(dtype=dtype)
+        return self.mlp(t_freq)
+
+
+class JiTT2ITextEmbedder(nn.Module):
+    def __init__(self, text_embed_dim, hidden_size, bias=True):
+        super().__init__()
+        self.proj = nn.Linear(text_embed_dim, hidden_size, bias=bias)
+        self.norm = RMSNorm(hidden_size, eps=1e-6)
+
+    def forward(self, y):
+        return self.norm(self.proj(y))
+
+
+class JiTT2IBottleneckEmbed(nn.Module):
+    def __init__(self, in_chans, embed_dim, bottleneck_dim, bias=True):
+        super().__init__()
+        self.proj1 = nn.Linear(in_chans, bottleneck_dim, bias=False)
+        self.proj2 = nn.Linear(bottleneck_dim, embed_dim, bias=bias)
+
+    def forward(self, x):
+        return self.proj2(self.proj1(x))
+
+
+class JiTT2IAttention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0, eps=1e-6):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.qkv_x = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.kv_y = nn.Linear(dim, dim * 2, bias=qkv_bias)
+        self.q_norm = RMSNorm(head_dim, eps=eps)
+        self.k_norm = RMSNorm(head_dim, eps=eps)
+        self.attn_drop = attn_drop
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x, y, pos):
+        batch_size, num_tokens, channels = x.shape
+        qkv_x = self.qkv_x(x).reshape(batch_size, num_tokens, 3, self.num_heads, channels // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, kx, vx = qkv_x[0], qkv_x[1], qkv_x[2]
+        q = self.q_norm(q.contiguous())
+        kx = self.k_norm(kx.contiguous())
+        q, kx = apply_rotary_emb(q, kx, freqs_cis=pos)
+        kv_y = self.kv_y(y).reshape(batch_size, -1, 2, self.num_heads, channels // self.num_heads).permute(2, 0, 3, 1, 4)
+        ky, vy = kv_y[0], kv_y[1]
+        ky = self.k_norm(ky.contiguous())
+        k = torch.cat([kx, ky], dim=2)
+        v = torch.cat([vx, vy], dim=2)
+        dropout_p = self.attn_drop if self.training else 0.0
+        out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+        out = out.transpose(1, 2).reshape(batch_size, num_tokens, channels)
+        return self.proj_drop(self.proj(out))
+
+
+class JiTT2ITextRefineAttention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0, eps=1e-6):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.q_norm = RMSNorm(head_dim, eps=eps)
+        self.k_norm = RMSNorm(head_dim, eps=eps)
+        self.attn_drop = attn_drop
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x):
+        batch_size, num_tokens, channels = x.shape
+        qkv = self.qkv(x).reshape(batch_size, num_tokens, 3, self.num_heads, channels // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+        dropout_p = self.attn_drop if self.training else 0.0
+        out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+        out = out.transpose(1, 2).reshape(batch_size, num_tokens, channels)
+        return self.proj_drop(self.proj(out))
+
+
+class JiTT2ISwiGLU(nn.Module):
+    def __init__(self, dim, hidden_dim, drop=0.0, bias=False):
+        super().__init__()
+        self.w12 = nn.Linear(dim, 2 * hidden_dim, bias=bias)
+        self.w3 = nn.Linear(hidden_dim, dim, bias=bias)
+        self.ffn_dropout = nn.Dropout(drop)
+
+    def forward(self, x):
+        x12 = self.w12(x)
+        x1, x2 = x12.chunk(2, dim=-1)
+        return self.w3(self.ffn_dropout(F.silu(x1) * x2))
+
+
+class JiTT2IBlock(nn.Module):
+    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, eps=1e-6):
+        super().__init__()
+        self.norm1 = RMSNorm(hidden_size, eps=eps)
+        self.attn = JiTT2IAttention(hidden_size, num_heads=num_heads, eps=eps)
+        self.norm2 = RMSNorm(hidden_size, eps=eps)
+        self.mlp = JiTT2ISwiGLU(hidden_size, int(hidden_size * mlp_ratio))
+        self.adaLN_modulation = nn.Linear(hidden_size, 6 * hidden_size, bias=True)
+
+    def forward(self, x, y, c, pos):
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=-1)
+        x = x + gate_msa * self.attn(modulate(self.norm1(x), shift_msa, scale_msa, unsqueeze=False), y, pos)
+        x = x + gate_mlp * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp, unsqueeze=False))
+        return x
+
+
+class JiTT2ITextRefineBlock(nn.Module):
+    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, eps=1e-6):
+        super().__init__()
+        self.norm1 = RMSNorm(hidden_size, eps=eps)
+        self.attn = JiTT2ITextRefineAttention(hidden_size, num_heads=num_heads, eps=eps)
+        self.norm2 = RMSNorm(hidden_size, eps=eps)
+        self.mlp = JiTT2ISwiGLU(hidden_size, int(hidden_size * mlp_ratio))
+        self.adaLN_modulation = nn.Linear(hidden_size, 6 * hidden_size, bias=True)
+
+    def forward(self, x, c):
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=-1)
+        x = x + gate_msa * self.attn(modulate(self.norm1(x), shift_msa, scale_msa, unsqueeze=False))
+        x = x + gate_mlp * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp, unsqueeze=False))
+        return x
+
+
+class JiTT2IFinalLayer(nn.Module):
+    def __init__(self, hidden_size, out_channels):
+        super().__init__()
+        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.linear = nn.Linear(hidden_size, out_channels, bias=True)
+
+    def forward(self, x):
+        return self.linear(self.norm_final(x))
+
+
+class JiTTransformer2DT2IModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: int = 512,
+        patch_size: int = 16,
+        in_channels: int = 3,
+        hidden_size: int = 1536,
+        num_layers: int = 16,
+        num_attention_heads: int = 24,
+        num_text_layers: int = 4,
+        text_embed_dim: int = 2048,
+        text_max_length: int = 128,
+        bottleneck_dim: int = 256,
+        mlp_ratio: float = 4.0,
+        model_type: str | None = None,
+    ):
+        super().__init__()
+        if model_type in PIXELGEN_T2I_PRESET_CONFIGS:
+            preset = PIXELGEN_T2I_PRESET_CONFIGS[model_type]
+            sample_size = int(preset["sample_size"])
+            patch_size = int(preset["patch_size"])
+            in_channels = int(preset["in_channels"])
+            hidden_size = int(preset["hidden_size"])
+            num_layers = int(preset["num_layers"])
+            num_attention_heads = int(preset["num_attention_heads"])
+            num_text_layers = int(preset["num_text_layers"])
+            text_embed_dim = int(preset["text_embed_dim"])
+            text_max_length = int(preset["text_max_length"])
+            bottleneck_dim = int(preset["bottleneck_dim"])
+
+        self.sample_size = sample_size
+        self.patch_size = patch_size
+        self.in_channels = in_channels
+        self.out_channels = in_channels
+        self.hidden_size = hidden_size
+        self.gradient_checkpointing = False
+
+        patch_dim = in_channels * patch_size * patch_size
+        self.x_embedder = JiTT2IBottleneckEmbed(patch_dim, hidden_size, bottleneck_dim)
+        self.t_embedder = JiTT2ITimestepEmbedder(hidden_size)
+        self.y_embedder = JiTT2ITextEmbedder(text_embed_dim, hidden_size)
+
+        self.num_patches = (sample_size // patch_size) ** 2
+        pos_embed = get_2d_sincos_pos_embed(hidden_size, int(self.num_patches**0.5))
+        self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=True)
+        self.y_pos_embedding = nn.Parameter(torch.randn(1, text_max_length, hidden_size) * 0.02)
+
+        self.blocks = nn.ModuleList(
+            [JiTT2IBlock(hidden_size, num_attention_heads, mlp_ratio=mlp_ratio) for _ in range(num_layers)]
+        )
+        self.text_refine_blocks = nn.ModuleList(
+            [JiTT2ITextRefineBlock(hidden_size, num_attention_heads, mlp_ratio=mlp_ratio) for _ in range(num_text_layers)]
+        )
+        self.final_layer = JiTT2IFinalLayer(hidden_size, patch_dim)
+        self._pos_cache: Dict[Tuple[int, int], torch.Tensor] = {}
+
+    def _fetch_pos(self, height: int, width: int, device: torch.device, scale_h: float = 1.0, scale_w: float = 1.0):
+        key = (height, width, scale_h, scale_w)
+        if key not in self._pos_cache:
+            self._pos_cache[key] = precompute_freqs_cis_2d(
+                self.hidden_size // self.config.num_attention_heads,
+                height,
+                width,
+                scale=(scale_h, scale_w),
+            )
+        return self._pos_cache[key].to(device)
+
+    def unpatchify(self, x, height, width):
+        return F.fold(
+            x.transpose(1, 2).contiguous(),
+            output_size=(height, width),
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+        )
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        timestep: torch.LongTensor,
+        encoder_hidden_states: torch.Tensor,
+        scale_h: float = 1.0,
+        scale_w: float = 1.0,
+        return_dict: bool = True,
+    ):
+        batch_size, _, height, width = sample.shape
+        x = F.unfold(sample, kernel_size=self.patch_size, stride=self.patch_size).transpose(1, 2)
+        xpos = self._fetch_pos(height // self.patch_size, width // self.patch_size, sample.device, scale_h, scale_w)
+
+        t = self.t_embedder(timestep.reshape(-1), dtype=sample.dtype).view(batch_size, -1, self.hidden_size)
+        y = self.y_embedder(encoder_hidden_states) + self.y_pos_embedding.to(encoder_hidden_states.dtype)
+        condition = F.silu(t)
+
+        for block in self.text_refine_blocks:
+            y = block(y, condition)
+
+        x = self.x_embedder(x) + self.pos_embed.to(x.dtype)
+        for block in self.blocks:
+            if self.training and self.gradient_checkpointing:
+                x = torch.utils.checkpoint.checkpoint(block, x, y, condition, xpos, use_reentrant=False)
+            else:
+                x = block(x, y, condition, xpos)
+
+        x = self.final_layer(x)
+        output = self.unpatchify(x, height, width)
+
+        if not return_dict:
+            return (output,)
+        return Transformer2DModelOutput(sample=output)
+
+    @classmethod
+    def from_pixelgen_checkpoint(
+        cls,
+        checkpoint_path: str,
+        weights_prefix: str = "ema_denoiser.",
+        model_type: str = "PixelGen-XXL/16-T2I",
+        map_location: str = "cpu",
+        strict: bool = False,
+    ) -> Tuple["JiTTransformer2DT2IModel", Dict[str, object]]:
+        checkpoint = torch.load(checkpoint_path, map_location=map_location, weights_only=False)
+        state_dict = checkpoint.get("state_dict", checkpoint)
+        filtered = {k[len(weights_prefix) :]: v for k, v in state_dict.items() if k.startswith(weights_prefix)}
+        if not filtered:
+            raise ValueError(f"No parameters found with prefix '{weights_prefix}' in checkpoint.")
+
+        config = dict(PIXELGEN_T2I_PRESET_CONFIGS[model_type])
+        config["model_type"] = model_type
+        model = cls(**config)
+        remapped = remap_t2i_state_dict(filtered)
+        model.load_state_dict(remapped, strict=strict)
+        return model, {"checkpoint_path": checkpoint_path, "model_type": model_type}

README.md

@@ -0,0 +1,193 @@
+---
+license: mit
+library_name: diffusers
+pipeline_tag: text-to-image
+tags:
+  - diffusers
+  - image-generation
+  - class-conditional
+  - text-to-image
+  - imagenet
+  - pixelgen
+  - flow-matching
+  - pixel-space
+  - jit
+widget:
+  - text: golden retriever
+    output:
+      url: PixelGen-XL-16-256/demo.png
+language:
+  - en
+---
+
+# BiliSakura/PixelGen-diffusers
+
+Self-contained PixelGen checkpoints for Hugging Face diffusers. Each variant folder ships its own pipeline code, component modules, and weights.
+
+Converted from upstream PixelGen checkpoints using [PixelGen-diffusers](https://github.com/Bili-Sakura/Visual-Generative-Foundation-Model-Collection/tree/main/libs/PixelGen-diffusers) in [Visual-Generative-Foundation-Model-Collection](https://github.com/Bili-Sakura/Visual-Generative-Foundation-Model-Collection).
+
+## Available checkpoints
+
+| Subfolder | Pipeline | Task | Resolution | Model type |
+| --- | --- | --- | ---: | --- |
+| [`PixelGen-XL-16-256/`](PixelGen-XL-16-256/) | `PixelGenC2IPipeline` | class-to-image | 256×256 | PixelGen-XL/16 |
+| [`PixelGen-XXL-16-512-t2i/`](PixelGen-XXL-16-512-t2i/) | `PixelGenT2IPipeline` | text-to-image | 512×512 | PixelGen-XXL/16-T2I |
+
+## Repo layout
+
+```text
+BiliSakura/PixelGen-diffusers/
+├── README.md
+├── PixelGen-XL-16-256/
+│   ├── pipeline.py
+│   ├── model_index.json
+│   ├── demo.png
+│   ├── scheduler/
+│   │   ├── scheduler_config.json
+│   │   └── scheduling_pixelgen.py
+│   └── transformer/
+│       ├── config.json
+│       └── transformer_jit.py
+└── PixelGen-XXL-16-512-t2i/
+    ├── pipeline.py
+    ├── model_index.json
+    ├── conversion_metadata.json
+    ├── scheduler/
+    │   ├── scheduler_config.json
+    │   └── scheduling_pixelgen.py
+    ├── text_encoder/
+    ├── tokenizer/
+    └── transformer/
+        ├── config.json
+        ├── diffusion_pytorch_model.safetensors
+        └── transformer_jit_t2i.py
+```
+
+Each class-conditional variant is self-contained: load with `custom_pipeline=.../pipeline.py` and `trust_remote_code=True`. PixelGen denoises directly in pixel space (no VAE).
+
+## ImageNet class labels
+
+For [`PixelGen-XL-16-256/`](PixelGen-XL-16-256/), `id2label` is embedded in `model_index.json` (DiT-style).
+
+- `pipe.id2label` — inspect id → English label correspondence
+- `pipe.labels` — reverse map (English synonym → id)
+- `pipe.get_label_ids("golden retriever")`
+- `pipe(class_labels="golden retriever", ...)` — string labels resolved automatically
+
+## Demo
+
+![PixelGen-XL-16-256 demo](PixelGen-XL-16-256/demo.png)
+
+Class 207 — golden retriever, 256×256, 50 steps, `guidance_scale=2.25`, Heun solver, `timeshift=2.0`.
+
+## Load from Hugging Face
+
+### Class-to-image (`PixelGen-XL-16-256`)
+
+```python
+import torch
+from diffusers import DiffusionPipeline
+
+pipe = DiffusionPipeline.from_pretrained(
+    "BiliSakura/PixelGen-diffusers/PixelGen-XL-16-256",
+    trust_remote_code=True,
+    torch_dtype=torch.bfloat16,
+).to("cuda")
+
+print(pipe.id2label[207])
+print(pipe.get_label_ids("golden retriever"))
+
+generator = torch.Generator(device="cuda").manual_seed(0)
+images = pipe(
+    class_labels="golden retriever",
+    num_inference_steps=50,
+    guidance_scale=2.25,
+    generator=generator,
+).images
+```
+
+### Text-to-image (`PixelGen-XXL-16-512-t2i`)
+
+Uses a bundled Qwen3 text encoder when `text_encoder/` is present; otherwise downloads from the path recorded in `conversion_metadata.json`.
+
+```python
+import torch
+from diffusers import DiffusionPipeline
+
+pipe = DiffusionPipeline.from_pretrained(
+    "BiliSakura/PixelGen-diffusers/PixelGen-XXL-16-512-t2i",
+    trust_remote_code=True,
+    torch_dtype=torch.bfloat16,
+)
+
+generator = torch.Generator(device="cuda").manual_seed(42)
+images = pipe(
+    prompt="A golden retriever playing in a sunny garden",
+    num_inference_steps=50,
+    guidance_scale=4.0,
+    generator=generator,
+).images
+```
+
+## Load from a local clone
+
+### Class-to-image (`PixelGen-XL-16-256`)
+
+```python
+from pathlib import Path
+import torch
+from diffusers import DiffusionPipeline
+
+model_dir = Path("./PixelGen-XL-16-256").resolve()
+pipe = DiffusionPipeline.from_pretrained(
+    str(model_dir),
+    local_files_only=True,
+    custom_pipeline=str(model_dir / "pipeline.py"),
+    trust_remote_code=True,
+    torch_dtype=torch.bfloat16,
+).to("cuda")
+
+generator = torch.Generator(device="cuda").manual_seed(0)
+image = pipe(
+    class_labels="golden retriever",
+    num_inference_steps=50,
+    guidance_scale=2.25,
+    generator=generator,
+).images[0]
+image.save("demo.png")
+```
+
+## Recommended inference settings
+
+| Variant | Steps | CFG scale | Solver | Timeshift | CFG interval |
+| --- | ---: | ---: | --- | ---: | --- |
+| `PixelGen-XL-16-256` | 50 | 2.25 | heun | 2.0 | [0.1, 0.9] |
+| `PixelGen-XXL-16-512-t2i` | 25 | 4.0 | adam_lm | 3.0 | [0.0, 1.0] |
+
+`height` and `width` are fixed by each checkpoint's `sample_size`. Custom sizes are not supported for these exports.
+
+## Interface notes
+
+- Class-conditional generation uses `class_labels` (integer ImageNet id or English synonym).
+- `guidance_scale > 1.0` enables classifier-free guidance over a null class token.
+- `sampling_method` accepts `heun` or `euler` for C2I; T2I defaults to `adam_lm`.
+- `noise_scale` defaults to `1.0` at 256×256 and `2.0` at 512×512 when not specified.
+
+## Citation
+
+Source paper:
+
+- [PixelGen: Improving Pixel Diffusion with Perceptual Loss](https://arxiv.org/abs/2602.02493)
+- [Hugging Face Papers page](https://huggingface.co/papers/2602.02493)
+
+```bibtex
+@article{ma2026pixelgen,
+  title={PixelGen: Improving Pixel Diffusion with Perceptual Loss},
+  author={Zehong Ma and Ruihan Xu and Shiliang Zhang},
+  year={2026},
+  eprint={2602.02493},
+  archivePrefix={arXiv},
+  primaryClass={cs.CV},
+  url={https://arxiv.org/abs/2602.02493},
+}
+```