Add JLT-H-1 metadata and update B/L small files

.gitattributes

@@ -36,3 +36,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 demo.png filter=lfs diff=lfs merge=lfs -text
 JLT-B-1/demo.png filter=lfs diff=lfs merge=lfs -text
 JLT-L-1/demo.png filter=lfs diff=lfs merge=lfs -text
+JLT-H-1/demo.png filter=lfs diff=lfs merge=lfs -text

JLT-B-1/pipeline.py

@@ -69,9 +69,11 @@ def _solver_from_scheduler(scheduler) -> str:
 
 class JLTPipeline(DiffusionPipeline):
     r"""
-    Pipeline for class-conditional image generation with JLT (Clean-Latent JiT).
+    Class-conditional JLT sampler.
 
-    Latent models (`in_channels=128`) require a FLUX.2 VAE attached for `output_type="pil"`.
+    Starts from unit Gaussian noise in FLUX.2 latent space (`noise_scale=1.0`) at the
+    trained 256×256 resolution, integrates `JLTScheduler` (Heun, last step Euler),
+    applies interval CFG in velocity space, and decodes FLUX.2 latents when a VAE is attached.
     """
 
     model_cpu_offload_seq = "transformer->vae"
@@ -372,9 +374,8 @@ class JLTPipeline(DiffusionPipeline):
                 ImageNet class id(s) or English label(s).
             guidance_scale (`float`, *optional*):
                 CFG scale. Enabled when `guidance_scale > 1`.
-            noise_scale (`float`, *optional*, defaults to `None`):
-                Std of the initial noise. JiT/JLT scale this with resolution to keep SNR constant
-                (`1.0` at 256, `2.0` at 512, `4.0` at 1024). See huggingface/diffusers#13098.
+            noise_scale (`float`, *optional*, defaults to `1.0`):
+                Std of the initial noise. These checkpoints are trained at 256×256 with unit-scale FLUX.2 latents.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps.
             output_type (`str`, *optional*, defaults to `"pil"`):
@@ -400,13 +401,10 @@ class JLTPipeline(DiffusionPipeline):
         batch_size = len(class_label_ids)
         latent_size = int(self.transformer.config.sample_size)
         channels = int(self.transformer.config.in_channels)
-        # Official JiT uses pixel `sample_size`. JLT latent models store the latent grid.
-        image_size = latent_size * 16 if channels == 128 else latent_size
         null_class_val = int(
             getattr(self.transformer.config, "num_classes", getattr(self.transformer.config, "num_class_embeds", 1000))
         )
-
-        noise_scale = noise_scale if noise_scale is not None else (image_size / 256.0)
+        noise_scale = 1.0 if noise_scale is None else float(noise_scale)
 
         latents = noise_scale * randn_tensor(
             shape=(batch_size, channels, latent_size, latent_size),

JLT-B-1/scheduler/scheduling_jlt.py

@@ -102,6 +102,7 @@ class JLTScheduler(SchedulerMixin, ConfigMixin):
         num_inference_steps: int,
         device: Union[str, torch.device, None] = None,
     ) -> None:
+        """Build the `t: 0 → 1` knot schedule and Heun/Euler step plan."""
         if num_inference_steps < 1:
             raise ValueError("num_inference_steps must be >= 1.")
         if self.config.solver == "heun" and num_inference_steps < 2:
@@ -208,6 +209,7 @@ class JLTScheduler(SchedulerMixin, ConfigMixin):
         prediction_type: Optional[str] = None,
         t_eps: Optional[float] = None,
     ) -> Union[JLTSchedulerOutput, Tuple[torch.Tensor]]:
+        """Advance one scheduled ODE step. `model_output` is clean-`x` or velocity."""
         del generator
         step_index = self._resolve_step_index(timestep)
         if step_index >= len(self._step_plan):

JLT-B-1/transformer/transformer_jlt.py

@@ -1,4 +1,8 @@
-"""JLT (Clean-Latent) Transformer for diffusers."""
+"""JLT transformer: class-conditional clean-latent (or pixel) DiT.
+
+Architecture hyperparameters are read from `config.json` / constructor kwargs.
+Named training presets such as `JiT-B/1` are not part of this module.
+"""
 
 from __future__ import annotations
 
@@ -24,6 +28,8 @@ def rotate_half(x):
 
 
 class JLTRotaryEmbedding(nn.Module):
+    """2D axial RoPE for JLT attention, matching the official JiT implementation."""
+
     def __init__(self, dim, pt_seq_len=16, ft_seq_len=None, theta=10000):
         super().__init__()
         freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
@@ -52,6 +58,8 @@ def modulate(x, shift, scale):
 
 
 class JLTPatchEmbed(nn.Module):
+    """Conv patch embedder for latents (`patch_size=1`) or pixels."""
+
     def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, bias=True):
         super().__init__()
         img_size = (img_size, img_size)
@@ -62,15 +70,12 @@ class JLTPatchEmbed(nn.Module):
         self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias)
 
     def forward(self, x):
-        batch, _, height, width = x.shape
-        if height != self.img_size[0] or width != self.img_size[1]:
-            raise ValueError(
-                f"Input image size ({height}*{width}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
-            )
         return self.proj(x).flatten(2).transpose(1, 2)
 
 
 class JLTTimestepEmbedder(nn.Module):
+    """Sinusoidal timestep MLP used for adaLN conditioning."""
+
     def __init__(self, hidden_size, frequency_embedding_size=256):
         super().__init__()
         self.mlp = nn.Sequential(
@@ -100,6 +105,8 @@ class JLTTimestepEmbedder(nn.Module):
 
 
 class JLTLabelEmbedder(nn.Module):
+    """Class embedding table with one extra null class for CFG."""
+
     def __init__(self, num_classes, hidden_size):
         super().__init__()
         self.embedding_table = nn.Embedding(num_classes + 1, hidden_size)
@@ -110,6 +117,8 @@ class JLTLabelEmbedder(nn.Module):
 
 
 class JLTAttnProcessor:
+    """Official JiT-style attention: split `to_q` / `to_k` / `to_v` + `dispatch_attention_fn`."""
+
     _attention_backend = None
     _parallel_config = None
 
@@ -141,6 +150,8 @@ class JLTAttnProcessor:
 
 
 class JLTAttention(nn.Module, AttentionModuleMixin):
+    """Multi-head self-attention with QK RMSNorm, RoPE, and separate QKV projections."""
+
     _default_processor_cls = JLTAttnProcessor
     _available_processors = [JLTAttnProcessor]
 
@@ -170,6 +181,8 @@ class JLTAttention(nn.Module, AttentionModuleMixin):
 
 
 class JLTSwiGLUFFN(nn.Module):
+    """SwiGLU feed-forward block (`w12` fused, then `w3`)."""
+
     def __init__(self, dim: int, hidden_dim: int, drop=0.0, bias=True) -> None:
         super().__init__()
         hidden_dim = int(hidden_dim * 2 / 3)
@@ -184,6 +197,8 @@ class JLTSwiGLUFFN(nn.Module):
 
 
 class JLTBlock(nn.Module):
+    """Pre-norm transformer block with adaLN-Zero modulation."""
+
     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)
@@ -209,6 +224,24 @@ class JLTBlock(nn.Module):
 
 
 class JLTTransformer2DModel(ModelMixin, ConfigMixin, AttentionMixin):
+    r"""
+    2D transformer for class-conditional clean-latent generation ([JLT](https://arxiv.org/abs/2605.27102)).
+
+    The model predicts the clean sample `x`. Sampling uses [`JLTScheduler`] (`t`: 0 noise → 1 data). Latent
+    checkpoints use FLUX.2 latents (`in_channels=128`, `sample_size` is the latent grid).
+
+    Args:
+        sample_size (`int`, defaults to 16): Spatial size of the input grid.
+        patch_size (`int`, defaults to 1): Patch embed kernel / stride.
+        in_channels (`int`, defaults to 128): Input/output channels.
+        hidden_size (`int`, defaults to 768): Transformer width.
+        num_layers (`int`, defaults to 12): Number of `JLTBlock`s.
+        num_attention_heads (`int`, defaults to 12): Attention heads.
+        mlp_ratio (`float`, defaults to 4.0): SwiGLU expansion.
+        num_classes (`int`, defaults to 1000): ImageNet classes (plus a null CFG class).
+        model_type (`str`, *optional*): Training-name label stored in `config.json` only.
+    """
+
     _supports_gradient_checkpointing = True
     _skip_layerwise_casting_patterns = ["pos_embed", "norm"]
     _no_split_modules = ["JLTBlock"]
@@ -335,6 +368,7 @@ class JLTTransformer2DModel(ModelMixin, ConfigMixin, AttentionMixin):
         return_features: bool = False,
         feature_layers: Optional[List[int]] = None,
     ):
+        """Predict the clean sample from a noisy latent / image `sample` at JLT time `timestep`."""
         t = torch.as_tensor(timestep, device=sample.device)
         if t.ndim == 0:
             t = t.repeat(sample.shape[0])

JLT-H-1/model_index.json

@@ -0,0 +1,1021 @@
+{
+  "_class_name": [
+    "pipeline",
+    "JLTPipeline"
+  ],
+  "_diffusers_version": "0.39.0",
+  "transformer": [
+    "transformer_jlt",
+    "JLTTransformer2DModel"
+  ],
+  "scheduler": [
+    "scheduling_jlt",
+    "JLTScheduler"
+  ],
+  "vae": [
+    "diffusers",
+    "AutoencoderKLFlux2"
+  ],
+  "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"
+  }
+}

JLT-H-1/pipeline.py

@@ -0,0 +1,456 @@
+"""Hub custom pipeline: JLTPipeline.
+Load with native Hugging Face diffusers and trust_remote_code=True.
+"""
+
+from __future__ import annotations
+
+import importlib
+import importlib.util
+import json
+import sys
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+from diffusers import AutoencoderKLFlux2
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from diffusers.utils.torch_utils import randn_tensor
+
+
+def _load_jlt_scheduler_module():
+    module_path = Path(__file__).resolve().parent / "scheduler" / "scheduling_jlt.py"
+    spec = importlib.util.spec_from_file_location("scheduling_jlt", module_path)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+JLTScheduler = _load_jlt_scheduler_module().JLTScheduler
+
+
+def _unpatchify_latents(latents: torch.Tensor) -> torch.Tensor:
+    batch_size, num_channels_latents, height, width = latents.shape
+    latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), 2, 2, height, width)
+    latents = latents.permute(0, 1, 4, 2, 5, 3)
+    return latents.reshape(batch_size, num_channels_latents // (2 * 2), height * 2, width * 2)
+
+
+def _flux2_bn_stats(vae: AutoencoderKLFlux2, ref: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+    mean = vae.bn.running_mean.view(1, -1, 1, 1).to(ref.device, ref.dtype)
+    std = torch.sqrt(
+        vae.bn.running_var.view(1, -1, 1, 1).to(ref.device, ref.dtype) + vae.config.batch_norm_eps
+    )
+    return mean, std
+
+
+@torch.no_grad()
+def decode_flux2_latents(vae: AutoencoderKLFlux2, latents: torch.Tensor) -> torch.Tensor:
+    vae_param = next(vae.parameters())
+    latents = latents.to(device=vae_param.device, dtype=torch.float32)
+    mean, std = _flux2_bn_stats(vae, latents)
+    latents = latents * std + mean
+    latents = _unpatchify_latents(latents)
+    return vae.decode(latents.to(vae_param.dtype), return_dict=False)[0].float()
+
+
+def _new_jlt_scheduler(solver: str = "heun") -> JLTScheduler:
+    if solver not in {"heun", "euler"}:
+        raise ValueError("solver must be one of: 'heun', 'euler'.")
+    return JLTScheduler(solver=solver)
+
+
+def _solver_from_scheduler(scheduler) -> str:
+    solver = getattr(getattr(scheduler, "config", None), "solver", None)
+    if solver in {"heun", "euler"}:
+        return solver
+    return "heun"
+
+
+class JLTPipeline(DiffusionPipeline):
+    r"""
+    Class-conditional JLT sampler.
+
+    Starts from unit Gaussian noise in FLUX.2 latent space (`noise_scale=1.0`) at the
+    trained 256×256 resolution, integrates `JLTScheduler` (Heun, last step Euler),
+    applies interval CFG in velocity space, and decodes FLUX.2 latents when a VAE is attached.
+    """
+
+    model_cpu_offload_seq = "transformer->vae"
+    _optional_components = ["vae"]
+
+    def __init__(
+        self,
+        transformer,
+        scheduler,
+        vae=None,
+        id2label=None,
+        prediction_type="sample",
+        t_eps=5e-2,
+        solver="heun",
+    ):
+        super().__init__()
+        if scheduler is None or isinstance(scheduler, list):
+            scheduler = _new_jlt_scheduler(solver)
+        self.register_modules(
+            transformer=transformer,
+            scheduler=scheduler,
+            vae=vae,
+        )
+        self.image_processor = VaeImageProcessor(vae_scale_factor=1)
+        self.prediction_type = prediction_type
+        self.t_eps = t_eps
+        self.solver = _solver_from_scheduler(scheduler)
+        self._id2label = self._normalize_id2label(id2label)
+        self.labels = self._build_label2id(self._id2label)
+        self._labels_loaded_from_model_index = bool(self._id2label)
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path=None, subfolder=None, **kwargs):
+        repo_root = Path(__file__).resolve().parent
+
+        if pretrained_model_name_or_path in (None, "", "."):
+            variant = repo_root
+        else:
+            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()
+            if subfolder:
+                variant = variant / subfolder
+
+        model_kwargs = dict(kwargs)
+        inserted: List[str] = []
+
+        def _load_transformer():
+            comp_dir = None
+            candidate = variant / "transformer"
+            if (candidate / "transformer_jlt.py").exists() and (candidate / "config.json").exists():
+                comp_dir = candidate
+            else:
+                comp_dir = None
+            if comp_dir is None:
+                return None
+            module_path = comp_dir / "transformer_jlt.py"
+            if not module_path.exists() or not (comp_dir / "config.json").exists():
+                return None
+
+            comp_path = str(comp_dir)
+            if comp_path not in sys.path:
+                sys.path.insert(0, comp_path)
+                inserted.append(comp_path)
+
+            module = importlib.import_module("transformer_jlt")
+            transformer_cls = getattr(module, "JLTTransformer2DModel")
+            return transformer_cls.from_pretrained(str(comp_dir), **model_kwargs)
+
+        def _load_scheduler():
+            scheduler_dir = variant / "scheduler"
+            config_path = scheduler_dir / "scheduler_config.json"
+            if not config_path.exists():
+                return _new_jlt_scheduler("heun")
+
+            scheduler_entry = None
+            model_index_path = variant / "model_index.json"
+            if model_index_path.exists():
+                scheduler_entry = json.loads(model_index_path.read_text(encoding="utf-8")).get("scheduler")
+
+            if scheduler_entry is None:
+                class_name = json.loads(config_path.read_text(encoding="utf-8")).get("_class_name")
+                scheduler_entry = ["diffusers", class_name]
+
+            module_name, class_name = scheduler_entry
+            if module_name == "diffusers":
+                import diffusers.schedulers as schedulers_pkg
+
+                scheduler_cls = getattr(schedulers_pkg, class_name)
+                return scheduler_cls.from_pretrained(str(scheduler_dir), **model_kwargs)
+
+            comp_path = str(scheduler_dir)
+            if comp_path not in sys.path:
+                sys.path.insert(0, comp_path)
+                inserted.append(comp_path)
+            module = importlib.import_module(module_name)
+            scheduler_cls = getattr(module, class_name)
+            return scheduler_cls.from_pretrained(str(scheduler_dir), **model_kwargs)
+
+        def _load_vae():
+            vae_dir = variant / "vae"
+            if vae_dir.exists() and (vae_dir / "config.json").exists():
+                return AutoencoderKLFlux2.from_pretrained(str(vae_dir), **model_kwargs)
+            return None
+
+        try:
+            transformer = _load_transformer()
+            if transformer is None:
+                raise ValueError(f"No loadable transformer found under {variant}")
+
+            scheduler = _load_scheduler()
+            vae = _load_vae()
+            id2label = cls._read_id2label_from_model_index(str(variant))
+            solver = _solver_from_scheduler(scheduler)
+            pipe = cls(
+                transformer=transformer,
+                scheduler=scheduler,
+                vae=vae,
+                id2label=id2label or None,
+                solver=solver,
+            )
+            if hasattr(pipe, "register_to_config"):
+                pipe.register_to_config(_name_or_path=str(variant))
+            return pipe
+        finally:
+            for comp_path in inserted:
+                if comp_path in sys.path:
+                    sys.path.remove(comp_path)
+
+    @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 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 label(s): {missing}. Examples: {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):
+            return self.get_label_ids(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_at_t(
+        self,
+        z_value: torch.Tensor,
+        t_jlt: Union[float, 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:
+        t_jlt = torch.as_tensor(t_jlt, device=z_value.device, dtype=z_value.dtype)
+        if t_jlt.ndim == 0:
+            t_jlt = t_jlt.reshape(1)
+
+        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_jlt.flatten().expand(z_in.shape[0]).to(dtype=self.transformer.dtype)
+        model_out = self.transformer(
+            z_in.to(self.transformer.dtype),
+            timestep=t_batch,
+            class_labels=labels,
+        ).sample
+
+        # velocity in fp32; clamp the 1 / (1 - t) denominator at JLT's `t_eps` (0.05).
+        v = self.scheduler.velocity_from_prediction(
+            z_in.float(),
+            model_out.float(),
+            t_jlt,
+            prediction_type=self.prediction_type,
+            t_eps=self.t_eps,
+        )
+
+        if not do_classifier_free_guidance:
+            return v
+
+        v_cond, v_uncond = v.chunk(2, dim=0)
+        interval_mask = t_jlt < guidance_interval_max
+        if guidance_interval_min != 0.0:
+            interval_mask = interval_mask & (t_jlt > guidance_interval_min)
+        scale = torch.where(
+            interval_mask,
+            torch.tensor(guidance_scale, device=z_value.device, dtype=torch.float32),
+            torch.tensor(1.0, device=z_value.device, dtype=torch.float32),
+        )
+        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,
+    ) -> torch.Tensor:
+        self.scheduler.set_timesteps(num_inference_steps, device=latents.device)
+        for t_jlt in self.progress_bar(self.scheduler.timesteps):
+            velocity = self._predict_velocity_at_t(
+                latents,
+                t_jlt,
+                class_labels,
+                class_null,
+                do_classifier_free_guidance,
+                guidance_scale,
+                guidance_interval_min,
+                guidance_interval_max,
+            )
+            latents = self.scheduler.step(
+                velocity,
+                t_jlt,
+                latents,
+                prediction_type="velocity",
+                t_eps=self.t_eps,
+            ).prev_sample
+        return latents
+
+    @torch.no_grad()
+    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 = 1.0,
+        noise_scale: Optional[float] = None,
+        t_eps: 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]:
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            class_labels (`int`, `str`, or list):
+                ImageNet class id(s) or English label(s).
+            guidance_scale (`float`, *optional*):
+                CFG scale. Enabled when `guidance_scale > 1`.
+            noise_scale (`float`, *optional*, defaults to `1.0`):
+                Std of the initial noise. These checkpoints are trained at 256×256 with unit-scale FLUX.2 latents.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                `"pil"`, `"np"`, `"pt"`, or `"latent"`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return an [`ImagePipelineOutput`].
+        """
+        if num_inference_steps < 2:
+            raise ValueError("num_inference_steps must be >= 2.")
+        if output_type not in {"pil", "np", "pt", "latent"}:
+            raise ValueError("output_type must be one of: 'pil', 'np', 'pt', 'latent'.")
+
+        if t_eps is not None:
+            self.t_eps = t_eps
+            if hasattr(self.scheduler, "register_to_config"):
+                self.scheduler.register_to_config(t_eps=t_eps)
+
+        class_label_ids = self._normalize_class_labels(class_labels)
+        if guidance_scale is None:
+            guidance_scale = 1.0
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        batch_size = len(class_label_ids)
+        latent_size = int(self.transformer.config.sample_size)
+        channels = int(self.transformer.config.in_channels)
+        null_class_val = int(
+            getattr(self.transformer.config, "num_classes", getattr(self.transformer.config, "num_class_embeds", 1000))
+        )
+        noise_scale = 1.0 if noise_scale is None else float(noise_scale)
+
+        latents = noise_scale * randn_tensor(
+            shape=(batch_size, channels, latent_size, latent_size),
+            generator=generator,
+            device=self._execution_device,
+            dtype=torch.float32,
+        )
+
+        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,
+        )
+
+        if output_type == "latent":
+            if not return_dict:
+                return (latents,)
+            return ImagePipelineOutput(images=latents)
+
+        if self.vae is not None:
+            images_pt = decode_flux2_latents(self.vae, latents)
+        else:
+            images_pt = latents
+
+        if not torch.isfinite(images_pt).all():
+            raise RuntimeError(
+                "JLT generation produced non-finite values before image conversion. "
+                "Restart the notebook kernel and re-run the pipeline load cell."
+            )
+
+        if output_type in {"latent", "pt"}:
+            images = images_pt
+        else:
+            images = self.image_processor.postprocess(images_pt, output_type=output_type)
+
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (images,)
+        return ImagePipelineOutput(images=images)

JLT-H-1/scheduler/scheduler_config.json

@@ -0,0 +1,8 @@
+{
+  "_class_name": "JLTScheduler",
+  "_diffusers_version": "0.39.0",
+  "num_train_timesteps": 1000,
+  "prediction_type": "sample",
+  "solver": "heun",
+  "t_eps": 0.05
+}

JLT-H-1/scheduler/scheduling_jlt.py

@@ -0,0 +1,250 @@
+"""Hub custom scheduler: JLTScheduler.
+
+JLT integrates a clean-latent ODE. Time runs from 0 (noise) to 1 (data):
+
+    z_t = t * x + (1 - t) * eps
+
+The transformer predicts clean latent ``x`` by default (``prediction_type="sample"``).
+``step()`` integrates ``dz/dt``. Heun uses a predictor/corrector pair; the last
+interval is always Euler because velocity is singular at ``t = 1``.
+"""
+
+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 velocity_from_prediction(
+    sample: torch.Tensor,
+    model_output: torch.Tensor,
+    timestep: Union[float, torch.Tensor],
+    *,
+    prediction_type: str = "sample",
+    t_eps: float = 5e-2,
+) -> torch.Tensor:
+    """Convert a JLT model output to ODE velocity ``dz/dt``.
+
+    Follows the official JiT pipeline (diffusers#13098): compute velocity in fp32
+    and clamp the ``1 / (1 - t)`` denominator at ``t_eps`` (default 0.05).
+    """
+    if prediction_type == "velocity":
+        return model_output.float()
+    if prediction_type != "sample":
+        raise ValueError("prediction_type must be one of: 'sample', 'velocity'.")
+
+    sample = sample.float()
+    model_output = model_output.float()
+    t = torch.as_tensor(timestep, device=sample.device, dtype=torch.float32)
+    while t.ndim < sample.ndim:
+        t = t.unsqueeze(-1)
+    denom = (1.0 - t).clamp(min=t_eps)
+    return (model_output - sample) / denom
+
+
+@dataclass
+class JLTSchedulerOutput(BaseOutput):
+    prev_sample: torch.Tensor
+
+
+class JLTScheduler(SchedulerMixin, ConfigMixin):
+    """Clean-latent ODE scheduler for JLT (``t``: 0 -> 1)."""
+
+    config_name = "scheduler_config.json"
+    _compatibles = []
+    order = 2
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        solver: str = "heun",
+        prediction_type: str = "sample",
+        t_eps: float = 5e-2,
+    ):
+        if solver not in {"heun", "euler"}:
+            raise ValueError("solver must be one of: 'heun', 'euler'.")
+        if prediction_type not in {"sample", "velocity"}:
+            raise ValueError("prediction_type must be one of: 'sample', 'velocity'.")
+
+        self.timesteps: Optional[torch.Tensor] = None
+        self.schedule: Optional[torch.Tensor] = None
+        self.num_inference_steps: Optional[int] = None
+        self._step_plan: List[dict] = []
+        self._step_index: Optional[int] = None
+        self._begin_index: Optional[int] = None
+        self._heun_sample: Optional[torch.Tensor] = None
+        self._heun_velocity: Optional[torch.Tensor] = None
+        self._reset_heun_state()
+
+    @property
+    def init_noise_sigma(self) -> float:
+        return 1.0
+
+    @property
+    def step_index(self) -> Optional[int]:
+        return self._step_index
+
+    @property
+    def begin_index(self) -> Optional[int]:
+        return self._begin_index
+
+    def set_begin_index(self, begin_index: int = 0) -> None:
+        self._begin_index = begin_index
+
+    def set_timesteps(
+        self,
+        num_inference_steps: int,
+        device: Union[str, torch.device, None] = None,
+    ) -> None:
+        """Build the `t: 0 → 1` knot schedule and Heun/Euler step plan."""
+        if num_inference_steps < 1:
+            raise ValueError("num_inference_steps must be >= 1.")
+        if self.config.solver == "heun" and num_inference_steps < 2:
+            raise ValueError("Heun solver requires num_inference_steps >= 2.")
+
+        knots = torch.linspace(0.0, 1.0, num_inference_steps + 1, device=device, dtype=torch.float32)
+        plan: List[dict] = []
+        if self.config.solver == "heun":
+            for idx in range(num_inference_steps - 1):
+                t_cur = float(knots[idx])
+                t_next = float(knots[idx + 1])
+                plan.append({"kind": "heun_predict", "t_model": t_cur, "t_cur": t_cur, "t_next": t_next})
+                plan.append({"kind": "heun_correct", "t_model": t_next, "t_cur": t_cur, "t_next": t_next})
+            plan.append(
+                {
+                    "kind": "euler",
+                    "t_model": float(knots[-2]),
+                    "t_cur": float(knots[-2]),
+                    "t_next": float(knots[-1]),
+                }
+            )
+        else:
+            for idx in range(num_inference_steps):
+                plan.append(
+                    {
+                        "kind": "euler",
+                        "t_model": float(knots[idx]),
+                        "t_cur": float(knots[idx]),
+                        "t_next": float(knots[idx + 1]),
+                    }
+                )
+
+        self.num_inference_steps = num_inference_steps
+        self.schedule = knots
+        self._step_plan = plan
+        self.timesteps = torch.tensor([item["t_model"] for item in plan], device=device, dtype=torch.float32)
+        self._step_index = 0 if self._begin_index is None else self._begin_index
+        self._reset_heun_state()
+
+    def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor:
+        del timestep
+        return sample
+
+    def scale_noise(
+        self,
+        sample: torch.Tensor,
+        timestep: Union[float, torch.Tensor],
+        noise: torch.Tensor,
+    ) -> torch.Tensor:
+        """Forward process ``z = t * x + (1 - t) * eps``."""
+        t = torch.as_tensor(timestep, device=sample.device, dtype=sample.dtype)
+        while t.ndim < sample.ndim:
+            t = t.unsqueeze(-1)
+        return t * sample + (1.0 - t) * noise
+
+    def velocity_from_prediction(
+        self,
+        sample: torch.Tensor,
+        model_output: torch.Tensor,
+        timestep: Union[float, torch.Tensor],
+        *,
+        prediction_type: Optional[str] = None,
+        t_eps: Optional[float] = None,
+    ) -> torch.Tensor:
+        return velocity_from_prediction(
+            sample,
+            model_output,
+            timestep,
+            prediction_type=prediction_type or self.config.prediction_type,
+            t_eps=self.config.t_eps if t_eps is None else t_eps,
+        )
+
+    def _reset_heun_state(self) -> None:
+        self._heun_sample = None
+        self._heun_velocity = None
+
+    def _resolve_step_index(self, timestep: Union[float, torch.Tensor, None]) -> int:
+        if self.timesteps is None or not self._step_plan:
+            raise ValueError("Call `set_timesteps` before `step`.")
+        if self._step_index is not None:
+            return self._step_index
+        if timestep is None:
+            return 0
+
+        t_value = float(timestep) if not isinstance(timestep, torch.Tensor) else float(timestep.reshape(-1)[0])
+        matches = (self.timesteps - t_value).abs() < 1e-6
+        if matches.any():
+            return int(matches.nonzero(as_tuple=False)[0].item())
+        return 0
+
+    def _broadcast_dt(self, t_cur: float, t_next: float, sample: torch.Tensor) -> torch.Tensor:
+        dt = torch.as_tensor(t_next - t_cur, device=sample.device, dtype=sample.dtype)
+        while dt.ndim < sample.ndim:
+            dt = dt.unsqueeze(-1)
+        return dt
+
+    def step(
+        self,
+        model_output: torch.Tensor,
+        timestep: Union[float, torch.Tensor, None],
+        sample: torch.Tensor,
+        return_dict: bool = True,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        prediction_type: Optional[str] = None,
+        t_eps: Optional[float] = None,
+    ) -> Union[JLTSchedulerOutput, Tuple[torch.Tensor]]:
+        """Advance one scheduled ODE step. `model_output` is clean-`x` or velocity."""
+        del generator
+        step_index = self._resolve_step_index(timestep)
+        if step_index >= len(self._step_plan):
+            raise ValueError("Scheduler has already reached the final timestep.")
+
+        plan = self._step_plan[step_index]
+        sample = sample.to(dtype=torch.float32)
+        velocity = self.velocity_from_prediction(
+            sample,
+            model_output.to(dtype=sample.dtype, device=sample.device),
+            plan["t_model"],
+            prediction_type=prediction_type,
+            t_eps=t_eps,
+        )
+        dt = self._broadcast_dt(plan["t_cur"], plan["t_next"], sample)
+
+        kind = plan["kind"]
+        if kind == "heun_predict":
+            self._heun_sample = sample
+            self._heun_velocity = velocity
+            prev_sample = sample + dt * velocity
+        elif kind == "heun_correct":
+            if self._heun_sample is None or self._heun_velocity is None:
+                raise RuntimeError("Heun corrector called before predictor. Call `set_timesteps` and step in order.")
+            prev_sample = self._heun_sample + dt * 0.5 * (self._heun_velocity + velocity)
+            self._reset_heun_state()
+        else:
+            prev_sample = sample + dt * velocity
+            self._reset_heun_state()
+
+        self._step_index = step_index + 1
+
+        if not return_dict:
+            return (prev_sample,)
+        return JLTSchedulerOutput(prev_sample=prev_sample)
+
+    def __len__(self) -> int:
+        return int(self.config.num_train_timesteps)

JLT-H-1/transformer/config.json

@@ -0,0 +1,21 @@
+{
+  "_class_name": "JLTTransformer2DModel",
+  "_diffusers_version": "0.39.0",
+  "attention_dropout": 0.0,
+  "dropout": 0.2,
+  "hidden_size": 1280,
+  "in_channels": 128,
+  "loop_count": 0,
+  "loop_indices": null,
+  "mask_prob": 0.0,
+  "mask_ratio": 0.0,
+  "mlp_ratio": 4.0,
+  "model_type": "JiT-H/1",
+  "norm_eps": 1e-06,
+  "num_attention_heads": 20,
+  "num_class_embeds": null,
+  "num_classes": 1000,
+  "num_layers": 32,
+  "patch_size": 1,
+  "sample_size": 16
+}

JLT-H-1/transformer/transformer_jlt.py

@@ -0,0 +1,421 @@
+"""JLT transformer: class-conditional clean-latent (or pixel) DiT.
+
+Architecture hyperparameters are read from `config.json` / constructor kwargs.
+Named training presets such as `JiT-B/1` are not part of this module.
+"""
+
+from __future__ import annotations
+
+import math
+from typing import List, Optional, Tuple
+
+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.attention import AttentionMixin, AttentionModuleMixin
+from diffusers.models.attention_dispatch import dispatch_attention_fn
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import RMSNorm
+
+
+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 JLTRotaryEmbedding(nn.Module):
+    """2D axial RoPE for JLT attention, matching the official JiT implementation."""
+
+    def __init__(self, dim, pt_seq_len=16, ft_seq_len=None, theta=10000):
+        super().__init__()
+        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_h = freqs[:, None, :].expand(-1, freqs.shape[0], -1)
+        freqs_w = freqs[None, :, :].expand(freqs.shape[0], -1, -1)
+        freqs = torch.cat((freqs_h, freqs_w), dim=-1)
+
+        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(dtype=tensor.dtype)
+        freqs_sin = self.freqs_sin[:seq_len].to(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 JLTPatchEmbed(nn.Module):
+    """Conv patch embedder for latents (`patch_size=1`) or pixels."""
+
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, bias=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.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias)
+
+    def forward(self, x):
+        return self.proj(x).flatten(2).transpose(1, 2)
+
+
+class JLTTimestepEmbedder(nn.Module):
+    """Sinusoidal timestep MLP used for adaLN conditioning."""
+
+    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, device=t.device) / half
+        )
+        args = t.float().unsqueeze(-1) * freqs
+        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.to(t.dtype)
+
+    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 JLTLabelEmbedder(nn.Module):
+    """Class embedding table with one extra null class for CFG."""
+
+    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 JLTAttnProcessor:
+    """Official JiT-style attention: split `to_q` / `to_k` / `to_v` + `dispatch_attention_fn`."""
+
+    _attention_backend = None
+    _parallel_config = None
+
+    def __call__(self, attn: "JLTAttention", hidden_states: torch.Tensor, rope=None) -> torch.Tensor:
+        query = attn.to_q(hidden_states).unflatten(2, (attn.num_heads, attn.head_dim))
+        key = attn.to_k(hidden_states).unflatten(2, (attn.num_heads, attn.head_dim))
+        value = attn.to_v(hidden_states).unflatten(2, (attn.num_heads, attn.head_dim))
+
+        query = attn.norm_q(query)
+        key = attn.norm_k(key)
+
+        if rope is not None:
+            query = rope(query)
+            key = rope(key)
+
+        hidden_states = dispatch_attention_fn(
+            query,
+            key,
+            value,
+            dropout_p=attn.attn_drop if attn.training else 0.0,
+            is_causal=False,
+            backend=self._attention_backend,
+            parallel_config=self._parallel_config,
+        )
+        hidden_states = hidden_states.flatten(2, 3)
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+        return hidden_states
+
+
+class JLTAttention(nn.Module, AttentionModuleMixin):
+    """Multi-head self-attention with QK RMSNorm, RoPE, and separate QKV projections."""
+
+    _default_processor_cls = JLTAttnProcessor
+    _available_processors = [JLTAttnProcessor]
+
+    def __init__(
+        self, dim, num_heads=8, qkv_bias=True, qk_norm=True, attn_drop=0.0, proj_drop=0.0, eps=1e-6, processor=None
+    ):
+        super().__init__()
+        self.num_heads = num_heads
+        self.heads = num_heads
+        self.head_dim = dim // num_heads
+        self.attn_drop = attn_drop
+
+        self.norm_q = RMSNorm(self.head_dim, eps=eps) if qk_norm else nn.Identity()
+        self.norm_k = RMSNorm(self.head_dim, eps=eps) if qk_norm else nn.Identity()
+
+        self.to_q = nn.Linear(dim, dim, bias=qkv_bias)
+        self.to_k = nn.Linear(dim, dim, bias=qkv_bias)
+        self.to_v = nn.Linear(dim, dim, bias=qkv_bias)
+        self.to_out = nn.ModuleList([nn.Linear(dim, dim), nn.Dropout(proj_drop)])
+
+        if processor is None:
+            processor = self._default_processor_cls()
+        self.set_processor(processor)
+
+    def forward(self, x, rope=None):
+        return self.processor(self, x, rope=rope)
+
+
+class JLTSwiGLUFFN(nn.Module):
+    """SwiGLU feed-forward block (`w12` fused, then `w3`)."""
+
+    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)
+        return self.w3(self.ffn_dropout(F.silu(x1) * x2))
+
+
+class JLTBlock(nn.Module):
+    """Pre-norm transformer block with adaLN-Zero modulation."""
+
+    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 = JLTAttention(
+            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 = JLTSwiGLUFFN(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):
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(self.act(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
+
+
+class JLTTransformer2DModel(ModelMixin, ConfigMixin, AttentionMixin):
+    r"""
+    2D transformer for class-conditional clean-latent generation ([JLT](https://arxiv.org/abs/2605.27102)).
+
+    The model predicts the clean sample `x`. Sampling uses [`JLTScheduler`] (`t`: 0 noise → 1 data). Latent
+    checkpoints use FLUX.2 latents (`in_channels=128`, `sample_size` is the latent grid).
+
+    Args:
+        sample_size (`int`, defaults to 16): Spatial size of the input grid.
+        patch_size (`int`, defaults to 1): Patch embed kernel / stride.
+        in_channels (`int`, defaults to 128): Input/output channels.
+        hidden_size (`int`, defaults to 768): Transformer width.
+        num_layers (`int`, defaults to 12): Number of `JLTBlock`s.
+        num_attention_heads (`int`, defaults to 12): Attention heads.
+        mlp_ratio (`float`, defaults to 4.0): SwiGLU expansion.
+        num_classes (`int`, defaults to 1000): ImageNet classes (plus a null CFG class).
+        model_type (`str`, *optional*): Training-name label stored in `config.json` only.
+    """
+
+    _supports_gradient_checkpointing = True
+    _skip_layerwise_casting_patterns = ["pos_embed", "norm"]
+    _no_split_modules = ["JLTBlock"]
+    _repeated_blocks = ["JLTBlock"]
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: int = 16,
+        patch_size: int = 1,
+        in_channels: int = 128,
+        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,
+        norm_eps: float = 1e-6,
+        model_type: str | None = None,
+        num_class_embeds: int | None = None,
+        mask_prob: float = 0.0,
+        mask_ratio: float = 0.0,
+        loop_indices: Optional[List[int]] = None,
+        loop_count: int = 0,
+    ):
+        super().__init__()
+        if num_class_embeds is not None:
+            num_classes = int(num_class_embeds)
+
+        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.mask_prob = float(mask_prob)
+        self.mask_ratio = float(mask_ratio)
+        self.loop_count = int(loop_count)
+        self.gradient_checkpointing = False
+
+        self.mask_token = nn.Parameter(torch.zeros(1, 1, hidden_size))
+        nn.init.normal_(self.mask_token, std=0.02)
+
+        self.loop_indices: Optional[Tuple[int, ...]] = None
+        if loop_indices and self.loop_count > 0:
+            sorted_idx = sorted(int(x) for x in loop_indices)
+            if len(set(sorted_idx)) != len(sorted_idx):
+                raise ValueError(f"loop_indices must be unique, got {loop_indices}.")
+            for i in range(1, len(sorted_idx)):
+                if sorted_idx[i] != sorted_idx[i - 1] + 1:
+                    raise ValueError(f"loop_indices must be consecutive integers, got {sorted_idx}.")
+            if sorted_idx[0] < 0 or sorted_idx[-1] >= num_layers:
+                raise ValueError(f"loop_indices must lie in [0, {num_layers}), got {sorted_idx}.")
+            self.loop_indices = tuple(sorted_idx)
+
+        self.t_embedder = JLTTimestepEmbedder(hidden_size)
+        self.y_embedder = JLTLabelEmbedder(num_classes, hidden_size)
+        self.x_embedder = JLTPatchEmbed(sample_size, patch_size, in_channels, hidden_size, bias=True)
+
+        half_head_dim = hidden_size // num_attention_heads // 2
+        hw_seq_len = sample_size // patch_size
+        self.feat_rope = JLTRotaryEmbedding(dim=half_head_dim, pt_seq_len=hw_seq_len)
+
+        self.blocks = nn.ModuleList(
+            [
+                JLTBlock(
+                    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)
+
+        self.initialize_weights()
+
+    def initialize_weights(self):
+        def _basic_init(module):
+            if isinstance(module, nn.Linear):
+                torch.nn.init.xavier_uniform_(module.weight)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0)
+
+        self.apply(_basic_init)
+
+        w = self.x_embedder.proj.weight.data
+        nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
+        nn.init.constant_(self.x_embedder.proj.bias, 0)
+        nn.init.normal_(self.y_embedder.embedding_table.weight, std=0.02)
+        nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
+        nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
+
+        for block in self.blocks:
+            nn.init.constant_(block.adaLN_modulation.weight, 0)
+            nn.init.constant_(block.adaLN_modulation.bias, 0)
+
+        nn.init.constant_(self.adaLN_modulation_final.weight, 0)
+        nn.init.constant_(self.adaLN_modulation_final.bias, 0)
+        nn.init.constant_(self.linear_final.weight, 0)
+        nn.init.constant_(self.linear_final.bias, 0)
+
+    def _block_schedule(self) -> List[int]:
+        if self.loop_indices is not None and self.loop_count > 0:
+            a, b = self.loop_indices[0], self.loop_indices[-1]
+            return list(range(0, a)) + list(range(a, b + 1)) * self.loop_count + list(range(b + 1, len(self.blocks)))
+        return list(range(len(self.blocks)))
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        timestep: torch.Tensor,
+        class_labels: torch.Tensor,
+        return_dict: bool = True,
+        return_features: bool = False,
+        feature_layers: Optional[List[int]] = None,
+    ):
+        """Predict the clean sample from a noisy latent / image `sample` at JLT time `timestep`."""
+        t = torch.as_tensor(timestep, device=sample.device)
+        if t.ndim == 0:
+            t = t.repeat(sample.shape[0])
+        elif t.ndim == 1 and t.shape[0] == 1 and sample.shape[0] > 1:
+            t = t.repeat(sample.shape[0])
+
+        t_emb = self.t_embedder(t, dtype=sample.dtype)
+        y_emb = self.y_embedder(class_labels).to(dtype=sample.dtype)
+        if t_emb.ndim == 3:
+            y_emb = y_emb.unsqueeze(1)
+        c = t_emb + y_emb
+
+        x = self.x_embedder(sample)
+        if self.training and self.mask_prob > 0.0 and self.mask_ratio > 0.0:
+            batch, num_tokens, _ = x.shape
+            s_mask = torch.rand(batch, device=x.device) < self.mask_prob
+            t_mask = torch.rand(batch, num_tokens, device=x.device) < self.mask_ratio
+            mask = (s_mask.unsqueeze(1) & t_mask).unsqueeze(-1)
+            x = torch.where(mask, self.mask_token.to(x.dtype), x)
+
+        target_layers = set(feature_layers or [])
+        features = {} if return_features else None
+
+        for layer_idx in self._block_schedule():
+            if self.training and self.gradient_checkpointing:
+                x = torch.utils.checkpoint.checkpoint(
+                    self.blocks[layer_idx], x, c, self.feat_rope, use_reentrant=False
+                )
+            else:
+                x = self.blocks[layer_idx](x, c, self.feat_rope)
+            if return_features and layer_idx in target_layers:
+                features[layer_idx] = x
+
+        shift, scale = self.adaLN_modulation_final(self.act_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(x.shape[0], height, width, self.patch_size, self.patch_size, self.out_channels)
+        x = torch.einsum("nhwpqc->nchpwq", x)
+        output = x.reshape(x.shape[0], self.out_channels, height * self.patch_size, width * self.patch_size)
+
+        if return_features:
+            if not return_dict:
+                return (output, features)
+            return Transformer2DModelOutput(sample=output), features
+
+        if not return_dict:
+            return (output,)
+        return Transformer2DModelOutput(sample=output)

JLT-H-1/vae/config.json

@@ -0,0 +1,40 @@
+{
+  "_class_name": "AutoencoderKLFlux2",
+  "_diffusers_version": "0.37.0.dev0",
+  "_name_or_path": "black-forest-labs/FLUX.2-dev",
+  "act_fn": "silu",
+  "batch_norm_eps": 0.0001,
+  "batch_norm_momentum": 0.1,
+  "block_out_channels": [
+    128,
+    256,
+    512,
+    512
+  ],
+  "down_block_types": [
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D"
+  ],
+  "force_upcast": true,
+  "in_channels": 3,
+  "latent_channels": 32,
+  "layers_per_block": 2,
+  "mid_block_add_attention": true,
+  "norm_num_groups": 32,
+  "out_channels": 3,
+  "patch_size": [
+    2,
+    2
+  ],
+  "sample_size": 1024,
+  "up_block_types": [
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D"
+  ],
+  "use_post_quant_conv": true,
+  "use_quant_conv": true
+}

JLT-L-1/pipeline.py

@@ -69,9 +69,11 @@ def _solver_from_scheduler(scheduler) -> str:
 
 class JLTPipeline(DiffusionPipeline):
     r"""
-    Pipeline for class-conditional image generation with JLT (Clean-Latent JiT).
+    Class-conditional JLT sampler.
 
-    Latent models (`in_channels=128`) require a FLUX.2 VAE attached for `output_type="pil"`.
+    Starts from unit Gaussian noise in FLUX.2 latent space (`noise_scale=1.0`) at the
+    trained 256×256 resolution, integrates `JLTScheduler` (Heun, last step Euler),
+    applies interval CFG in velocity space, and decodes FLUX.2 latents when a VAE is attached.
     """
 
     model_cpu_offload_seq = "transformer->vae"
@@ -372,9 +374,8 @@ class JLTPipeline(DiffusionPipeline):
                 ImageNet class id(s) or English label(s).
             guidance_scale (`float`, *optional*):
                 CFG scale. Enabled when `guidance_scale > 1`.
-            noise_scale (`float`, *optional*, defaults to `None`):
-                Std of the initial noise. JiT/JLT scale this with resolution to keep SNR constant
-                (`1.0` at 256, `2.0` at 512, `4.0` at 1024). See huggingface/diffusers#13098.
+            noise_scale (`float`, *optional*, defaults to `1.0`):
+                Std of the initial noise. These checkpoints are trained at 256×256 with unit-scale FLUX.2 latents.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps.
             output_type (`str`, *optional*, defaults to `"pil"`):
@@ -400,13 +401,10 @@ class JLTPipeline(DiffusionPipeline):
         batch_size = len(class_label_ids)
         latent_size = int(self.transformer.config.sample_size)
         channels = int(self.transformer.config.in_channels)
-        # Official JiT uses pixel `sample_size`. JLT latent models store the latent grid.
-        image_size = latent_size * 16 if channels == 128 else latent_size
         null_class_val = int(
             getattr(self.transformer.config, "num_classes", getattr(self.transformer.config, "num_class_embeds", 1000))
         )
-
-        noise_scale = noise_scale if noise_scale is not None else (image_size / 256.0)
+        noise_scale = 1.0 if noise_scale is None else float(noise_scale)
 
         latents = noise_scale * randn_tensor(
             shape=(batch_size, channels, latent_size, latent_size),

JLT-L-1/scheduler/scheduling_jlt.py

@@ -102,6 +102,7 @@ class JLTScheduler(SchedulerMixin, ConfigMixin):
         num_inference_steps: int,
         device: Union[str, torch.device, None] = None,
     ) -> None:
+        """Build the `t: 0 → 1` knot schedule and Heun/Euler step plan."""
         if num_inference_steps < 1:
             raise ValueError("num_inference_steps must be >= 1.")
         if self.config.solver == "heun" and num_inference_steps < 2:
@@ -208,6 +209,7 @@ class JLTScheduler(SchedulerMixin, ConfigMixin):
         prediction_type: Optional[str] = None,
         t_eps: Optional[float] = None,
     ) -> Union[JLTSchedulerOutput, Tuple[torch.Tensor]]:
+        """Advance one scheduled ODE step. `model_output` is clean-`x` or velocity."""
         del generator
         step_index = self._resolve_step_index(timestep)
         if step_index >= len(self._step_plan):

JLT-L-1/transformer/transformer_jlt.py

@@ -1,4 +1,8 @@
-"""JLT (Clean-Latent) Transformer for diffusers."""
+"""JLT transformer: class-conditional clean-latent (or pixel) DiT.
+
+Architecture hyperparameters are read from `config.json` / constructor kwargs.
+Named training presets such as `JiT-B/1` are not part of this module.
+"""
 
 from __future__ import annotations
 
@@ -24,6 +28,8 @@ def rotate_half(x):
 
 
 class JLTRotaryEmbedding(nn.Module):
+    """2D axial RoPE for JLT attention, matching the official JiT implementation."""
+
     def __init__(self, dim, pt_seq_len=16, ft_seq_len=None, theta=10000):
         super().__init__()
         freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
@@ -52,6 +58,8 @@ def modulate(x, shift, scale):
 
 
 class JLTPatchEmbed(nn.Module):
+    """Conv patch embedder for latents (`patch_size=1`) or pixels."""
+
     def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, bias=True):
         super().__init__()
         img_size = (img_size, img_size)
@@ -62,15 +70,12 @@ class JLTPatchEmbed(nn.Module):
         self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias)
 
     def forward(self, x):
-        batch, _, height, width = x.shape
-        if height != self.img_size[0] or width != self.img_size[1]:
-            raise ValueError(
-                f"Input image size ({height}*{width}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
-            )
         return self.proj(x).flatten(2).transpose(1, 2)
 
 
 class JLTTimestepEmbedder(nn.Module):
+    """Sinusoidal timestep MLP used for adaLN conditioning."""
+
     def __init__(self, hidden_size, frequency_embedding_size=256):
         super().__init__()
         self.mlp = nn.Sequential(
@@ -100,6 +105,8 @@ class JLTTimestepEmbedder(nn.Module):
 
 
 class JLTLabelEmbedder(nn.Module):
+    """Class embedding table with one extra null class for CFG."""
+
     def __init__(self, num_classes, hidden_size):
         super().__init__()
         self.embedding_table = nn.Embedding(num_classes + 1, hidden_size)
@@ -110,6 +117,8 @@ class JLTLabelEmbedder(nn.Module):
 
 
 class JLTAttnProcessor:
+    """Official JiT-style attention: split `to_q` / `to_k` / `to_v` + `dispatch_attention_fn`."""
+
     _attention_backend = None
     _parallel_config = None
 
@@ -141,6 +150,8 @@ class JLTAttnProcessor:
 
 
 class JLTAttention(nn.Module, AttentionModuleMixin):
+    """Multi-head self-attention with QK RMSNorm, RoPE, and separate QKV projections."""
+
     _default_processor_cls = JLTAttnProcessor
     _available_processors = [JLTAttnProcessor]
 
@@ -170,6 +181,8 @@ class JLTAttention(nn.Module, AttentionModuleMixin):
 
 
 class JLTSwiGLUFFN(nn.Module):
+    """SwiGLU feed-forward block (`w12` fused, then `w3`)."""
+
     def __init__(self, dim: int, hidden_dim: int, drop=0.0, bias=True) -> None:
         super().__init__()
         hidden_dim = int(hidden_dim * 2 / 3)
@@ -184,6 +197,8 @@ class JLTSwiGLUFFN(nn.Module):
 
 
 class JLTBlock(nn.Module):
+    """Pre-norm transformer block with adaLN-Zero modulation."""
+
     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)
@@ -209,6 +224,24 @@ class JLTBlock(nn.Module):
 
 
 class JLTTransformer2DModel(ModelMixin, ConfigMixin, AttentionMixin):
+    r"""
+    2D transformer for class-conditional clean-latent generation ([JLT](https://arxiv.org/abs/2605.27102)).
+
+    The model predicts the clean sample `x`. Sampling uses [`JLTScheduler`] (`t`: 0 noise → 1 data). Latent
+    checkpoints use FLUX.2 latents (`in_channels=128`, `sample_size` is the latent grid).
+
+    Args:
+        sample_size (`int`, defaults to 16): Spatial size of the input grid.
+        patch_size (`int`, defaults to 1): Patch embed kernel / stride.
+        in_channels (`int`, defaults to 128): Input/output channels.
+        hidden_size (`int`, defaults to 768): Transformer width.
+        num_layers (`int`, defaults to 12): Number of `JLTBlock`s.
+        num_attention_heads (`int`, defaults to 12): Attention heads.
+        mlp_ratio (`float`, defaults to 4.0): SwiGLU expansion.
+        num_classes (`int`, defaults to 1000): ImageNet classes (plus a null CFG class).
+        model_type (`str`, *optional*): Training-name label stored in `config.json` only.
+    """
+
     _supports_gradient_checkpointing = True
     _skip_layerwise_casting_patterns = ["pos_embed", "norm"]
     _no_split_modules = ["JLTBlock"]
@@ -335,6 +368,7 @@ class JLTTransformer2DModel(ModelMixin, ConfigMixin, AttentionMixin):
         return_features: bool = False,
         feature_layers: Optional[List[int]] = None,
     ):
+        """Predict the clean sample from a noisy latent / image `sample` at JLT time `timestep`."""
         t = torch.as_tensor(timestep, device=sample.device)
         if t.ndim == 0:
             t = t.repeat(sample.shape[0])

README.md

@@ -19,7 +19,7 @@ language:
 
 # JLT-diffusers
 
-Native [diffusers](https://github.com/huggingface/diffusers) checkpoints for **JLT** (Clean-Latent Prediction in Latent Diffusion Transformers). Each variant folder is self-contained:
+Native [diffusers](https://github.com/huggingface/diffusers) checkpoints for **JLT** (Clean-Latent Prediction in Latent Diffusion Transformers). Paper: [arXiv:2605.27102](https://arxiv.org/abs/2605.27102). Code: [akatsuki-neo/JLT](https://github.com/akatsuki-neo/JLT). Each variant folder is self-contained:
 
 - `pipeline.py` — `JLTPipeline`
 - `scheduler/scheduling_jlt.py` — `JLTScheduler` (JLT time `t: 0 → 1`, Heun / Euler)
@@ -32,7 +32,7 @@ Load with PyPI `diffusers` and `trust_remote_code=True`. No extra JLT package is
 
 ![JLT-B-1 demo](demo.png)
 
-Class: *golden retriever* (ImageNet 207) — JLT-B/1 at 256×256, 50 Heun steps, `guidance_scale=2.9`, default `noise_scale=image_size/256` (`1.0` at 256), `torch_dtype=bfloat16`, seed 42.
+Class: *golden retriever* (ImageNet 207) — JLT-B/1 at 256×256, 50 Heun steps, `guidance_scale=2.9`, `noise_scale=1.0`, `torch_dtype=bfloat16`, seed 42. These checkpoints are trained and sampled at 256×256 only.
 
 ## Available checkpoints
 
@@ -40,6 +40,7 @@ Class: *golden retriever* (ImageNet 207) — JLT-B/1 at 256×256, 50 Heun steps,
 | --- | --- | --- | --- | ---: |
 | JLT-B/1 | [`JLT-B-1/`](JLT-B-1/) | 256×256 | FLUX.2, patch /1 | 2.9 |
 | JLT-L/1 | [`JLT-L-1/`](JLT-L-1/) | 256×256 | FLUX.2, patch /1 | 2.9 |
+| JLT-H/1 | [`JLT-H-1/`](JLT-H-1/) | 256×256 | FLUX.2, patch /1 | 2.9 |
 
 ## ImageNet class labels
 
@@ -53,17 +54,14 @@ Each variant keeps an English `id2label` map in `model_index.json`.
 ## Inference
 
 ```python
-import importlib.util
-from pathlib import Path
 import torch
+from diffusers import DiffusionPipeline
 
-model_dir = Path("./JLT-B-1")
-spec = importlib.util.spec_from_file_location("jlt_pipeline", model_dir / "pipeline.py")
-mod = importlib.util.module_from_spec(spec)
-spec.loader.exec_module(mod)
-
-pipe = mod.JLTPipeline.from_pretrained(
-    str(model_dir),
+pipe = DiffusionPipeline.from_pretrained(
+    "BiliSakura/JLT-diffusers",
+    subfolder="JLT-B-1",  # or "JLT-L-1" / "JLT-H-1"
+    custom_pipeline="pipeline.py",
+    trust_remote_code=True,
     torch_dtype=torch.bfloat16,
 ).to("cuda")
 
@@ -77,18 +75,13 @@ image = pipe(
 image.save("demo.png")
 ```
 
-Or with native diffusers:
-
-```python
-from pathlib import Path
-from diffusers import DiffusionPipeline
-import torch
+## Citation
 
-model_dir = Path("./JLT-B-1")
-pipe = DiffusionPipeline.from_pretrained(
-    str(model_dir),
-    custom_pipeline=str(model_dir / "pipeline.py"),
-    trust_remote_code=True,
-    torch_dtype=torch.bfloat16,
-).to("cuda")
+```bibtex
+@article{fu2026jlt,
+  title={{JLT}: {C}lean-{L}atent {P}rediction in {L}atent {D}iffusion {T}ransformers},
+  author={Fu, Funing and Wang, Tenghui and Zhou, Guanyu and Cen, Junyong and Zhu, Qichao},
+  journal = {arXiv preprint arXiv:2605.27102},
+  year={2026}
+}
 ```