File size: 8,223 Bytes
2552774 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 | #!/usr/bin/env python3
"""Queue MOSS-Image prompts against the local ComfyUI (http://127.0.0.1:8188).
python3 moss.py "a gorilla in a nightclub..." --seeds 1 2 3
python3 moss.py -f prompts.txt --enhance --lora r256
"""
import argparse, copy, json, os, sys, time, urllib.request
HERE = os.path.dirname(os.path.abspath(__file__))
URL = "http://127.0.0.1:8188"
OUT = os.path.join(HERE, "..", "..", "output")
MODELS = {
"base": "moss_image_base_bf16.safetensors",
"sft": "moss_image_sft_bf16.safetensors",
}
LORAS = {
"b21": "moss_image_sft_lora_blocks_21_to_49_rank_512_bf16.safetensors",
"r256": "moss_image_sft_lora_rank_256_bf16.safetensors",
"r512": "moss_image_sft_lora_rank_512_bf16.safetensors",
}
def call(path, data=None):
req = urllib.request.Request(URL + path, data=json.dumps(data).encode() if data is not None else None,
headers={"Content-Type": "application/json"})
with urllib.request.urlopen(req) as r:
return json.load(r)
def build(base, prompt, seed, a):
g = copy.deepcopy(base)
g["1"]["inputs"]["unet_name"] = MODELS.get(a.model, a.model)
if a.swap_layers is not None or a.swap_other:
# base model + exact donor weights for chosen layers/parts (LoRA chain is skipped)
g["40"] = {"class_type": "UNETLoader", "inputs": {"unet_name": MODELS.get(a.donor, a.donor), "weight_dtype": "default"}}
g["41"] = {"class_type": "MOSSBlockSwap", "inputs": {"model": ["1", 0], "donor": ["40", 0], "layers": a.swap_layers or "",
"parts": a.swap_parts, "include_other": a.swap_other, "ratio": a.swap_ratio}}
g["16"]["inputs"]["model"] = ["41", 0]
del g["30"]
elif a.lora == "none":
g["16"]["inputs"]["model"] = ["1", 0]
del g["30"]
else:
g["30"]["inputs"]["lora_name"] = LORAS.get(a.lora, a.lora)
g["30"]["inputs"]["strength_model"] = a.strength
if a.lora2:
g["31"] = {"class_type": "LoraLoaderModelOnly", "inputs": {"model": g["16"]["inputs"]["model"],
"lora_name": LORAS.get(a.lora2, a.lora2), "strength_model": a.strength2}}
g["16"]["inputs"]["model"] = ["31", 0]
if a.enhance:
g["17"]["inputs"]["prompt"] = prompt
g["17"]["inputs"]["sampling_mode.seed"] = seed
else:
g["5"]["inputs"]["text"] = prompt
for k in ("17", "18", "19"):
del g[k]
if a.probe:
g["42"] = {"class_type": "MOSSStripeProbe", "inputs": {"model": g["16"]["inputs"]["model"], "head_from": MODELS.get(a.model, a.model),
"tag": f"{a.probe}_s{seed}", "width": a.width, "height": a.height}}
g["16"]["inputs"]["model"] = ["42", 0]
if a.no_tsr:
g["2"]["inputs"]["model"] = g["16"]["inputs"]["model"]
del g["16"]
if a.no_slg:
for k in ("20", "22"):
g[k]["inputs"]["model"] = g["2"]["inputs"]["model"]
del g["2"]
if a.sampler != "er_sde":
g["21"] = {"class_type": "KSamplerSelect", "inputs": {"sampler_name": a.sampler}}
if a.negative is not None:
g["6"]["inputs"]["text"] = a.negative
g["20"]["inputs"].update(noise_seed=seed, cfg=a.cfg)
g["22"]["inputs"]["steps"] = a.steps
g["7"]["inputs"].update(width=a.width, height=a.height)
g["10"]["inputs"]["filename_prefix"] = "moss/" + a.prefix
if a.window_steps:
window(g, a)
return g
def window(g, a):
"""Apply --window-lora only for sampling steps a..b (1-indexed) via ComfyUI hook keyframes on both conds.
--window-steps "1-N" applies it for the whole run; strength 0 everywhere is a no-op control."""
lo, hi = (int(v) for v in a.window_steps.split("-"))
n = a.steps
g["70"] = {"class_type": "CreateHookLoraModelOnly", "inputs": {"lora_name": a.window_lora, "strength_model": a.window_strength}}
kf = ([(0.0, 0.0)] if lo > 1 else []) + [((lo - 1) / n, 1.0)] + ([(hi / n, 0.0)] if hi < n else [])
prev = None
for i, (pct, mult) in enumerate(kf):
node = {"class_type": "CreateHookKeyframe", "inputs": {"strength_mult": mult, "start_percent": pct}}
if prev:
node["inputs"]["prev_hook_kf"] = [prev, 0]
prev = str(71 + i)
g[prev] = node
g["75"] = {"class_type": "SetHookKeyframes", "inputs": {"hooks": ["70", 0], "hook_kf": [prev, 0]}}
for cid, src in (("76", "5"), ("77", "6")):
g[cid] = {"class_type": "ConditioningSetProperties", "inputs": {"cond_NEW": [src, 0], "strength": 1.0, "set_cond_area": "default", "hooks": ["75", 0]}}
g["20"]["inputs"]["positive"], g["20"]["inputs"]["negative"] = ["76", 0], ["77", 0]
def main():
p = argparse.ArgumentParser()
p.add_argument("prompt", nargs="*")
p.add_argument("-f", "--file", help="one prompt per paragraph (blank-line separated)")
p.add_argument("--seeds", type=int, nargs="+", default=[16])
p.add_argument("--enhance", action="store_true", help="rewrite the prompt with Qwen3.5 first")
p.add_argument("--model", default="sft", help="base | sft | <filename>")
p.add_argument("--lora", default="b21", help="b21 | r256 | none | <filename>")
p.add_argument("--strength", type=float, default=-0.5)
p.add_argument("--donor", default="sft", help="model whose exact weights are swapped in (with --swap-layers)")
p.add_argument("--swap-layers", help='layers to take from --donor, e.g. "21-41" or "3,4"; enables block swap')
p.add_argument("--swap-parts", default="all", help="all | comma list of qkv,o,ffn_up_gate,ffn_down,norm")
p.add_argument("--swap-other", action="store_true", help="also swap non-layer weights (embedders, modulation, head)")
p.add_argument("--swap-ratio", type=float, default=1.0)
p.add_argument("--probe", help="record per-layer stripe spectra to output/moss/probe/<probe>_s<seed>.npz")
p.add_argument("--window-steps", help='e.g. "5-12": apply --window-lora only for these sampling steps (1-indexed)')
p.add_argument("--window-lora", default="moss_image_sft_banding_fix_L37-41_bf16.safetensors")
p.add_argument("--window-strength", type=float, default=1.0)
p.add_argument("--lora2", help="second LoRA chained after the first, e.g. --lora r512 --strength 1 --lora2 b21 --strength2 -0.5")
p.add_argument("--strength2", type=float, default=1.0)
p.add_argument("--no-tsr", action="store_true")
p.add_argument("--no-slg", action="store_true")
p.add_argument("--sampler", default="er_sde", help="er_sde (workflow's SamplerER_SDE) or any KSampler sampler name")
p.add_argument("--steps", type=int, default=25)
p.add_argument("--cfg", type=float, default=3.0)
p.add_argument("--width", type=int, default=1024)
p.add_argument("--height", type=int, default=1024)
p.add_argument("--negative")
p.add_argument("--prefix", default="moss")
a = p.parse_args()
prompts = [" ".join(a.prompt)] if a.prompt else []
if a.file:
prompts += [x.strip() for x in open(a.file).read().split("\n\n") if x.strip()]
if not prompts:
p.error("give a prompt or -f file")
base = json.load(open(os.path.join(HERE, "moss_image_t2i_api.json")))
jobs = []
for prompt in prompts:
for seed in a.seeds:
r = call("/prompt", {"prompt": build(base, prompt, seed, a)})
if r.get("node_errors"):
sys.exit(json.dumps(r["node_errors"], indent=1))
jobs.append((r["prompt_id"], prompt, seed))
print(f"queued {len(jobs)}", flush=True)
for pid, prompt, seed in jobs:
while True:
h = call(f"/history/{pid}").get(pid)
if h and h.get("status", {}).get("completed") is not None:
break
time.sleep(1)
if h["status"].get("status_str") != "success":
print(f"FAILED seed {seed}: {json.dumps(h['status'].get('messages', [])[-1:])[:800]}")
continue
outs = h["outputs"]
for img in outs.get("10", {}).get("images", []):
print(os.path.normpath(os.path.join(OUT, img["subfolder"], img["filename"])), f"seed={seed}")
if "19" in outs:
print(" enhanced:", outs["19"]["text"][0][:1500])
if __name__ == "__main__":
main()
|