Files
admin 7753939a3b feat: ai-manju skill 初始提交
AI 漫剧流水线技能完整落地:文件驱动、可校验、单镜可重做。

七阶段:基线资产 → 策划叙事 → 分镜中枢 → 配音先行 → 出图/出视频 → 缝合 → 审片归档。

- scripts/: manju.py 编排、submit_t2i/submit_i2v 提交、cleanup_comfyui 清理
- assets/comfyui-manju-cleanup/: ComfyUI 清理扩展,归属校验 fail-closed
- references/: models.md 模型节点图全表、workflow/director/storyboard/pitfalls
- .gitignore: 排除 __pycache__ 与运行产物
2026-11-14 17:49:39 +08:00

409 lines
17 KiB
Python
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
#!/usr/bin/env python3
"""manju T2I:内网 MiniMax H3-Image 文生图提交器。
零第三方依赖。读 config/pipeline.json 的 imageModel,输出 PNG 到
work/<EP>/images/<SHOT>.png —— 正是 tools/submit_i2v.py 期望的首帧路径。
服务端明文 HTTP,必须绕过本机代理(见 references/models.md §3.2)。
用法:
python3 tools/submit_t2i.py --ep EP01 --shot S01 --prompt-file 40_prompts/S01.t2i.txt
python3 tools/submit_t2i.py --prompt "A lone figure on a misty pier, ..." --out assets/characters/ling.png
python3 tools/submit_t2i.py --ep EP01 --shot S03 --width 4096 --height 4096 --force
约定:给 --ep/--shot 时,未给 --prompt-file 则读
projects/<EP>/40_prompts/<SHOT>.t2i.txt(出图提示词,与 .i2v.txt 运动提示词分开留档)。
★ 权重自适应(2026-11-13)
------------------------
不同端点的 H3-Image 量化权重名不同,写死必然踩坑:
DCU 192.168.1.24:30120 -> minimax_h3_image_pruned_w4a8.safetensors
K8s 192.168.1.30:30113 -> minimax_h3_image_pruned_int8_convrot.safetensors
所以本脚本**不写死权重名**:提交前读目标端点的 /object_info/<node>,
按优先级挑一个真正存在的(见 UNET_PRIORITY)。端点换权重、或将来新增第三个
端点,不用改代码就能跑。
⚠️ 两条线的坑完全不同,别互相照搬:
- DCU(ROCm gfx928):**不要用 int8_convrot** —— convrot 的 rint 有 bug → NaN → 纯黑图。
- K8s(CUDA sm_75):**没有 w4a8**(w4a8 的 CUDA 加速路径要求 sm_80),只有 int8_convrot。
该端点跑在 4Gi cgroup 硬限里,靠 22 处内存墙补丁 + 显存换入换出才跑得动,
详见 references/models.md §2.1。
"""
import argparse
import hashlib
import json
import os
import sys
import time
import uuid
import urllib.parse
import urllib.request
DEFAULTS = {
"endpoint": "http://192.168.1.30:30113",
"width": 2048,
"height": 2048,
"steps": 6,
"seed": 42,
"prefix": "H3IMG",
"timeout": 3600,
}
# 权重候选,按优先级。实际用哪个由目标端点 /object_info 决定。
UNET_PRIORITY = [
"minimax_h3_image_pruned_w4a8.safetensors", # DCU 30120
"minimax_h3_image_pruned_int8_convrot.safetensors", # K8s 30113
]
CLIP_PRIORITY = [
"qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
]
VAE_PRIORITY = [
"minimax_h3_image_vae_fp16.safetensors",
]
# 本机代理会把明文 HTTP 响应改写成 text/html,必须显式绕过
opener = urllib.request.build_opener(urllib.request.ProxyHandler({}))
def pick_weight(host, node_type, priority):
"""从端点的 /object_info/<node> 里挑一个真实存在的权重名。
端点换权重不用改代码 —— 这是本脚本能在 DCU / K8s 两套基础设施上通吃的关键。
探不到就回退到 priority[0],让服务端报 node_errors(比静默猜错好)。
"""
try:
d = get(host, "/object_info/" + node_type, timeout=20)
node = d.get(node_type) or {}
req = ((node.get("input") or {}).get("required") or {})
available = []
for key in ("unet_name", "clip_name", "vae_name"):
v = req.get(key)
if isinstance(v, list) and v and isinstance(v[0], list):
available = list(v[0])
break
for name in priority:
if name in available:
return name
# 端点上有别的 H3-Image 权重(将来新增档位)—— 认前缀兜底
for name in available:
if name.startswith("minimax_h3_image"):
return name
return priority[0]
except Exception:
return priority[0]
def find_root():
"""工作空间根:MANJU_ROOT > cwd 向上找 config/pipeline.json > 脚本上级。"""
env = os.environ.get("MANJU_ROOT")
if env:
return os.path.abspath(os.path.expanduser(env))
d = os.getcwd()
while True:
if os.path.isfile(os.path.join(d, "config", "pipeline.json")):
return d
nd = os.path.dirname(d)
if nd == d:
break
d = nd
return os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
def load_cfg(root):
cfg = dict(DEFAULTS)
p = os.path.join(root, "config", "pipeline.json")
if os.path.isfile(p):
cfg.update(json.load(open(p, encoding="utf-8")).get("imageModel", {}))
return cfg
def get(host, path, timeout=60):
return json.loads(opener.open(host + path, timeout=timeout).read())
def post(host, path, payload, timeout=60):
req = urllib.request.Request(host + path, data=json.dumps(payload).encode(),
headers={"Content-Type": "application/json"})
return json.loads(opener.open(req, timeout=timeout).read())
def build(prompt, width, height, steps, seed, prefix, unet, clip, vae):
"""实机 /object_info 核对:width/height step=32;vae、images 为 optional。
unet/clip/vae 由 pick_weight() 从目标端点探出来,不写死。
"""
return {
"1": {"class_type": "UNETLoader",
"inputs": {"unet_name": unet, "weight_dtype": "default"}},
"2": {"class_type": "CLIPLoader",
"inputs": {"clip_name": clip, "type": "minimax", "device": "default"}},
"3": {"class_type": "VAELoader", "inputs": {"vae_name": vae}},
"4": {"class_type": "MiniMaxH3ImageConditioning",
"inputs": {"clip": ["2", 0], "vae": ["3", 0], "prompt": prompt,
"width": width, "height": height}},
"5": {"class_type": "ConditioningZeroOut", "inputs": {"conditioning": ["4", 0]}},
"6": {"class_type": "KSampler",
"inputs": {"model": ["1", 0], "seed": seed, "steps": steps, "cfg": 1.0,
"sampler_name": "euler", "scheduler": "simple",
"positive": ["4", 0], "negative": ["5", 0],
"latent_image": ["4", 1], "denoise": 1.0}},
"7": {"class_type": "VAEDecode", "inputs": {"samples": ["6", 0], "vae": ["3", 0]}},
"8": {"class_type": "SaveImage", "inputs": {"images": ["7", 0], "filename_prefix": prefix}},
}
def wait(host, pid, timeout):
t0 = time.time()
gaps = 0
while True:
try:
d = get(host, "/history/" + urllib.parse.quote(pid), timeout=30)
except Exception as e:
gaps += 1
print(" ...服务暂不可达 (%d): %s" % (gaps, str(e)[:90]), flush=True)
if gaps >= 8 or time.time() - t0 > timeout:
raise TimeoutError("等待出图超时或服务重启")
time.sleep(8)
continue
gaps = 0
item = d.get(pid)
if item:
st = item.get("status", {})
errs = [m for m in st.get("messages", []) if m[0] == "execution_error"]
if errs:
raise RuntimeError("渲染失败: " + json.dumps(errs, ensure_ascii=False)[:2000])
if st.get("completed") and item.get("outputs"):
return item
if time.time() - t0 > timeout:
raise TimeoutError("等待出图超时 %s 秒" % timeout)
time.sleep(5)
def find_image(item):
for out in (item.get("outputs") or {}).values():
for f in out.get("images", []) or []:
if isinstance(f, dict) and f.get("filename"):
return f
raise RuntimeError("未找到图片输出: " + json.dumps(item.get("outputs"), ensure_ascii=False)[:1200])
def download(host, f, out):
q = urllib.parse.urlencode({"filename": f["filename"],
"subfolder": f.get("subfolder", ""),
"type": f.get("type", "output")})
raw = opener.open(host + "/view?" + q, timeout=300).read()
os.makedirs(os.path.dirname(os.path.abspath(out)) or ".", exist_ok=True)
with open(out, "wb") as fh:
fh.write(raw)
return len(raw)
def sanity_check(path):
"""出图自检:全黑 / 全白 / 单色 一律判失败。
为什么必须有这一步(2026-11-13 实测踩过):
H3-Image 在量化档位不对时会**出图成功但内容是纯黑/单色** ——
ComfyUI 侧 status.completed=True、outputs 有文件,不报错。
不自检就会把废图当成成品记进台账,后面审片才发现。
只用标准库读 PNG:解 IHDR 拿尺寸,采样像素算颜色数/亮度/标准差。
"""
try:
with open(path, "rb") as fh:
data = fh.read()
if data[:8] != b"\x89PNG\r\n\x1a\n":
return "not-png"
# PNG IHDR: 8 字节签名 + 4 长度 + 4 类型 + 8 宽高
w = int.from_bytes(data[16:20], "big")
h = int.from_bytes(data[20:24], "big")
# 解 IDAT(PNG 用 zlib deflate)
import zlib
idat = b""
i = 8
while i < len(data) - 8:
ln = int.from_bytes(data[i:i + 4], "big")
typ = data[i + 4:i + 8]
if typ == b"IDAT":
idat += data[i + 8:i + 8 + ln]
elif typ == b"IEND":
break
i += 12 + ln
if not idat:
return "no-idat"
raw = zlib.decompress(idat)
# 假设 8bit RGB/RGBA,逐行反滤波(filter 0..4)
bpp = 4 if len(raw) >= h * (1 + w * 4) else 3
stride = w * bpp
prev = bytearray(stride)
colors = set()
bright_sum = 0.0
bright_sq = 0.0
cnt = 0
pos = 0
for _y in range(h):
ft = raw[pos]
pos += 1
line = bytearray(raw[pos:pos + stride])
pos += stride
if ft == 1:
for x in range(bpp, stride):
line[x] = (line[x] + line[x - bpp]) & 0xFF
elif ft == 2:
for x in range(stride):
line[x] = (line[x] + prev[x]) & 0xFF
elif ft == 3:
for x in range(stride):
a = line[x - bpp] if x >= bpp else 0
line[x] = (line[x] + ((a + prev[x]) >> 1)) & 0xFF
elif ft == 4:
for x in range(stride):
a = line[x - bpp] if x >= bpp else 0
c = prev[x - bpp] if x >= bpp else 0
b = prev[x]
p = a + b - c
pa, pb, pc = abs(p - a), abs(p - b), abs(p - c)
pr = a if (pa <= pb and pa <= pc) else (b if pb <= pc else c)
line[x] = (line[x] + pr) & 0xFF
prev = line
# 采样:每 8 行取一行,避免大图太慢
for x in range(0, stride, bpp * 7):
px = (line[x], line[x + 1], line[x + 2])
colors.add(px)
v = (px[0] * 299 + px[1] * 587 + px[2] * 114) / 1000.0
bright_sum += v
bright_sq += v * v
cnt += 1
if cnt == 0:
return "empty"
mean = bright_sum / cnt
std = (max(0.0, bright_sq / cnt - mean * mean)) ** 0.5
# 判定:颜色数极少 或 标准差极小 → 近乎纯色(黑/白/单色)
if len(colors) <= 8:
return "single-color(%d colors)" % len(colors)
if std < 3.0:
return "flat(std=%.2f, %d colors)" % (std, len(colors))
return "ok(%dx%d, %d colors, mean=%.1f, std=%.1f)" % (w, h, len(colors), mean, std)
except Exception as e:
return "check-failed(%s)" % str(e)[:60]
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--ep", default="", help="集号,如 EP01")
ap.add_argument("--shot", default="", help="镜号,如 S01")
ap.add_argument("--prompt", default="")
ap.add_argument("--prompt-file", default="")
ap.add_argument("--out", default="", help="输出 PNG;默认 work/<EP>/images/<SHOT>.png")
ap.add_argument("--width", type=int, default=0)
ap.add_argument("--height", type=int, default=0)
ap.add_argument("--steps", type=int, default=0)
ap.add_argument("--seed", type=int, default=-1)
ap.add_argument("--timeout", type=int, default=0)
ap.add_argument("--force", action="store_true", help="已存在也重跑")
ap.add_argument("--cleanup", action="store_true",
help="导出后清理 ComfyUI 的 history/queue 记录与产物(等价于 cleanup.auto=true)")
ap.add_argument("--no-cleanup", action="store_true",
help="本次强制不清理,覆盖 config 里的 cleanup.auto")
a = ap.parse_args()
root = find_root()
cfg = load_cfg(root)
host = str(cfg["endpoint"]).rstrip("/")
full = json.load(open(os.path.join(root, "config", "pipeline.json"), encoding="utf-8")) \
if os.path.isfile(os.path.join(root, "config", "pipeline.json")) else {}
do_cleanup = (a.cleanup or full.get("cleanup", {}).get("auto", False)) and not a.no_cleanup
pf = a.prompt_file
if not pf and a.ep and a.shot and not a.prompt:
pf = os.path.join(root, "projects", a.ep, "40_prompts", a.shot + ".t2i.txt")
prompt = open(pf, encoding="utf-8").read().strip() if pf else a.prompt.strip()
if not prompt:
return "需要 --prompt / --prompt-file,或备好 projects/<EP>/40_prompts/<SHOT>.t2i.txt"
width = a.width or int(cfg["width"])
height = a.height or int(cfg["height"])
steps = a.steps or int(cfg["steps"])
seed = a.seed if a.seed >= 0 else int(cfg["seed"])
timeout = a.timeout or int(cfg["timeout"])
if width % 32 or height % 32:
return "宽高必须 32 对齐(当前 %dx%d)" % (width, height)
if a.out:
out = a.out
elif a.ep and a.shot:
out = os.path.join(root, "work", a.ep, "images", a.shot + ".png")
else:
out = os.path.join(root, "h3_image_out.png")
h8 = hashlib.sha256(prompt.encode("utf-8")).hexdigest()[:8]
if os.path.exists(out) and os.path.getsize(out) > 100000 and not a.force:
print("SKIP 已存在 %s(--force 可重跑)" % out, flush=True)
return 0
sysinfo = get(host, "/system_stats", timeout=20).get("system", {})
unet = pick_weight(host, "UNETLoader", UNET_PRIORITY)
clip = pick_weight(host, "CLIPLoader", CLIP_PRIORITY)
vae = pick_weight(host, "VAELoader", VAE_PRIORITY)
print("ComfyUI %s @ %s | %dx%d %d步 seed=%d | promptHash=%s" %
(sysinfo.get("comfyui_version", "?"), host, width, height, steps, seed, h8), flush=True)
print("weights unet=%s clip=%s vae=%s" % (unet, clip, vae), flush=True)
graph = build(prompt, width, height, steps, seed, str(cfg["prefix"]), unet, clip, vae)
t0 = time.time()
# client_id 会被 ComfyUI 持久化在 history 项 prompt[3],是清理时的归属凭证
cid = "manju-t2i-" + uuid.uuid4().hex[:8]
r = post(host, "/prompt", {"prompt": graph, "client_id": cid})
if r.get("node_errors"):
raise RuntimeError("节点错误: " + json.dumps(r["node_errors"], ensure_ascii=False)[:3000])
pid = r["prompt_id"]
print("[submit] prompt_id=%s -> %s" % (pid, out), flush=True)
item = wait(host, pid, timeout)
size = download(host, find_image(item), out)
verdict = sanity_check(out)
bad = not verdict.startswith("ok")
if bad:
# 废图(纯黑/单色)照样落地,但要显式报警 —— 不能让「completed」等于「可用」
print("[SANITY] FAIL %s -> %s" % (out, verdict), file=sys.stderr, flush=True)
print("[SANITY] 出图成功但内容异常。H3-Image 在量化档位不对时会静默出纯色图:"
"DCU 禁用 int8_convrot;K8s(30113) 只有 int8_convrot 可用。"
"先看 models.md §2.1 的量化档位表,别急着重跑。", file=sys.stderr, flush=True)
else:
print("[SANITY] %s -> %s" % (out, verdict), flush=True)
log_path = os.path.join(os.path.dirname(os.path.abspath(out)), "t2i_runs.json")
runs = json.load(open(log_path, encoding="utf-8")) if os.path.exists(log_path) else {}
key = (a.ep + "/" + a.shot) if (a.ep and a.shot) else os.path.basename(out)
runs[key] = {"out": os.path.relpath(out, root), "promptHash": h8, "prompt": prompt,
"model": unet, "endpoint": host, "width": width, "height": height,
"steps": steps, "seed": seed, "promptId": pid, "clientId": cid,
"bytes": size, "sanity": verdict, "sanityOK": not bad,
"seconds": round(time.time() - t0, 1),
"at": time.strftime("%Y-%m-%d %H:%M:%S")}
with open(log_path, "w", encoding="utf-8") as f:
json.dump(runs, f, ensure_ascii=False, indent=2)
print("[ok] %.2f MB in %.1fs" % (size / 1048576, time.time() - t0), flush=True)
if do_cleanup and bad:
# 废图不清服务端 —— 保留 history 与产物当排查证据(量化档位/权重问题要看原始产物)
print("[cleanup] SKIP 自检未通过,保留服务端记录与产物供排查", flush=True)
elif do_cleanup:
try:
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from cleanup_comfyui import after_export
print("[cleanup] " + after_export(root, pid, host, log_path, key, out, cid),
flush=True)
except Exception as e:
# 成片已安全落地;清理失败只提示,绝不影响主流程
print("[cleanup] FAILED %s(图片已落地,可稍后跑 cleanup_comfyui.py 补清)"
% e, file=sys.stderr, flush=True)
return 1 if bad else 0
if __name__ == "__main__":
try:
sys.exit(main())
except Exception as e:
print("ERROR: %s" % e, file=sys.stderr, flush=True)
sys.exit(1)