黑图、彩色噪声、NaN、偏色怎么排查
这篇文章解决什么问题
工作流完成但输出全黑、彩色噪声、花屏、过曝或偏色，先查数值、协议与保存链路，也要核对输入是否本就黑暗。症状只给检查方向，不能单独证明模型或某个算子出了错。先固定 prompt 与 seed，避免不断换输入掩盖同一问题。
先给结论：先判断 latent 是否有效，再检查 VAE协议、dtype、数值范围和颜色/媒体链。
症状分类
第一步：检查日志中的 NaN/Inf
下面是未运行的 PyTorch 诊断提示，适用于真实数值张量。该示意会创建布尔掩码与有限值副本，可能增加显存和同步开销。大张量应按目标实现分块采集，并说明检查覆盖范围；抽样结果不能写成全量统计。它保留原始 NaN/Inf 个数，仅在有限值子集上计算极值；本次没有执行 PyTorch 或 GPU，真实张量统计仍待目标环境采集：
import torch
def tensor_health(name, x):
finite = torch.isfinite(x)
finite_values = x[finite]
count = x.numel()
finite_count = int(finite.sum())
state = ("EMPTY" if count == 0 else "NO_FINITE_VALUES" if finite_count == 0
else "NON_FINITE_PRESENT" if finite_count < count else "FINITE_ONLY")
print(name, "shape=", tuple(x.shape), "dtype=", x.dtype,
"state=", state, "finite=", finite_count, "count=", count,
"finite_min=", float(finite_values.min()) if finite_count else None,
"finite_max=", float(finite_values.max()) if finite_count else None,
"nan=", int(torch.isnan(x).sum()), "inf=", int(torch.isinf(x).sum()))
检查：
- 初始 noise；
- 首帧 latent；
- 每若干采样步的 latent；
- 最终 latent；
- VAE输出。
首次观测到 NaN 的检查点比最终黑图更有定位价值；若只间隔采样，它还不能证明实际首次产生位置。
VAE/latent协议错配
VAE/latent 错配是一个检查方向，本文没有故障样本分布证明它的发生频率。检查：
模型家族
checkpoint版本
latent通道
空间/时间压缩
scale factor
mean/std
帧数对齐
VAE因果padding
现象：
- latent数值看似有限；
- Decode后完全噪声；
- shape通过节点但内容无意义；
- 换回官方VAE立即正常。
不要用图片模型VAE逐帧解视频latent，除非模型明确这样设计。
模型任务错配
以下是待核的组合类型示例，实际兼容性由对应模型版本的官方配置裁决：
- T2V checkpoint接I2V条件节点；
- Wan high/low expert顺序颠倒；
- LTX dev和distilled组件混用；
- Hunyuan原版VAE配1.5；
- Text Encoder版本不匹配；
- LoRA基座不同。
错误可能不报shape异常，因为自定义节点做了适配。最小化到官方模板验证。
dtype与数值溢出
FP16动态范围较小，某些VAE或归一化层可能溢出。排查顺序：
- 保持主DiT设置不变；
- 按目标 VAE 官方支持选择 BF16 以扩大动态范围，或使用 FP32 作数值对照；记录实际生效 dtype 与输出变化；
- 关闭不可靠FP8/INT4转换；
- 关闭fast math或实验kernel；
- 使用官方attention；
- 检查LoRA scale；
- 比较base权重。
PyTorch AMP会按算子选择精度，但第三方节点可能强制cast。1
量化错误
社区量化可能出现：
- 转换脚本不支持某层；
- scale异常；
- embedding/输出层被错误量化；
- high/low expert映射错；
- loader版本不匹配；
- LoRA在线注入失败；
- 量化格式名相同但实现不同。
验证：
官方BF16基线正常?
→ 同workflow只换量化文件
→ 只运行最小规格
→ 检查层缺失/unknown key
→ 比较数值范围
不要同时加入蒸馏LoRA和cache，否则无法定位。
Scheduler/模型参数化错配
若模型预测flow velocity，却被当作epsilon处理，或timestep/sigma映射错误，输出可能保持噪声。常见现象：
- 采样每步数值变化异常小或爆炸；
- steps增加仍是噪声；
- 官方sampler正常，通用sampler异常；
- 低步蒸馏使用错误时间点。
回到官方workflow的sampler、scheduler、shift、steps和CFG。
全黑并不总是NaN
可能是：
- RGB值范围是 [-1,1]，保存器按[0,1]错误截断；
- alpha全0；
- YUV色彩范围错误；
- HDR metadata让播放器显示异常；
- 16-bit帧被8-bit编码器错误读取；
- 通道维顺序错误；
- 图像张量仍在BGR。
分别保存：
- VAE输出Tensor统计；
- 单帧PNG；
- 无压缩或常规H.264 MP4；
- 在不同播放器查看。
若核对的是同一批源帧，PNG 正常而最终 MP4 异常，优先缩小到帧读取、颜色转换、编码、封装和播放这一段。它不能证明具体编码器是根因，也不能排除抽样之外的源帧异常。
偏色排查
红蓝互换
检查 RGB/BGR。
整体灰白
检查归一化、limited/full range、HDR→SDR。
暗部发紫/发绿
检查VAE精度、量化、色彩矩阵和subsampling。
随时间漂色
检查temporal tile、自动调色、VAE时间一致和光源Prompt。
先用标准色块检查保存与颜色转换；若该 VAE 官方支持编码与解码，再做受控 round-trip，并分别保存输入与输出。两种对照不要混写成同一个已执行实验。
Tile接缝与分段错误
若画面空间上固定分区颜色不同，检查：
- spatial tile overlap；
- blending window；
- tile normalization；
- 边界padding。
若某个时间点后整体偏色，检查：
- temporal chunk初始上下文；
- overlap； -因果历史；
- 每段独立归一化；
- 分段写盘色彩设置。
黑图故障树
文件完整性
核对：
- SHA-256；
- 下载大小；
- safetensors header；
- Git LFS是否只下载了pointer；
- 磁盘坏块/下载中断；
- 权重分片是否齐全；
- 文本编码器和VAE路径；
- 模型卡版本。
Git LFS pointer被当权重时通常会在加载阶段报错，但部分自定义loader错误信息不清晰。
最小诊断顺序
- 官方workflow；
- 官方完整精度权重；
- 官方VAE和Text Encoder；
- 最小合法尺寸/帧；
- 无LoRA；
- 无cache；
- 官方sampler；
- 保存单帧PNG；
- 检查tensor数值；
- 再逐项恢复优化。
原始异常值不能被“修好”的统计覆盖
旧的统计方式先调用 nan_to_num，再求最小值和最大值，会把 NaN 替成零、把无穷替成有限极值，读者看到的范围便不再代表原始有限样本。若整组都是 NaN，替换后的范围可能变成零到零，很容易误读成“正常但画面黑”。本稿将两种目的分开：诊断先保留异常个数，有限值范围只从原始有限子集中提取；实际修复或展示需要替换时，另存转换后的数据并说明规则，不能覆盖诊断证据。
空张量、全非有限和混合异常也必须区分。空张量表示该检查点没有元素，范围不可计算；全非有限表示有元素，但没有可用于有限范围的数值；混合异常则既报告异常个数，又报告剩余有限值范围。None 代表没有范围，不是零。有限值全部为零只能证明这些数值没有 NaN/Inf，它可能对应黑画面，也可能是合法中间量，不能命名为“图像健康”。具体范围是否符合 latent、VAE 输出或图像保存协议，需要另查组件契约。
真实 Tensor 诊断还应记 shape、dtype、设备、检查点和运行编号。文本编码器输出、采样 latent 和最终 RGB 的范围不能套用同一个阈值，dtype 改变也不能自动算通过。上面的 PyTorch 函数适用于实数数值张量的诊断示意；它没有处理所有稀疏、复数或第三方容器，不能直接放进任意节点并声称通用。本次仅在下面的纯标量教学例验证有限值统计原则，没有验证真实 Tensor 的设备同步、性能成本或后端兼容性。
最小例：保留 NaN/Inf，再缩小检查边界
这段标准库程序不依赖 PyTorch，不解码 PNG 或 MP4，也不加载模型。health 接收作者构造的普通 int/float 列表；locate 接收完整字典，缺键会 KeyError。输入被限定为本文有限样例，不是用于任意外部 JSON 的产品接口。媒体是否正常是作者填写的观测标记，而不是程序对文件的检查结果。它的价值是把空、全异常、混合异常和有限数值分开，并让媒体段的结论停在“继续检查”。
在 examples 目录运行：
python3 check_numeric_boundary.py
完整程序如下：
"""Finite scalar fixtures; no tensors, image decoding, video or model execution."""
import math
def health(values):
# shortcut: author lists of ordinary int/float scalars, not arbitrary tensor input.
if any(type(v) not in (int, float) for v in values):
raise ValueError("real_scalars_required")
finite = [v for v in values if math.isfinite(v)]
nan_count = sum(math.isnan(v) for v in values)
inf_count = sum(math.isinf(v) for v in values)
state = ("EMPTY" if not values else "NO_FINITE_VALUES" if not finite
else "NON_FINITE_PRESENT" if nan_count or inf_count else "FINITE_ONLY")
return dict(state=state, count=len(values), finite_count=len(finite),
nan_count=nan_count, inf_count=inf_count,
finite_min=min(finite) if finite else None,
finite_max=max(finite) if finite else None)
def locate(record):
for key in ("png_ok", "mp4_ok"):
if record[key] is not None and type(record[key]) is not bool:
raise ValueError("media_observation_must_be_bool_or_none")
latent, decoded = health(record["latent"]), health(record["decoded"])
if latent["state"] in ("EMPTY", "NO_FINITE_VALUES"):
return "INSUFFICIENT_LATENT_FINITE_VALUES"
if latent["state"] == "NON_FINITE_PRESENT":
return "NON_FINITE_AT_RECORDED_LATENT"
if decoded["state"] in ("EMPTY", "NO_FINITE_VALUES"):
return "INSUFFICIENT_DECODED_FINITE_VALUES"
if decoded["state"] == "NON_FINITE_PRESENT":
return "LATENT_TO_DECODED_BOUNDARY_PENDING"
if record["png_ok"] is False:
return "RANGE_OR_SAVE_BOUNDARY_PENDING"
if record["png_ok"] is True and record["mp4_ok"] is False:
return "MEDIA_BOUNDARY_PENDING"
return "NUMERIC_OBSERVATIONS_ONLY"
def fixture(**changes):
row = dict(latent=[-1.0, 0.0, 1.0], decoded=[0.0, 0.5, 1.0],
png_ok=True, mp4_ok=True)
row.update(changes)
return row
def rejects(action, expected):
try:
action()
except ValueError as exc:
assert str(exc) == expected
else:
raise AssertionError("invalid fixture accepted")
def main():
cases = [([], "EMPTY", None, None),
([float("nan")], "NO_FINITE_VALUES", None, None),
([float("inf"), -float("inf")], "NO_FINITE_VALUES", None, None),
([-2.0, 3.0, float("nan"), float("inf")], "NON_FINITE_PRESENT", -2.0, 3.0),
([0.0, 0.0], "FINITE_ONLY", 0.0, 0.0)]
for values, state, low, high in cases:
result = health(values)
assert (result["state"], result["finite_min"], result["finite_max"]) == (state, low, high)
print(f"fixture_health={state} finite={result['finite_count']} nan={result['nan_count']} inf={result['inf_count']} range={low},{high}")
samples = [(fixture(latent=[]), "INSUFFICIENT_LATENT_FINITE_VALUES"),
(fixture(latent=[0.0, float("nan")]), "NON_FINITE_AT_RECORDED_LATENT"),
(fixture(decoded=[]), "INSUFFICIENT_DECODED_FINITE_VALUES"),
(fixture(decoded=[0.0, float("inf")]), "LATENT_TO_DECODED_BOUNDARY_PENDING"),
(fixture(png_ok=False), "RANGE_OR_SAVE_BOUNDARY_PENDING"),
(fixture(mp4_ok=False), "MEDIA_BOUNDARY_PENDING")]
for row, expected in samples:
assert locate(row) == expected
assert locate(fixture()) == "NUMERIC_OBSERVATIONS_ONLY"
rejects(lambda: health([True]), "real_scalars_required")
rejects(lambda: health(["nan"]), "real_scalars_required")
rejects(lambda: locate(fixture(png_ok=1)), "media_observation_must_be_bool_or_none")
rejects(lambda: locate(fixture(mp4_ok="black")), "media_observation_must_be_bool_or_none")
print("fixture_png_ok_mp4_bad=MEDIA_BOUNDARY_PENDING root_cause=UNKNOWN")
print("self_check=16_assertions_passed scope=author_scalar_fixture_only")
if __name__ == "__main__":
main()
本次 CPU 真实执行退出码为 0，标准错误为空。标准输出为：
fixture_health=EMPTY finite=0 nan=0 inf=0 range=None,None
fixture_health=NO_FINITE_VALUES finite=0 nan=1 inf=0 range=None,None
fixture_health=NO_FINITE_VALUES finite=0 nan=0 inf=2 range=None,None
fixture_health=NON_FINITE_PRESENT finite=2 nan=1 inf=1 range=-2.0,3.0
fixture_health=FINITE_ONLY finite=2 nan=0 inf=0 range=0.0,0.0
fixture_png_ok_mp4_bad=MEDIA_BOUNDARY_PENDING root_cause=UNKNOWN
self_check=16_assertions_passed scope=author_scalar_fixture_only
16 个断言包括五类数值样例、六个分层方向、一个有限观测边界与四种拒绝输入。混合样例保留一处 NaN 和一处 Inf，有限范围只来自 -2 和 3，未先替换异常值。全 NaN、全无穷和空列表都没有有限范围；全零列为 FINITE_ONLY，并未宣布画面可用。最后的媒体样例只输出待检查媒体边界和根因未知。没有任何真实生成图片、张量或视频参与这些断言。
按同一运行的中间物定位
先给失败运行一个编号，再将 latent 统计、VAE 输出统计、原生单帧、保存后的 PNG 和最终视频关联到相同编号与帧位置。若预览来自前一轮缓存，视频来自后一轮，它们的差异无法定位媒体段；若只看首帧 PNG，而后半视频全黑，也不能宣布全部源帧正常。至少覆盖首段、异常前后和末段，记录所查帧的时间或索引，不把抽样目检写成全帧通过。
假设记录到 latent 有有限值且不含 NaN/Inf，VAE 输出开始出现 Inf，下一项应检查解码边界的输入协议、精度、权重和相关算子。但这只是两个被记录检查点之间的观察差异；它不能证明一定是 VAE 权重损坏，也不能证明解码前完全没有短暂异常。若检查点是每五步一次，就说明这是一段区间；需要更细定位时，再缩小采集间隔或按目标实现添加检查点，而不是从稀疏日志推断准确的首次产生步骤。
若两个数值检查点都有限，下一项是核范围、通道和保存规则。有限不代表正确：把原本约定的负值当成零到一图像直接截断，可以改变输出；把 BGR 当 RGB 会改变颜色；alpha 为零可能让显示透明。这些变化没有必要制造 NaN。故障单应把原始统计和转换后统计分开，并记变换代码或节点版本；只有成片一张截图，通常不足以区分输入、解码和保存问题。
若同一组 PNG 在相同帧索引下看起来正常，而 MP4 异常，就比较送入编码器的实际文件序列、读取像素格式、颜色转换设置、输出封装和播放器。检查输出分辨率、帧数与帧率，避免把错误序列或不完整文件当作模型黑图。换播放器正常可以支持播放兼容性方向，但它也不证明源文件所有元数据都正确；再次编码正常同样会改变多个条件，仍需记录具体差异，不能省略“待查”两字。
把兼容性核验变成一个最小对照
先确认模型家族、任务类型与实际组件哈希，再核 VAE、文本编码器、scheduler 和节点版本。仓库名字相同、文件名有相同版本数字，不能证明它们构成官方支持的组合；同样，节点不报 shape 错误也不代表 latent 的缩放、均值或时序协议兼容。第三方节点可能补维度或做转换，这会让错误到解码才显现。核验需要官方模板与真实配置并排比，本文没有执行任何模型组合验证。
提高解码精度时只改该段，保留主模型与输入，核实际节点是否按设定运行。BF16 与 FP16 的动态范围和精度特征不同，某个实现允许配置不等于所有层都采用该精度；AMP 的算子选择也不能替第三方强制 cast 的节点保证数值稳定。若改了精度后现象改善，就写出具体配置与观察，再保留原失败日志；不要把一次改善写成“FP16 总是不稳定”或“FP32 已根治”。1
量化对照先回到受支持的完整精度组合，再只更换量化文件或对应加载路径。若量化版本同时需要不同 loader，就把这两项当一个组合差异报告，不能假装只改了权重精度。检查缺失层、unknown key、格式与基座条件；LoRA、缓存、采样器和实验 kernel 在这个对照阶段应固定。只要多项一起变，黑图消失仍无法给其中某项单独归因。本文的统计例也没有验证任何量化算法质量。
时间分块与空间分块不能随意套用。目标 VAE 的因果上下文、overlap 和官方模板决定可以怎么拆；HunyuanVideo-1.5 官方 ComfyUI 指南在本次核验仍对 tiled 解码的 temporal_size 提出明确要求，不能照搬“切得更小总更省”的经验。对接缝或后半变色，应保存分块边界前后源帧、节点配置和编码前后结果，先确定异常在哪一层出现，再决定是否调整 overlap 或时间上下文。2
诊断交接与完成条件
交接包应同时保存：失败配置与组件哈希、原始统计及检查点、被查看的帧索引、原生与最终媒体样本、一次只改变一项的对照表。对照表记退出码、输出是否完整、哪一层首先被观测到异常，以及尚未解释的候选。若暂时只有混合非有限统计，就保留异常个数与有限范围，不自动替换数据后上传一个“修好了”的范围；若只有播放器截图，就承认尚缺原生帧证据。
完成条件由实际需求确定：不是节点跑完就通过，也不是 NaN 个数变为零就通过。至少应确认目标规格完整、输出可读、数值与媒体链约束满足，并由实际内容检查判断是否达到业务质量。标准库自检只验证本例统计与分层逻辑，没有能力满足这些成片条件。没有 NaN、节点跑完或容量检查通过，都不能替代目标成片的完整性与视觉质量验收。
本文结论
黑图、噪声、NaN 和偏色首先是待解释的症状。保留原始非有限统计，再分 latent、解码帧、保存与媒体输出逐层核对；模型协议、数值精度、颜色与编码都是候选。固定输入做单变量对照，才能继续判断原因。
参考资料
Footnotes
- Automatic Mixed Precision package — torch.amp，PyTorch，持续更新，访问日期 2026-10-14。docs.pytorch.org/docs/2.14/a… autocast。 ↩ ↩2
- ComfyUI Usage Guide for HunyuanVideo-1.5，Tencent Hunyuan，2025-11，访问日期 2026-10-14。github.com/Tencent-Hun… ComfyUI 模板、采样配置和 Tiled VAE 注意事项。 ↩
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
What about black maps, colored noises, NAN, color bias?
What's the problem with this article?
workflow is completed but the output is full black, colour noise, screen, exposure or colour bias, first checking values, protocols and preservation links and also checking whether the input is dark. Symptoms are given only in the direction of the examination and cannot be independently shown to have gone wrong in the model or in a particular algorithm. Fix prompt and seed first, and avoid changing input to cover up the same problem.
First to conclude: to judge the validity of the latent, then to check the VAE protocol, dtype, numerical range and colour/media chain.
Classification of symptoms
Step 1: NAN/Inf in the Check Log
The following is the unrun PyTorch diagnostic hint, which applies to the true number scale. This will create a copy of the Boolean mask and limited values, possibly adding VRAM and synchronized expenses. Masses should be collected by segment of the target, with an indication of the scope of the inspection; the results of the sample could not be fully counted. It maintains the original NAN/ Inf number, calculating the polar value only on the limited subset; no PyTorch or GPU is being executed this time, and real coverage statistics remain to be collected in the target environment:
import torch
def tensor_health(name, x):
finite = torch.isfinite(x)
finite_values = x[finite]
count = x.numel()
finite_count = int(finite.sum())
state = ("EMPTY" if count == 0 else "NO_FINITE_VALUES" if finite_count == 0
else "NON_FINITE_PRESENT" if finite_count < count else "FINITE_ONLY")
print(name, "shape=", tuple(x.shape), "dtype=", x.dtype,
"state=", state, "finite=", finite_count, "count=", count,
"finite_min=", float(finite_values.min()) if finite_count else None,
"finite_max=", float(finite_values.max()) if finite_count else None,
"nan=", int(torch.isnan(x).sum()), "inf=", int(torch.isinf(x).sum()))
Inspection:
- Initial;
- first frame late;
- every sample step;
- Eventually, the last;
- VAE output.
The sites first observed in NN were more locational than the final black map; if sampled at intervals, it could not prove the actual first-time location.
VAE/latent protocol mismatch
The VAE/latent mismatch is the direction of the inspection, and the frequency of its occurrence is not supported by the sample distribution. Inspection:
Model family.
Checkpoint version
latent channel
Space/time compression
scale factor
mean/std
Frame Alignment
VAE cause and effect
The phenomenon:
- The value of the slot appears to be limited;
- Full noise after Decode;
- Shape adopts nodes without meaning;
- Back to official VAE immediately normal.
Do not use the image model VAE to decipher the video latent unless the model is clearly designed so.
Model job mismatch
The following are examples of the type of combination to be approved, with actual compatibility determined by the official configuration of the corresponding model version:
- T2V checkpoint to I2V condition node;
- Wan high/low opposite order;
- The combination of LTX dev and distilled components;
- The original version of the VAE with 1.5;
- Text Encoder does not match;
- The Loras are different.
An error may not be reported because the custom nodes fit. Minimize to official template validation.
dtype and value spill
FP16 has a relatively small dynamic range, and some VAEs or subclassification layers may spill over. Question order:
- Maintaining the main DiT settings;
- Selection of BF16 by target VAE official support to expand the dynamic scope or to use FP32 as a numerical comparison; and recording of dtype and output changes in effect;
- Turning off unreliable FP8/INT4 conversion;
- Close the fast path or experimentkernel;
- Use of official attention;
- Check LoRA scale;
- Compare base weight.
PyTorch AMP selects precision by algorithm, but third-party nodes may forcecast. One.
quantization Error
Community quantization may appear:
- The conversion of scripts does not support a layer;
- Scale abnormal;
- Embeding/output layer was wrong quantization;
- High/low error mapping;
- The loader version does not match;
- LoRA online injection failed;
- The name of quantization is the same but different.
Authentication:
The official BF16 baseline is normal?
quantization files only.
Run minimum specifications only
~ Inspection layer missing/unknown key
Comparison range
Do not add distilled LoRA and cache at the same time, otherwise it is impossible to locate.
Scheduler/model parameter mismatch
If the model predicts low development, it is treated as an epsilon, or timestep/sigma map error, the output may maintain noise. Common:
- An abnormally small change or explosion per step in the sample;
- Steps increase is still noise;
- Official sampler normal, general sampler abnormal;
- Low-step distillation using the wrong time point.
It is the first time that I've been able to get back to Sampler, Scheduler, Shift, Steps and CFG in the official workflow.
All black isn't always NAN.
Could be:
- The RGB range is [-1,1], and the saver presses [0,1] for an error;
- Alpha 0;
- Error in the YUV colour range;
- HDR metada to show anomalies on the player;
- The 16-bit frame was misread by the 8-bit encoder;
- An error in routing;
- Images are still in BGR.
Save separately:
- VAE output Tensor statistics;
- Single frame PNG;
- No compression or conventional H.264 MP4;
- Check in different players.
If checked against the same source frame, the PNG normal and final MP4 anomaly is reduced to the frame reading, colour conversion, coding, encapsulation and play. It does not prove that the specific encoder is the root cause, nor does it exclude source frame anomalies other than the sample.
Speculation
Red and blue swap
Check RGB/BGR.
Whole gray
Checking for integration, limited/full Range, HDR→SDR.
Dark purple/green
Check VAE accuracy, quantization, Color Matrix and subsampling.
Floating over time
Checks the temporal file, automatic tuning, VAE time consistency and light source Prompt.
Start with a standard colour block to save with colour conversion; if the VAE officially supports encoding and decoding, then do a controlled round-trip and save input and output separately. The two contrasts should not be combined into the same implemented experiment.
Tile stitches and segment errors
If there is a difference in the colour of the partition in the image space, check:
- spatial tile overlap；
- blending window；
- tile normalization；
- Boundary padding.
If the whole color is coloured after a certain point in time, check:
- initial context;
- Overlap; - causal history;
- Individualization of each paragraph;
- Divisiond disc colour settings.
Black Chart Fault Tree
Document integrity
Check:
- SHA-256；
- Download size;
- safetensors header；
- Is Git LFS just downloading pointer;
- Disk break/down break;
- Whether the weighting is complete;
- Text encoder and VAE path;
- Model card version.
Git LFS pointer is usually reported at the loading stage when he/she is given weight, but the partially defined loader error information is not clear.
Minimum diagnostic order
- Official workflow;
- Officially complete accuracy weights;
- Official VAE and Text Encoder;
- Minimum legal size/frame;
- No Lora;
- No cache;
- Official sampler;
- Preservation of single frame PNG;
- Check the tensor values;
- Retrofitting each item.
Original anomaly values cannot be covered by "fixed" statistics
The old statistical method first calls nan to num, then minimizes and max, replaces NN with zero and infinity with limited extremes, and the reader sees the range no longer representing the original limited sample. If the whole group is NAN, the replacement range could be zero to zero, which could easily be misread as “normal but black”. The present draft distinguishes between two purposes: the diagnosis retains an abnormal number and the limited value range is extracted from the original limited sub-group only; the actual repair or display, where replacement is required, saves converted data and sets out the rules and does not cover diagnostic evidence.
A distinction must also be made between empty space, all non-limited and hybrid anomalies. The blank space indicates that the checkpoint is non-elementary and the range cannot be calculated; the whole is not limited to the number of elements but is not available for the limited range; the mixed anomaly reports both the abnormalities and the remaining limited range. None represents no scope, not zero. All the limited values are zero only to prove that they are not NAN/Inf, which may be a black image or a legitimate intermediate and cannot be named “image health”. The specific range is in accordance with the protocol for the latent, VAE output or image preservation and requires a separate check of the component contract.
The true Tensor diagnosis should also bear in mind the shape, dtype, equipment, checkpoints and operational numbers. The range of text encoder output, sample lott and eventually RGB cannot be applied to the same threshold, and dtype changes cannot be automatically counted. The PyTorch function above applies to diagnostic indications of the true number of numerical scales; it does not treat all thin, complex or third-party containers and cannot be placed directly into any node and claims to be generic. This time only the following pure standard teaching example validates the principle of limited value statistics and does not verify the synchronisation, performance cost or back-end compatibility of the real Tensor equipment.
Minimal example: retain NAN/Inf, then narrow the check borders
This standard library program does not rely on PyTorch, does not decode PNG or MP4, and does not load models. Health receives the normal int/ float list constructed by the author; locate receives the complete dictionary, the missing key is KeyError. Enter a limited sample of this document that is not used for any external JSON product interface. Whether the media is normal is an observation mark filled out by the author, rather than the results of the program ' s examination of the document. Its value is to separate empty, complete, mixed and limited values and to stop the conclusions of the media section from “continue to examine”.
Run in examples directory:
python3 check_numeric_boundary.py
The complete procedure is as follows:
"""Finite scalar fixtures; no tensors, image decoding, video or model execution."""
import math
def health(values):
# shortcut: author lists of ordinary int/float scalars, not arbitrary tensor input.
if any(type(v) not in (int, float) for v in values):
raise ValueError("real_scalars_required")
finite = [v for v in values if math.isfinite(v)]
nan_count = sum(math.isnan(v) for v in values)
inf_count = sum(math.isinf(v) for v in values)
state = ("EMPTY" if not values else "NO_FINITE_VALUES" if not finite
else "NON_FINITE_PRESENT" if nan_count or inf_count else "FINITE_ONLY")
return dict(state=state, count=len(values), finite_count=len(finite),
nan_count=nan_count, inf_count=inf_count,
finite_min=min(finite) if finite else None,
finite_max=max(finite) if finite else None)
def locate(record):
for key in ("png_ok", "mp4_ok"):
if record[key] is not None and type(record[key]) is not bool:
raise ValueError("media_observation_must_be_bool_or_none")
latent, decoded = health(record["latent"]), health(record["decoded"])
if latent["state"] in ("EMPTY", "NO_FINITE_VALUES"):
return "INSUFFICIENT_LATENT_FINITE_VALUES"
if latent["state"] == "NON_FINITE_PRESENT":
return "NON_FINITE_AT_RECORDED_LATENT"
if decoded["state"] in ("EMPTY", "NO_FINITE_VALUES"):
return "INSUFFICIENT_DECODED_FINITE_VALUES"
if decoded["state"] == "NON_FINITE_PRESENT":
return "LATENT_TO_DECODED_BOUNDARY_PENDING"
if record["png_ok"] is False:
return "RANGE_OR_SAVE_BOUNDARY_PENDING"
if record["png_ok"] is True and record["mp4_ok"] is False:
return "MEDIA_BOUNDARY_PENDING"
return "NUMERIC_OBSERVATIONS_ONLY"
def fixture(**changes):
row = dict(latent=[-1.0, 0.0, 1.0], decoded=[0.0, 0.5, 1.0],
png_ok=True, mp4_ok=True)
row.update(changes)
return row
def rejects(action, expected):
try:
action()
except ValueError as exc:
assert str(exc) == expected
else:
raise AssertionError("invalid fixture accepted")
def main():
cases = [([], "EMPTY", None, None),
([float("nan")], "NO_FINITE_VALUES", None, None),
([float("inf"), -float("inf")], "NO_FINITE_VALUES", None, None),
([-2.0, 3.0, float("nan"), float("inf")], "NON_FINITE_PRESENT", -2.0, 3.0),
([0.0, 0.0], "FINITE_ONLY", 0.0, 0.0)]
for values, state, low, high in cases:
result = health(values)
assert (result["state"], result["finite_min"], result["finite_max"]) == (state, low, high)
print(f"fixture_health={state} finite={result['finite_count']} nan={result['nan_count']} inf={result['inf_count']} range={low},{high}")
samples = [(fixture(latent=[]), "INSUFFICIENT_LATENT_FINITE_VALUES"),
(fixture(latent=[0.0, float("nan")]), "NON_FINITE_AT_RECORDED_LATENT"),
(fixture(decoded=[]), "INSUFFICIENT_DECODED_FINITE_VALUES"),
(fixture(decoded=[0.0, float("inf")]), "LATENT_TO_DECODED_BOUNDARY_PENDING"),
(fixture(png_ok=False), "RANGE_OR_SAVE_BOUNDARY_PENDING"),
(fixture(mp4_ok=False), "MEDIA_BOUNDARY_PENDING")]
for row, expected in samples:
assert locate(row) == expected
assert locate(fixture()) == "NUMERIC_OBSERVATIONS_ONLY"
rejects(lambda: health([True]), "real_scalars_required")
rejects(lambda: health(["nan"]), "real_scalars_required")
rejects(lambda: locate(fixture(png_ok=1)), "media_observation_must_be_bool_or_none")
rejects(lambda: locate(fixture(mp4_ok="black")), "media_observation_must_be_bool_or_none")
print("fixture_png_ok_mp4_bad=MEDIA_BOUNDARY_PENDING root_cause=UNKNOWN")
print("self_check=16_assertions_passed scope=author_scalar_fixture_only")
if __name__ == "__main__":
main()
This CPU real execution exit code is 0, standard error is empty. The standard output is:
fixture_health=EMPTY finite=0 nan=0 inf=0 range=None,None
fixture_health=NO_FINITE_VALUES finite=0 nan=1 inf=0 range=None,None
fixture_health=NO_FINITE_VALUES finite=0 nan=0 inf=2 range=None,None
fixture_health=NON_FINITE_PRESENT finite=2 nan=1 inf=1 range=-2.0,3.0
fixture_health=FINITE_ONLY finite=2 nan=0 inf=0 range=0.0,0.0
fixture_png_ok_mp4_bad=MEDIA_BOUNDARY_PENDING root_cause=UNKNOWN
self_check=16_assertions_passed scope=author_scalar_fixture_only
The claims included five numerical samples, six layers, a limited observation boundary and four denials of input. The hybrid sample retains one NN and one Inf with a limited range from -2 and 3, without replacing the anomaly first. All NAN, all infinity and empty lists have no limited scope; all are listed as FINITE ONLY and the images are not declared available. The final media sample will only be exported pending inspection of media boundaries and causes unknown. There is no real generation of pictures, banners or videos to participate in these assertions.
Orient the same run intermediate
Runs a number for failure, then links the last, VAE output, original frame, saved PNG and final video to the same number and frame position. If the preview comes from the previous round of cache, the video comes from the latter and their differences cannot locate the media segment; if only first frame PNG is seen, the latter half is completely black, the entire frame of origin cannot be declared normal. At least cover the front, back, back and end of the line, record the time or index of the frame, and do not write the sample in full frame.
Assuming that a limited value is recorded in the latent and does not contain NAN/Inf, VAE output begins to appear in the Inf, the next item should check the input protocol, precision, weight and related algorithms that decode the boundary. However, this was only an observation difference between the two recorded checkpoints; it could not prove that the VAE weights had been damaged or that there had been no transient anomalies before the decoded. If the checkpoint is every five steps, this means that it is a segment; if more detailed positioning is needed, the interval between collections is reduced or additional checkpoints are achieved by target, rather than extrapolating from the thin logs to the correct first-time-generated step.
If both numerical check points are limited, the next is the nuclear scope, access and preservation rules. Limited does not mean correct: using the originally agreed negative value as zero to one image is a direct cut that changes the output; using BGR as RGB changes the colour; alpha may make the display transparent. There is no need for these changes to make NAN. The failure sheet should separate the original and post-conversion statistics and record the version of the variable code or node; only a single screenshot is usually insufficient to distinguish between input, decoding and preservation problems.
If the same set of PNGs appears normal in the same frame index, and MP4 is abnormal, the actual file sequence, the pixel format, the colour conversion settings, the output envelope and the player are compared. Check the output of resolution, frame numbers and frame rates to avoid using the wrong sequence or incomplete file as a model black map. The replayer normally supports the play-compatibility orientation, but it does not support the correctness of all metadata in the source file; the re-encoding will also change many conditions, and specific differences will still need to be recorded and the word "to be checked" cannot be omitted.
Make compatibility check a minimum contrast.
The model family, the type of task and the actual component Hashi were first identified, and the VAE, text encoder, scheduler and nodal versions were re-established. The fact that the warehouse has the same name and file names have the same version of numbers does not prove that they constitute an officially supported combination; similarly, the failure to report shape errors does not represent the compatibility of the zoom, average or time-series agreements of the latent. Third-party nodes may complement or convert, which will allow errors to be decoded. An official template is required to match the actual configuration, and no model combination validation is performed.
Only the paragraph is changed to increase the accuracy of the decoder, to preserve the main model and input, and whether the actual nuclear node is running as set. BF16 Unlike the dynamic scope and accuracy characteristics of FP16, a certain enabling configuration does not equal the accuracy of all layers; AMP's alcalculator selection does not force the cast node to stabilize the value for a third party. If there is an improvement in precision, specific configurations and observations are written and the original log of failure is maintained; no one improvement is written as “FP16 is always unstable” or “FP32 is cured”. One.
quantization returns first to the supported complete accuracy group and replaces only quantization files or corresponding load paths. If the quantization version requires a different loader at the same time, the two are reported as a combination of differences, and it cannot be assumed that only weighting accuracy has been changed. Check the missing layer, unknown key, format and base conditions; LoRA, cache, sampler and experiment Kernel should be fixed during this control phase. As long as there are multiple variations, the disappearance of the black map cannot give a single attribution to one of them. The statistics do not verify the quality of any quantization algorithm.
Time and space segments cannot be used at random. The cause and effect of target VAE context, overlap and the official template determine how it can be removed; the HunyuanVideo-1.5 official ComfyUI guide still makes a clear request for tiled decoded temporal size in this verification, and cannot replicate the experience of “tougher total savings”. The result of the source frame, node configuration and coding of the fractional boundary should be preserved for the seams or the later half of the colour, first to determine at which level the anomaly occurs and then to decide whether to adjust overlap or time context. Two.
Diagnostic handover and completion conditions
The handing over package should be kept simultaneously: failed configuration with component Hashi, raw statistics and check points, frame index viewed, original and final media sample, and one-stop control table. The cross-checkout code, the completeness of the output, which layer was first observed unusual, and the unexplained candidate. For the time being, only a mix of non-limited statistics would preserve the anomaly and the limited range, without automatically replacing the data and uploading a “fixed” range; if only a screenshot of the player, it would admit that original frame evidence is not available.
The conditions for completion are determined by actual needs: it is not through running the nodes, it is not through the NAN to zero. At a minimum, the target specification should be confirmed as complete, output-readable, numerically and media chain-bound, with actual content checking to determine whether the quality of the operation is achieved. The SDS self-inspection only validates the statistical and hierarchical logic of the case and does not have the capacity to meet these partial conditions. Without NAN, nodal run or capacity clearance, there is no substitute for the integrity and visual mass of the target piece acceptance check.
Conclusion
Black maps, noise, NN and colour bias are first and foremost symptoms to be explained. Retain the original non-limited statistics, split the slots, decode the frame, save and check the media output layer by layer; model protocols, numerical precision, colours and codes are all candidates. Fixed input is used to compare the single variable to continue to determine the cause.
References
Footnotes
- Automatic Mixed Precision Package - Torch.amp, PyTorch, continuously updated, dates 2026-10-14. Docs.pytorch.org/docs/2.14/a... autocast. ↩2
- CommyUI Usage Guide for HunyuanVideo-1.5, Tencent Hunyuan, 2025-11, dates of access 2026-10-14. Gythub.com/Tencent-Hun... ComfyUI template, sample configuration and Tiled VAE attention. Zenium
Auto-fetched and translated by AI Atlas. All rights belong to the original author.