Frequently black/grey images and NaN latent errors across multiple models. Persists after factory reset. Affects Windows and Linux systems.

### Custom Node Testing

- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Your question

Field | Value
-- | --
ComfyUI Version | 0.17.2
ComfyUI Frontend | v1.41.20
ComfyUI Manager | V3.39.2
OS | Windows 32-bit (win32)
Python Version | 3.13.11 (Dec 5 2025) [MSC v.1944 64-bit AMD64]
PyTorch Version | 2.10.0+cu130
GPU | NVIDIA GeForce RTX 5080 (15.92 GB VRAM)
RAM | 61.6 GB

Generations across multiple model types (SDXL, Wan 2.1/2.2, Flux2 Klein 9b fp8) frequently produce black/grey images, NaN latent errors, or heavily distorted/noisy outputs. This is a **regression** - these issues were not present previously with the same hardware. The problem persists across clean installs, driver changes, and even after a full PC factory reset. It also persists across WSL2 and even dual-boot ubuntu installs.

## Issues by Model

### SDXL

- Generating at **1024×1024 or higher** frequently results in completely black or grey images (not 100% of the time)
- In a batch of 4 images, typically only **1 or 2 renders correctly** - the rest are black/grey
- Reducing resolution reliably decreases the error rate
- Reproduced in both **ComfyUI** and **sd-webui-forge-neo**

ComfyUI will *occasionally*, not consistently, report an error like this: ```ComfyUI_windows_portable\ComfyUI\nodes.py:1664: RuntimeWarning: invalid value encountered in cast
img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8))```

### Flux.2 Klein 9b fp8

- Generally works, but **~10% of generations** exhibit obvious distortion or random latent noise
- Issue is exacerbated by LORAs
- Occasionally enters a state where it **only generates black/grey images** - restarting ComfyUI usually resolves this
- Previously, Klein would *only* work with zero LORAs active; this was resolved after a full PC factory reset, but the other issues persist.

### Wan 2.2
- Once I reach a certain threshold with resolution and/or length, my generations either become super glitchy/noisy or just completely black.
- The threshold before it becomes too distorted is around 480x480 at length 49. Length at 53 becomes noisy, and 57 or higher turns black.
- These numbers are variable depending on the res and length, lower res = longer length and vice versa.

### Logs

```powershell
At launch logs:

comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_mxfp8', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'scaled_mm_mxfp8', 'scaled_mm_nvfp4']}
Found comfy_kitchen backend cuda: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'scaled_mm_nvfp4']}
Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': "ImportError: No module named 'triton'", 'capabilities': []}
Checkpoint files will always be loaded safely.
Total VRAM 16303 MB, total RAM 63081 MB
pytorch version: 2.10.0+cu130
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 5080 : cudaMallocAsync
Using async weight offloading with 2 streams
Enabled pinned memory 28386.0
working around nvidia conv3d memory bug.
Using pytorch attention
aimdo: src-win/cuda-detour.c:77:INFO:aimdo_setup_hooks: found driver at 00007FFDEFE30000, installing 4 hooks
aimdo: src-win/cuda-detour.c:61:DEBUG:install_hook_entrys: hooks successfully installed
aimdo: src/control.c:69:INFO:comfy-aimdo inited for GPU: NVIDIA GeForce RTX 5080 (VRAM: 16302 MB)
DynamicVRAM support detected and enabled
Python version: 3.13.11 (tags/v3.13.11:6278944, Dec 5 2025, 16:26:58) [MSC v.1944 64 bit (AMD64)]
ComfyUI version: 0.17.2
comfy-aimdo version: 0.2.10
comfy-kitchen version: 0.2.8
ComfyUI frontend version: 1.41.20

Log for first Wan2.2 gen (480x480 length 57, black video):

Using pytorch attention in VAE
Using pytorch attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
Found quantization metadata version 1
Using MixedPrecisionOps for text encoder
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load WanTEModel
Model WanTEModel prepared for dynamic VRAM loading. 6419MB Staged. 0 patches attached. Force pre-loaded 49 weights: 392 KB.
Model WanTEModel prepared for dynamic VRAM loading. 6419MB Staged. 0 patches attached. Force pre-loaded 49 weights: 392 KB.
Requested to load WanVAE
Model WanVAE prepared for dynamic VRAM loading. 242MB Staged. 0 patches attached. Force pre-loaded 52 weights: 28 KB.
Found quantization metadata version 1
Detected mixed precision quantization
Using mixed precision operations
model weight dtype torch.float16, manual cast: torch.float16
model_type FLOW
WARNING: No VAE weights detected, VAE not initalized.
no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.
Requested to load WAN21
Model WAN21 prepared for dynamic VRAM loading. 13853MB Staged. 0 patches attached.
100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 3.31it/s]
Found quantization metadata version 1
Detected mixed precision quantization
Using mixed precision operations
model weight dtype torch.float16, manual cast: torch.float16
model_type FLOW
WARNING: No VAE weights detected, VAE not initalized.
no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.
Requested to load WAN21
0 models unloaded.
Model WAN21 prepared for dynamic VRAM loading. 13853MB Staged. 0 patches attached.
100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:02<00:00, 1.07s/it]
0 models unloaded.
Model WanVAE prepared for dynamic VRAM loading. 242MB Staged. 0 patches attached. Force pre-loaded 52 weights: 28 KB.
Prompt executed in 63.09 seconds

Log for second wan2.2 gen (same settings apart from noise_seed, seemingly perfect non-glitchy video)

got prompt
Model WanVAE prepared for dynamic VRAM loading. 242MB Staged. 0 patches attached. Force pre-loaded 52 weights: 28 KB.
Model WAN21 prepared for dynamic VRAM loading. 13853MB Staged. 0 patches attached.
100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 3.36it/s]
0 models unloaded.
Model WAN21 prepared for dynamic VRAM loading. 13853MB Staged. 0 patches attached.
100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 3.69it/s]
0 models unloaded.
Model WanVAE prepared for dynamic VRAM loading. 242MB Staged. 0 patches attached. Force pre-loaded 52 weights: 28 KB.
Prompt executed in 24.73 seconds
```

### Other

## What Has Been Tried

### No Effect

- Installing older versions of ComfyUI
- Re-downloading all models, VAEs, and text encoders
- Clean reinstall of NVIDIA GPU drivers using DDU
- Installing older NVIDIA driver versions
- Switching from Game Ready to Studio drivers
- Updating AMD chipset drivers
- Closing all background applications
- Disabling GPU scheduling
- Disabling hardware acceleration in browser
- Disabling windowed optimizations
- Removing all external CUDA/PyTorch/Python installs outside the portable/venv directory
- Running ComfyUI with various precision/memory flags: `--force-fp32`, `--force-fp16`, `--fp16-vae`, `--bf16-vae`, `--split-cross-attention`, `--quad-cross-attention`, `--pytorch-cross-attention`, `--disable-xformers`, `--gpu-only`, `--lowvram`
- Running ComfyUI inside WSL2
- Running ComfyUI on a dual-boot Ubuntu install
- Full PC factory reset *(partially helped — see below)*

### Partially or Temporarily Effective

- Running VAE on CPU | Reliably fixes SDXL black image issue, but is far too slow for Wan or Klein (and slow for SDXL in practice)
- `--disable-smart-memory` | Appeared to improve Wan 2.2 temporarily; no longer has any noticeable effect. Consistently improves SDXL black image generation
- Enabling live preview (TAESD or latent) | The first time I tried this it completely resolved the issue until the next PC restart - no longer effective even after toggling
- Full PC factory reset | Fixed the "Klein refuses to run with any LORA" issue specifically; did not fix the other issues

## Key Concerns

1. **A full PC factory reset did not resolve the core issues** - suggesting this is not a software environment problem
2. **The issue persists inside WSL2 and Ubuntu installs**, likely ruling out most Windows-specific or driver-level causes
3. **Some fixes worked completely, then stopped working** after a restart (e.g. live preview enabling) - suggesting something non-deterministic or state-dependent is occurring
4. **No issues occur in other VRAM/CUDA-intensive software** (e.g. Blender rendering), making hardware defects unlikely
5. The issue is a **regression** - identical hardware and workflows that previously ran without problems are now consistently failing