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

[Sexhaver19]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes 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

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

[rattus128]

When you say installing old versions of ComfyUI are you doing a git rewind or are you starting fresh with self contained old version portables?

How far back did you go?

If you have been doing git rewind, try a clean old version windows portable from where you think the regression point might be as that rules out issues in the python dependencies.

[Sexhaver19]

I tested old versions with self-contained portables. I've tested 0.13.0 and 0.11.1, (as well as 2-3 of the most recent releases), which none of which seemed to show any difference.

[serious-44]

I've experienced similar issues before. There were two solutions: plug in a different GPU or swap in different memory sticks. I still don't understand it. The error only occurred with AI, but not with games or diagnostic programs.

[rattus128]

Yeah this is hard for it to be on the software as your result disprove software regression. This is most likely hardware failure. Undo any overclocking or undervolting you have in play if that is a thing. You can try maybe underclocking the GPU if possible or reducing thermal thresholds to something in the 50s just to test theories. Hardware errors can easily be AI specific or even usage pattern specific just in bad luck.

[Sexhaver19]

It makes sense that it could be something to do with the hardware setup I've got, which would explain why the issue persists across multiple installs and different OS. It doesn't explain why the issue has suddenly appeared recently though, why I was able to generate stuff without issue a few weeks ago.

I've tried swapping the slots of my RAM, it may have made a minor improvement to the resolution I can generate at with WAN models? Though I can't say for certain, as I do still get black generations at higher resolutions.
It doesn't seem to have made a difference with SDXL gens, I still get grey/black very consistently, unless I use the --disable-smart-memory arg which improves it vastly.
Unfortunately I don't have any other RAM sticks or GPUs to test with.

I also tried underclocking my GPU, doesn't seem to make any noticeable difference. I also attempted underclocking my GPU memory, bad idea - everything breaks at even lower resolutions (though its still possible to generate, but the limit seems to be around 400x400 for 1 SDXL image.

I don't understand entirely what the --disable-smart-memory command does, from my understanding it reduces the amount of VRAM being used. Does that mean the problem is more likely something with my VRAM?
Are there any args or tests that I should be running in order to help narrow down the cause of the issue?
Maybe there's some settings in NVIDIA control panel that might affect things? Or in BIOS? I'm really open to trying whatever because I wanna try and fix this.

[rattus128]

It makes sense that it could be something to do with the hardware setup I've got, which would explain why the issue persists across multiple installs and different OS. It doesn't explain why the issue has suddenly appeared recently though, why I was able to generate stuff without issue a few weeks ago.

I've tried swapping the slots of my RAM, it may have made a minor improvement to the resolution I can generate at with WAN models? Though I can't say for certain, as I do still get black generations at higher resolutions. It doesn't seem to have made a difference with SDXL gens, I still get grey/black very consistently, unless I use the --disable-smart-memory arg which improves it vastly. Unfortunately I don't have any other RAM sticks or GPUs to test with.

I also tried underclocking my GPU, doesn't seem to make any noticeable difference. I also attempted underclocking my GPU memory, bad idea - everything breaks at even lower resolutions (though its still possible to generate, but the limit seems to be around 400x400 for 1 SDXL image.

I don't understand entirely what the --disable-smart-memory command does, from my understanding it reduces the amount of VRAM being used. Does that mean the problem is more likely something with my VRAM? Are there any args or tests that I should be running in order to help narrow down the cause of the issue? Maybe there's some settings in NVIDIA control panel that might affect things? Or in BIOS? I'm really open to trying whatever because I wanna try and fix this.

--disable-smart-memory with no noticable regression path does possibly suggest a VRAM failure. Disable smart memory stops comfy from keeping multiple models in VRAM ever, even when there is space, so its possible its avoiding bad VRAM for you. --novram is a good test as that stops it using VRAM for model load completely (it still uses your GPU for compute). Big perf penalty on SDXL but might give more evidence.

Tried a graphics benchmark and any tearing etc? I dont really know how well they exercise the full VRAM range but they might prove something.

[Sexhaver19]

Well, I guess my VRAM is just messed up then. Running --novram seems to fix basically all the issues (although it creates new ones in that it breaks a bunch of custom nodes). Additionally, I tried running OCCT VRAM test and it was spitting out errors like crazy. Gonna look into getting the GPU replaced under warranty as I've only had the GPU for a couple of months.

The thing that still irks me though is that this is a problem that seemingly popped out of nowhere. One day, everything was genning fine, the next day I'm getting trouble with every gen. Is it possible that maybe pushing my GPU too hard could cause these errors? I haven't exactly been doing this extensively, probably like 10 mins gen time each day on average.
Another thing of concern is how some attempted fixes did actually fix the problem temporarily, like enabling live preview for the first time fixed things entirely for a short period of time. It just feels like I'm missing something here, or there's potentially something causing the GPU VRAM to mess up. I don't want to mess up the replacement GPU too if that's the case.

[PureRussianVodka]

Im having the exact same issue, astral 5090.

Although im not using comfy, im using swarmUI.

I need to ask an important question, does this happen AFTER you change a prompt? because for me i can generate forever until a prompt is changed in anyway.

this is the error i get

comfy\ComfyUI\comfy\utils.py:1024: RuntimeWarning: invalid value encountered in cast
[ComfyUI-0] [STDERR] images = [Image.fromarray(np.clip(255. * image.cpu().numpy(), 0, 255).astype(np.uint8)) for image in samples]

[Sexhaver19]

Im having the exact same issue, astral 5090.

Although im not using comfy, im using swarmUI.

I need to ask an important question, does this happen AFTER you change a prompt? because for me i can generate forever until a prompt is changed in anyway.

this is the error i get

comfy\ComfyUI\comfy\utils.py:1024: RuntimeWarning: invalid value encountered in cast [ComfyUI-0] [STDERR] images = [Image.fromarray(np.clip(255. * image.cpu().numpy(), 0, 255).astype(np.uint8)) for image in samples]

I have had this issue before, I mostly noticed it with Klein and WAN models, never with SDXL. It would be fine until I change the prompt, or changing the reference image in img2img.

[giandre]

Adding a detailed data point in case it helps narrow this down — I've been fighting the Qwen black-image / NaN issue for two days and ran a pretty exhaustive set of tests. TL;DR: for Qwen the NaN is produced during sampling (the latent itself comes back NaN), it correlates exactly with comfy_kitchen being installed, and it is not a hardware/VRAM problem in my case.

Setup: RTX 4090 (Ada sm_89) · Win11 portable · ComfyUI v0.24.0 (also v0.23/v0.22) · comfy_kitchen 0.2.10 · torch reproduced on both 2.10.0+cu130 and 2.12.0+cu130 · model qwen_image_edit_2509_fp8_e4m3fn.

Fingerprint: black output (luma ≈16); log ends model weight dtype float8_e4m3fn, manual cast: bfloat16 → Requested to load WanVAE → RuntimeWarning: invalid value encountered in cast (nodes.py:1657). Sampling completes with no error; black appears at decode/save.

Timeline: generated these exact images perfectly the morning of June 3; ComfyUI auto-updated and installed comfy_kitchen at 10:46 AM; every gen since is black.