Comfy generates noise after cancelling Qwen Image Edit Nunchaku generation

[MoreColors123]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Expected Behavior

Followup generations in Nunchaku Qwen Image Edit are working normally, after interrupting a generation.

Actual Behavior

Since the update i did on 24th of October for Comfy Portable (edit: also with Desktop Comfy) I have an issue whenever I do this:

1. Generate a pic with Nunchaku Qwen Image Edit
2. Interrupt the process
   -> Any following Nunchaku Qwen Image Edit generation will be noise (see below), and also the final picture is noise, not just the preview.

The only solution I found is to click the "clear models and node cache" button on the upper right between generations - which increases generation time a lot, of course.
And reverting to 0.3.65 helps too, still I don't understand why.

This does NOT happen with any other regular QWEN-Models, also not with the QWEN Edit FP8 Version or any other models at all. It even doesn't happen with Nunchaku Flux Kontext. Just with every Nunchaku Qwen Image Edit i have, and which worked perfectly before. The Workflow used is basically the Template Workflow for Qwen Image Edit, I just replaced the model loader with the Nunchaku one.

Steps to Reproduce

1. Generate a pic with Nunchaku Qwen Image Edit
2. Interrupt the process
   Problem: Any following Nunchaku Qwen Image Edit generation will be noise

Debug Logs

got prompt
Using pytorch attention in VAE
Using pytorch attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
Using scaled fp8: fp8 matrix mult: False, scale input: False
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load WanVAE
loaded completely 2307.8046875 242.02829551696777 True
Requested to load QwenImageTEModel_
loaded completely 13243.646704483031 7909.737449645996 True
loaded completely 13243.646704483031 7909.737449645996 True
model weight dtype torch.bfloat16, manual cast: None
model_type FLUX
Enabling CPU offload
Requested to load NunchakuQwenImage
 25%|█████████████████████████████████████▊                                                                                                                 | 2/8 [00:03<00:09,  1.60s/it]Interrupting prompt c4c29a20-716d-46c4-8614-29cd1f927f9e
 25%|█████████████████████████████████████▊                                                                                                                 | 2/8 [00:03<00:10,  1.72s/it]
Processing interrupted
Prompt executed in 23.99 seconds
got prompt
 12%|██████████████████▉                                                                                                                                    | 1/8 [00:01<00:11,  1.60s/it]Interrupting prompt 1b931387-9d97-4f32-bc3f-b87168645a08
 12%|██████████████████▉                                                                                                                                    | 1/8 [00:02<00:20,  2.87s/it]
Processing interrupted
Prompt executed in 2.89 seconds

Other

No response

[jovan2009]

I have found out that Qwen is very sensitive to some settings or optimizations. For me simply starting ComfyUI with "--fast" option gave noise as output in Qwen. This was weeks ago, I didn't try Qwen lately with later ComfyUI versions, perhaps it works now, but my starting ComfyUI_Qwen.bat in which I removed "--fast" from the starting command is still on my desktop as a reminder.

[rattus128]

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)

Expected Behavior

Followup generations in Nunchaku Qwen Image Edit are working normally, after interrupting a generation.

Actual Behavior

Since the update i did on 24th of October for Comfy Portable (edit: also with Desktop Comfy) I have an issue whenever I do this:

1. Generate a pic with Nunchaku Qwen Image Edit

2. Interrupt the process
   -> **Any following Nunchaku Qwen Image Edit generation will be noise** (see below), and also the final picture is noise, not just the preview.

The only solution I found is to click the "clear models and node cache" button on the upper right between generations - which increases generation time a lot, of course. And reverting to 0.3.65 helps too, still I don't understand why.

The behaviour suggests your loaded model weights is corrupted. Your workflow has custom model loader in play so its hard to debug. Can you attach a full workflow?

Also these noise bugs are very sensitive to some of the comfy startup options if you have any. Can you describe how you start comfy?

What is your GPU?

[rattus128]

I have found out that Qwen is very sensitive to some settings or optimizations. For me simply starting ComfyUI with "--fast" option gave noise as output in Qwen. This was weeks ago, I didn't try Qwen lately with later ComfyUI versions, perhaps it works now, but my starting ComfyUI_Qwen.bat in which I removed "--fast" from the starting command is still on my desktop as a reminder.

Did you have --async-offload in that --fast noise case? I'm interested in workflows if you can still reproduce.

I fixed a noise corruption in --async-offload this week that may have helped you (#10501)

[jovan2009]

I have found out that Qwen is very sensitive to some settings or optimizations. For me simply starting ComfyUI with "--fast" option gave noise as output in Qwen. This was weeks ago, I didn't try Qwen lately with later ComfyUI versions, perhaps it works now, but my starting ComfyUI_Qwen.bat in which I removed "--fast" from the starting command is still on my desktop as a reminder.

Did you have --async-offload in that --fast noise case? I'm interested in workflows if you can still reproduce.

I fixed a noise corruption in --async-offload this week that may have helped you (#10501)

No, --async-offload was on my "black list" before that because it was messing the output in WanVideoWraper workflows so it definitely wasn't on my Qwen workflow, if my memory serves me. I mean I would expected to have noise output with --async-offload, the surprise was that --fast was also doing it with Qwen.

I just tried to see what is the current situation with a Qwen edit 2509 + 4 steps Lora workflow on latest git pulled ComfyUI. With --fast fp16_accumulation fp8_matrix_mult pinned_memory and --async-offload enabled it works fine, no issue. (I don't use autotune in --fast options because I understood there is a bug in torch or cuda that makes it use a large amount of VRAM. And cublas_ops I can't build on a torch newer than 2.8.0 and I am currently on nightly 2.10.0+cu130).

[rattus128]

I have found out that Qwen is very sensitive to some settings or optimizations. For me simply starting ComfyUI with "--fast" option gave noise as output in Qwen. This was weeks ago, I didn't try Qwen lately with later ComfyUI versions, perhaps it works now, but my starting ComfyUI_Qwen.bat in which I removed "--fast" from the starting command is still on my desktop as a reminder.

Did you have --async-offload in that --fast noise case? I'm interested in workflows if you can still reproduce.
I fixed a noise corruption in --async-offload this week that may have helped you (#10501)

No, --async-offload was on my "black list" before that because it was messing the output in WanVideoWraper workflows so it definitely wasn't on my Qwen workflow, if my memory serves me. I mean I would expected to have noise output with --async-offload, the surprise was that --fast was also doing it with Qwen.

You shouldn't expect any noise with --async-offload as it's not any different in the math. It just tries to order things better for efficiency. You can always cut a bug report if async-offload degrades output quality.

I just tried to see what is the current situation with a Qwen edit 2509 + 4 steps Lora workflow on latest git pulled ComfyUI. With --fast fp16_accumulation fp8_matrix_mult pinned_memory and --async-offload enabled it works fine, no issue. (I don't use autotune in --fast options because I understood there is a bug in torch or cuda that makes it use a large amount of VRAM. And cublas_ops I can't build on a torch newer than 2.8.0 and I am currently on nightly 2.10.0+cu130).

Nice! Let me know any other interesting results positive or negative.

[MoreColors123]

The behaviour suggests your loaded model weights is corrupted. Your workflow has custom model loader in play so its hard to debug. Can you attach a full workflow?

I attached a workflow. I don't think the model is corrupted though, as I can run it with version 0.3.65.

Also these noise bugs are very sensitive to some of the comfy startup options if you have any. Can you describe how you start comfy?

I always start it with the original run_nvidia_gpu.bat and without any flags.

What is your GPU?

4070 TI super 16 GB

Thank you for your help. I assume it has to do with the new method to fast interrupt a generation. Is that possible?

nunchaku bug.json

[hoorf]

I have same problem in generate image , check your comfyui start command ,remove --fast or close this option

[driderr]

Not using any extra ComfyUI flags, I get this issue every single time I cancel a run using the unmodified example workflow. Clicking on ComfyUI's "Unload Models" and "Free Model" restore normal functionality.
Edit: Disabling cpu offloading for the main model "fixes" the issue. I can cancel with no issues with that disabled.

[Apacchi88]

I have the same issue. Problem with disabling cpu offloading is if you are running multiple loras it causes OOM errors if you have the average user hardware.

I found that modifying the ops.py to remove the faster workflow canceling feature added on 14 Oct fixes the issue.

Perhaps the faster workflow canceling feature can be added as toggle within comfy's settings? Or hopefully it can be fixed