Messed up images since trying 2511/v18, what am I doing wrong?
I get images like the above now. With v11 and all the versions before that I didn't get this issue. Also, qwen image edit 2511 works fine for me, when I switch to v18 though I get this glitched output. What am I doing wrong? I use the clip, vae, and safetensor that came within the checkpoint. I also use
- steps 4
- cfg 1
- eula ancestral, beta
- textencodeqwenimageeditplus
Unlike the recommended workflow here, the comfyui native one for 2511 uses fluxkontextmultireferencelatent and some other nodes like modelsamplingauraflow (set to 3) and cfgnorm (set to 1). I bypassed those just in case they were causing the issue, still no dice.
I am on pytorch 2.8.0+cu128, using sage attention and triton installed if those even matter. I also use distorch2/multigpu, but I believe this occurs even with multigpu turned off (using simple checkpoint loader instead)
Same thing happening to me
Try downloading comfyui portable and see if it is a problem with your install.
Same issue! The solution for me was to REMOVE all custom nodes from my comfy portable setup. A bunch of files were recently updated in comfy. Some listed below.
llama.cpython-312.pyc
clip_model.cpython-312.pyc
clip_vision.cpython-312.pyc
llama.py
clip_vision.py
clip_model.py
And others! For me, some of the custom nodes / modifications don't play well with the recent updates.
Backup a copy of your custom nodes, remove them, then re-add them in batches. Was damn frustrating when my images were coming out great then turned into a hyper weave of colors after restarting comfy!
Using torch 2.7.1+cu128, sageattention patch from KJ's nodes, windows-triton, distorch2/multigpu and transformers 4.57.3.
Same issue! The solution for me was to REMOVE all custom nodes from my comfy portable setup. A bunch of files were recently updated in comfy. Some listed below.
llama.cpython-312.pyc
clip_model.cpython-312.pyc
clip_vision.cpython-312.pyc
llama.py
clip_vision.py
clip_model.py
And others! For me, some of the custom nodes / modifications don't play well with the recent updates.
Backup a copy of your custom nodes, remove them, then re-add them in batches. Was damn frustrating when my images were coming out great then turned into a hyper weave of colors after restarting comfy!Using torch 2.7.1+cu128, sageattention patch from KJ's nodes, windows-triton, distorch2/multigpu and transformers 4.57.3.
I tried it myself, but it didn't work. :(
I had this same issue but I don't think my solution would help your problem. I'm posting it anyway just in case though. On the bright side, it did encourage me to explore undervolting and overclocking (while undervolted) my 5090 GPU. In my case, when I suddenly got random noise like OP posted, I found that I had to downclock my core clock slightly (about -50 MHz offset from default) to get stable outputs again. In my case I assume it's some kind of minor degradation on my GPU, unless there is actually some kind of rare bug happening. I did try fresh install too, no luck though, only worked fine after I reduced my core clock speed on the 5090 GPU. I had my GPU power limited too, before now undervolting it.
But it did encourage me to explore overclocking while undervolted, which has allowed me to keep virtually the same performance as before but with just over half the power consumption and less heat produced.
Infixed my issue by removing the --fast flag from my start up script
Was just popping in to say my launch args include --bf16-vae --bf16-text-enc, no --fast or fp16 accumulation. It was stable for me with just --bf16-vae but just today had to include --bf16-text-enc. Something got updated again and mucked it up!