Shape '[1, 52, 78, 16, 2, 2]' is invalid for input of size 519168 when using index_timestep_zero to fix image quality
hello, u y sure that u a using our model and model? it looks quilt different from our workflow
hello, u y sure that u a using our model and model? it looks quilt different from our workflow
With that workflow using the lighting lora also happens, the index_timestep_zero option works when loading the model in gguf or fp8 and fp16 formats, the first image is the input image and the second one is the output that has the workflow:
The official comfyui workflows of qwen 2511 also have the edit model reference method nodes here: https://docs.comfy.org/tutorials/image/qwen/qwen-image-edit-2511
ComfyUI Error Report
Error Details
- Node ID: 1704:2612
- Node Type: KSampler
- Exception Type: RuntimeError
- Exception Message: shape '[1, 112, 42, 16, 2, 2]' is invalid for input of size 602112
Same problem here without the "FluxKontextMultiReferenceLatentMethod" the workflow runs but the colors are all shifted towards red and the image is also less detailed.
With the standard comfy workflow with the non nunchaku version works without giving any error.
use the workflow of model card, and it will works, other guys face the same problem as they use the old workflow
node like: Edit Model Reference Method, cause the problem
node like: Edit Model Reference Method, cause the problem
Yes, the error is raised when the Edit Model Reference Method node is being used, but that node is absolutely necessary in order to preserve quality. I truly appreciate your work, but without this node the output is blurry and overexposed and, sadly, this makes your SVDQuant virtually unusable. :( The question is, whether the problem lies with the SDVQuant and is something that you can address on your end? Or should this issue be raised with the Nunchaku or ComfyUI teams?
Thank you for your work.
Yes, we need to use the node Edit Model Reference Method for better image quality. It would be great if someone fixes the error so we can use Nunchaku with this node.
The node's real name is FluxKontextMultiReferenceLatentMethod and the value is index_timestep_zero
Update: There already is a PR in the ComfyUI-nunchaku repository that addresses this issue. It hasn't been merged into the main branch yet, but I've tested it and it worked flawlessly on my system:
use the workflow of model card, and it will works, other guys face the same problem as they use the old workflow
Can confirm using workflow of model card generates red tinted blurry images every time. Adding "FluxKontextMultiReferenceLatentMethod" causes "Exception Message: shape '[1, 112, 42, 16, 2, 2]' is invalid for input of size 602112". No good inferences with the model.