Fast and simple. wan2.2 gguf img2vid with interpolation and upscale
Description
added duplicate clip to lower noise model for greater accuracy
FAQ
Comments (16)
WARNING - with wan2.2 i2v workflows, many get the error "groups=1...". It is caused by having the node flow2-wan-video installed. Use the manager in ComfyUI to delete it. The internet is full of faulty advice on how to fix this issue- but the node deletion is the correct solution. And no, I am not talking about workflows using this faulty node. This faulty node breaks your workflow by simply existing in your ComfyUI node library. ComfyUI allows non-active library code to break active code, and that is just so wrong from a CS POV, it is literally insane.
Thank you for this! I was trying to understand why I was getting this error all the time!!!
you are the best
Unfortunately, this didnt work for me - still getting "Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[1, 32, 18, 90, 60] to have 36 channels, but got 32 channels instead"
rickets_xxx I had the same problem. I was updating ComfyUI using the manager, but you need to use update in update folder, this one: update_comfyui.bat . I hope this helps. :)
I'm getting this error too and don't have flow2-wan-video installed. I have also updated comfyui via the update_comfyui.bat file. I'm not sure what else could be causing the issue, any ideas?
0atz At first I got that error as well without flow2-wan-video installed. "Update all Custom Nodes" via ComfyUI Manager didn't seem to help for me (it often does for errors like that).
What helped for me was:
Change the VAE to "wan_2.1_vae_fp8_e4m3fn.safetensors"
pottkopp yeah, I gave up on this one.
pottkopp I actually solved my issue, sorry for the long delay! It turns out after you uninstall flow2-wan-video, there is still a flow2 folder in the custom nodes folder that comfyui never removed. simply deleting it and restarting comfyui got everything working
Now I've finally got wan2.2 i2v working (see other message), let me tell everyone it is mindblowing. As this (and other) workflow shows, the best speedup LoRA works flawlessly. 4-6 total steps and the video is crazy good. BUT I've discovered the samplers and schedulers have an insane effect on the quality of Wan2.2 output, and I'd encourage everyone to investigate for themselves! (res_2s and bong_tangent are slow but transformative to Wan2.2 t2v, and must be used to be believed.)
I switch between euler, ddim, and uni_pc for samplers and scheduler is simple or beta, the latter especially with ddim
Any luck adding other Wan 2.1 Loras to these workflows? Like the "bounce" LORA etc?
all works even better you just have to duplicate lora settings for both low and high. this workflow is missing that. maybe in this one works but for example i get even better results than original wan 2.1 with the same loras. annoying thing to dupklicate all settings. except lightlora for that se something like 3.0 on high and 1.5 on low. experiment
fadog74502638 Are you using GGUF, FP8, or another version? Which LORA loader are you using? I keep getting load errors for the loras I load up due to mismatch.
fadog74502638 Thanks for being a stand-out help!