Workflow for quickly replacing the clothes in the reference image based on the kontext(nunchaku) model, with a high success rate.
How to use : https://civarchive.com/articles/16559
Description
FAQ
Comments (17)
Missing Node Types
When loading the graph, the following node types were not found
NunchakuFluxDiTLoader
NunchakuFluxLoraLoader
Read the decision here: https://github.com/mit-han-lab/ComfyUI-nunchaku/issues/159
Install nunchaku
Once PyTorch is installed, you can directly install nunchaku from Hugging Face, ModelScope or GitHub release. Be sure to select the appropriate wheel for your Python and PyTorch version. For example, for Python 3.11 and PyTorch 2.6: https://github.com/mit-han-lab/nunchaku?tab=readme-ov-file#installation
Great workflow and great tips.
Any clue about how to avoid big heads?
Fixed a big head in a similar setup:
Use a NAGKsampler and put this in the negative NAG clip text encode: "big head"
Tell me how it went!
@lnknou Unfortunately NAG isn't supported in nunchaku yet. The other brute force trick you can use is make the head in proportion to the body image, but you'll lose some face detail.
Thanks guys. I'll try that.
from my experience , use a thin ratio will gen a slim body
beats out flux.safetensors是什么模型？
uh, a nsfw lora 🤣
@0xqia where could i get this lora
@lakerwe exackly ?
também não achei essa lora
Got it working, but after one complete generation, I can't generate again without restarting the whole workflow or shutting down ComfyUI. Is there any sort of reset I can add to the process to ensure it lets me do two in a row?
try change seed after generation
you need to purge the Vram manualy or with a node
There are several settings that can block subsequent image generation -
1 - "SEED" "control after generate" settings set to "fixed"
2 - "Preview Bridge" module settings with "if_mask_empty" checked and no mask at all.
3 - and an error will be displayed - "'NoneType' object is not subscriptable" - "Image And Mask Preview" module with the "pass through" option unchecked
Don't collapse a node containing a prompt, in this case the LayerMask one.