RECOMMENDED LORA SCALE: 0.5-0.7
Trigger
Add details to this face, improve skin details
Description
FAQ
Comments (17)
Does it preserve the identity of the subject?
yeah, I recommend 0.4-0.5 scale for identity perservation
KeyError: 'lora_unet_final_layer_adaLN_modulation_1.lora_down.weight'
In official workflow
same error
Compatibility issue with ComfyUI-Nunchaku is documented: https://github.com/mit-han-lab/ComfyUI-nunchaku/issues/340
Created a HuggingFace repo to host the patch script and the patched version: https://huggingface.co/lym00/comfyui_nunchaku_lora_patch
you know you really should provide a before/after comparison with a lora like this. i cant say if its working correctly but if it is, its not good at all.
It is working fine. I recommend using it with starting 0.45 lora scale and gradually increase it as you like (I recommend max 0.8). It might not work outside of fal because it needs a conversion script with key/values inside lora. It should work on Comfy normally.
https://fal.ai/models/fal-ai/image-editing/realism/playground?share=e4e4742d-49df-4488-884f-0efa13562dc3
https://fal.ai/models/fal-ai/image-editing/realism/playground?share=c21d6743-a8e6-41e6-b031-627c42fd6304
@gokaygokay you keep showing images without a hint of metadata but there is still no before-after. which, given how kontext works, you should have.
@m4rcus the links I sent have before and after images inside it
@gokaygokay thanks, these links actually work. i also can see it works very well with head shots. problem is it doesnt restrict itself to skin. it goes over the whole picture. fissling out details on everything that aint skin.
@m4rcus I agree it is not perfect yet, I am working on making a way better version of this one. This was just an experiment
in comfyui with fp8 flux kontext and exact prmpt as on description it works horribly
Magnificent work youve done here. Looking forward to the future versions of this model.
Which workflow should I use for this
This is actually so fire 🔥🔥🔥🔥 recommend using it on a low strength together with the character lora (eg. rgthree power lora loader node). Thanks for this!
why not include before/after in your examples?