deroberon / demofusion-comfyui

An experimental implementation of a node that implements Demofusion technique in Comfyui
83 stars 3 forks source link

img2img node, please? #9

Open c4fun opened 4 months ago

c4fun commented 4 months ago

As you know demofusion can be used as a way to detailize current low-res images. So the image to image is quite a great implementation. The source of img2img from demofusion is like this: https://github.com/PRIS-CV/DemoFusion/blob/main/gradio_demo_img2img.py

Could you make a node out of it? Thanks.

ttulttul commented 4 months ago

My suggestion is to head over to the overpainting workflow for the Iterative Mixing Sampler: https://github.com/ttulttul/ComfyUI-Iterative-Mixer/blob/main/examples/Overpainting.json

If you want to add detail, I would further suggest taking your input image and overpainting it from step 0 to 50%, then skip two steps, and sample from there to end_step. For instance go from 0 to 20, then 22 to 40. Skipping a step causes the model to get out of step with the noise and generate all kinds of details that were never there to begin with.

kklors commented 2 weeks ago

That would be indeed great to have the img2img function. Also could be higher resolutions be supported than 4096x?