I've trained primitive i2i lora in an hour but already works nice
#6
by sorryhyun - opened
Hi, it feels a bit promoting by myself, but I wanted to share about this. I realized that qwen3.5 shows nice vision capability and there's reference dimension in moss model, I tried training a simple i2i task, colorizing manga image.
The approach was quite simple. I fed qwen(image, caption), vae(downscale(image, 2)) as source and image as target, and trained about an hour with local 5070ti, and the result is as below:
I uploaded lora weight and workflow in here and comfyui repo in here and training code is in here Anyway thanks for this excellent model.
Manga coloring is a fun first test.