Not good at all, why?
Hi,
I tested this in spaces with real test LoRAs, it's not good at all and didn't generate the images that looked like the images i fed it (tested and tried three times with: 3, 9 and 25 high quality images that i previously used to generate LoRAs with). Is it the space's fault? ( @AiSudo /Qwen-Image-to-LoRA)
Tried these 3 different things, the LoRA was not good at realism LoRAs:
- Object
- Character
- Clothing
But the idea is cool.
Hello, I'm the creator of the space you mentioned :)
You're right, I think the lora created by Qwen-Image-i2L is best suited for creating style/theme.
For example, if you train one on a comic book, the LoRA might not pick up a specific character from the comic, but it'll learn the drawing style instead.
In my opinion, for things like clothing, face, pose, and object: PuLID, Controlnet or the Inpainting method would be a better options.
I have Inpainting and Controlnet spaces, feel free to check them out. Maybe they're more suitable.
I was a bit wrong in the title, it does work better from ComfyUI when generating images with the LoRA it gave. But not as good as the AI-Toolkit LoRA training method.
I understand what you meant, sadly we can't edit the title after it's posted lol
Be patient, this model was just released few days ago, this is a brand new idea, more will come in the future.
We need to support DiffSynth-Studio so they can release faster and better models 😀
All the best Team