search

comfyanonymous / ComfyUI

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface.
https://www.comfy.org/
GNU General Public License v3.0
57.7k stars 6.11k forks source link

ConditioningSetArea not compatible with sd 3.5, but compatible well with flux.1 dev #5493

Open binarytahr opened 2 weeks ago

binarytahr commented 2 weeks ago

Your question

i test ConditioningSetArea and ConditioningCombine node with sd 3.5 large and flux.1 dev. flux.1 dev's results are what it supposed to be: dog in the center, robot in the right corner, without quality loss. sd 3.5 large's results are disordered. so i am thinking if there are some compatibility issue?

flux.1 dev: f2c7a522-5f85-4a1f-9fe9-a2b3c12b26f7

sd 3.5 large: c5ca66d9-8f3e-41ce-bf50-6950592a1119 f28a841d-30a9-45ac-b380-609ce964e579

Logs

No response

Other

No response

ltdrdata commented 2 weeks ago

Rather than being a compatibility issue, it's a characteristic of the model. Things like conditioning combine actually interfere with and intervene in the model's learned image generation process. FLUX handles this more flexibly and effectively, while SD3.5 doesn't handle it well.

Experimentation is needed, such as narrowing down the area or adjusting the number of steps.

binarytahr commented 2 weeks ago

thanks for your answer! i am sorry i originally thought it was because the code wasn't adapted for sd 3.5 embedding's operation, now I know it's because the model itself. i have made some simple test including narrow down the aera before asking the question and it doesn't seem to work, i will experiment with other parameters next.