Qwen 2.1 FinalCut (6.6GB) 25 Step
After 20 failed trainings including a training that took 4 A100 80GB a few days I finally have a version I am happy to release.
25 Steps
Reduced "QWEN Canvas" issue.
Works with most systems without issues and has lora support.
Note: I settled on group size 32 and standard qwen exclusions for INT8
I found the single DiT blocks of this model way worse to train then the double/single Krea2 - I personally do not like this model to train, and for use I would rate it above SDXL and Flux2. It handles text placement and guidance very well, but seems to fight new concepts.
Description
FAQ
Comments (7)
Is it good for image edit?
It should still take the multi-image and transparency input, but as that is not part of comfy core I did not test
Looks very good, but I'm gonna wait on users commenting on its editing capabilities before plunging on yet another model.
I did a couple of test. It works for image editing. Unfortunately for T2I I had more misformed limbs that with the base mode. what is the recommended scheduler and sampler?
I can't wait for V2
does it preserve faces
Normal/Euler
Thank you for the model.
It is near 2x slower than Int8 models (3060ti).
It does edit just fine.
It generates fine.
It can use loras i tested just fine (consistency, turbo, BFS head swap)
For edit, Turbo loras shift the image alot.
Overall, not for me since it takes twice as much to generate the same quality as base or other models.