Implementing support for PIXART-α Fine Tuning and DreamBooth

[FurkanGozukara]

PIXART-α: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis

https://pixart-alpha.github.io/

They have train scripts here : https://github.com/PixArt-alpha/PixArt-alpha/tree/master/train_scripts

This model is literally better than SDXL

I made a full tutorial for those who wonders.

PIXART-α : First Open Source Rival to Midjourney - Better Than Stable Diffusion SDXL - Full Tutorial

[mrT333]

This model is literally better than SDXL

And what is most amazing about it: It matches SDXL quality while having only 12% training time of SD v1.5 (remember the low quality of base SD v1.5?). This could indicate that once the community starts finetuning Pixart it will result in much higher quality improvements using way less training steps than SD 1.5 or SDXL.

Also we might not even know where the quality ceiling of Pixart is as it could be severely undertrained with just 12% training steps... meaning finetuning would have even greater impact on quality than it did on SD 1.5 and it could reach a quality plateau way later than all stable diffusion based models.

[FurkanGozukara]

This model is literally better than SDXL

And what is most amazing about it: It matches SDXL quality while having only 12% training time of SD v1.5 (remember the low quality of base SD v1.5?). This could indicate that once the community starts finetuning Pixart it will result in much higher quality improvements using way less training steps than SD 1.5 or SDXL.

Also we might not even know where the quality ceiling of Pixart is as it could be severely undertrained with just 12% training steps... meaning finetuning would have even greater impact on quality than it did on SD 1.5 and it could reach a quality plateau way later than all stable diffusion based models.

this is 100% accurate. huge potential there is.

[storuky]

This model is literally better than SDXL

It matches SDXL quality while having only 12% training time of SD v1.5

Lol, this model is totally undertrained. Just type something like "a giraffe" and generate few results with different seeds. Same with tiger, lion, etc. It's not variative.

Once again I'm convinced that if someone talks about reducing CO2 emissions, it's empty talk.

[FurkanGozukara]

CO2 emissions

who cares about CO2 emissions :)

we need better model