PixArt Sigma XL 2 MS: 2k, 1024, and 512 full finetune on custom captions.
INSTRUCTIONS: Place the .safetensors where the original model would go and select bunline.
Favorite sampling settings:
512/1024 models dpm++2s_a, simple, 24 steps, and CFG 3.1, 4.2, or sometimes more
2k model euler, sgm_uniform, 48 steps, CFG 3.5, 5, or sometimes more
Description
dpm++2s_a, 25-40 steps, and CFG 4-8
FAQ
Comments (14)
Tell me how to use this? Is this for A1111?
Coming soon to A1111 (edit: oops sd.next) but available now in ComfyUI with an extra plugin. Replace the Sigma model with the .safetensors to use
ComfyUI plugin https://github.com/city96/ComfyUI_ExtraModels?tab=readme-ov-file#pixart-sigma
@yayaman could we use it in StableSwarm with a custom workflow behind it? Or SDFX?
@thebrownsauce184 if StableSwarm is really comfy behind the scenes then idk why not, but haven't used it myself. I'll give it a shot sometime soon, though! The pixart discord is full of more knowledgeable people than myself as well. SDFX seems like a Gradio competitor and TIL thank you.
@thebrownsauce184 does work well in stable swarm i can confirm
@ReyArtAge Sweet I shall try it!!
Add this in place of PixartSigma-XL-2-1024-MS.pth?
Exactly! Keep the VAE and T5 the same, but add next to the original .pth. Select bunline from ComfyUI to use
this is actually pretty good, how did you train it? I see more potential in this than SD3. I would like to train it as well
Thanks! Using the official trainer and default config except real_prompt_ratio=1. That's to use only the one "prompt". Also extracting vae/t5 embeddings beforehand.
https://github.com/PixArt-alpha/PixArt-sigma/
@yayaman How did you do the extraction part first? What does the setup for training it look like in terms of GPU, VRAM and system RAM?
@anyMODE We spoke on discord, good luck w/ training!
@yayaman Is it neccessary to use official training code, or more user-friendly OneTrainer can produce the same quality with the same compute requirements?
@desm0nt OT works also! See the anime fine tune of Sigma (in suggested resources) for their configs