PixArt-900M-DiT v0.1

This is an experimental 900M DiT model that has been fully fine-tuned on a large-scale, high-quality dataset of images aligned with human preferences. It uses the same VAE and text encoder as PixArt-Sigma, so you only need to replace the transformer.

How to use it?

weight_dtype = torch.bfloat16
device = torch.device("cuda")
MODEL_ID = "PixArt-alpha/pixart_sigma_sdxlvae_T5_diffusers"
transformer = Transformer2DModel.from_pretrained("Owen777/pixart-900m", torch_dtype=weight_dtype) 
pipe = PixArtSigmaPipeline.from_pretrained(MODEL_ID, transformer=transformer, torch_dtype=weight_dtype) 
pipe.vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=weight_dtype).cuda()
pipe.to(device)
image = pipe("a cute cat with pure white background.",guidance_scale=3.5,num_inference_steps=30,height=1024,width=102)[0]
image.save("result.jpeg")

Credits

Original PixArt Sigma model by PIXART-Σ (https://pixart-alpha.github.io/PixArt-sigma-project/)

This model is created by Tian Ye and Peijie Dong from ROAS/DSA, HKUST-GZ.