TinySanaPreview

TinySanaPreview is a custom ComfyUI node that implements real-time previews during generation for Sana diffusion models.

Setup

1. Clone this respository into your custom_nodes directory
   git clone https://github.com/cake-ml/tiny-sana-preview.git ComfyUI/custom_nodes/tiny-sana-preview
2. Download the tsd.safetensors decoder model from Hugging Face and place it in your models/vae_approx directory
3. Restart ComfyUI and insert the latent > TinySanaPreview node to your workflow anywhere before the sampling node, then select the tsd.safetensors decoder model and appropriate dtype (bf16 is recommended, but older GPUs may require fp16 or fp32)

Decoder model

The TinySanaDecoder model can be found on Hugging Face at cake-ml/tsd. TSD decodes with a compression factor of 8, rather than the standard 32 of the Sana DC-AE, and utilises far fewer parameters (9.6M vs 159M). This results in preview images with a quarter of the width/height that the DC-AE decoder produces, but with a roughly 46x speedup (8.42ms vs 389ms for a 1, 32, 32, 32 latent on an Nvidia RTX A4000) and a significantly lower memory footprint. Latents of dimensions B, 32, H, W will decode to images of dimensions B, 3, H*8, W*8. The TSD model was trained on an Nvidia RTX A4000 for approximately 60 hours.