sm079/nanosaur2-web

Nanosaur2 web builds

Smaller files of Nanosaur2-670M for running it in the browser on WebGPU. Nanosaur2 is by its author and MIT-licensed; these are unofficial, modified versions of its files.

file	contents
nanosaur2_dmad_4step_diffusion_model.int8.safetensors	the 4-step diffusion model; attention and MLP linears in int8
nanosaur2_diffusion_model.int8.safetensors	the normal diffusion model; attention and MLP linears in int8
nanosaur2_vae_decoder.safetensors	the VAE's decoder half (bf16, unchanged weights; the DINOv2 encoder is left out)

The quantized linears use ComfyUI's int8_tensorwise format with ConvRot (a 256-point Hadamard rotation per group of inputs, one scale per output channel); everything else is bf16 as released. The text encoder isn't included: use the original nanosaur2_text_encoder.safetensors.

Image fidelity against the bf16 diffusion models (3 prompts, fixed seeds, 768×768): 28.7 dB PSNR for the 4-step model, 29.6 dB for the normal model, with practically identical images.