Modelpublic

brad-agi/sana-0.6b-onnx-webgpu

sourceHugging Faceapache-2.0updated 6mo agoView on Hugging Face

3likes

Model Card

Sana 0.6B — ONNX for In-Browser WebGPU Inference

Generate 1024x1024 to 4096x4096 images entirely in the browser using WebGPU.

Component	File	Size	Precision
CLIP text encoder	onnx-community/clip-vit-large-patch14-ONNX	432 MB	uint8
DiT 1024	1024/sanadit1024.onnx + .data	2.3 GB	float32
DiT 2048	2048/sanadit2048.onnx + .data	2.3 GB	float32
DiT 4096	4096/sanadit4096.onnx + .data	2.3 GB	float32
VAE 1024	1024/sanavae1024.onnx + .data	608 MB	float32
VAE 2048	2048/sanavae2048.onnx + .data	608 MB	float32
VAE 4096	4096/sanavae4096.onnx + .data	608 MB	float32

Note: DiT must be float32 — Sana's linear attention produces NaN in fp16.