Modelpublic

skillsafeai/sdxs-512-dreamshaper-onnx-webgpu

sourceHugging Faceopenrail++updated 4d agoView on Hugging Face

0likes

Model Card

SDXS-512-DreamShaper, fp16 ONNX for the browser

A one-step 512x512 text-to-image model for onnxruntime-web over WebGPU, exported from IDKiro/sdxs-512-dreamshaper at commit 76f720262bb051da75666b22c902a78c8e16c763 (distilled from Lykon's DreamShaper 8; image decoder TAESD by madebyollin, MIT). No weights were retrained or modified - only exported.

File	Size	Graph
`text_encoder.onnx`	246 MB	CLIP ViT-L/14 text model, fp16, `input_ids` int64 [1,77] -> `last_hidden_state` float32 [1,77,768]
`unet.onnx`	632 MB	UNet, fp16 inside, float32 IO: `sample` [1,4,64,64], `timestep` [1], `encoder_hidden_states` [1,77,768] -> `out_sample`
`vae_decoder.onnx`	2.5 MB	TAESD decoder, fp16 inside: `latent` [1,4,64,64] -> `image` [1,3,512,512] in [-1,1]

Standard ai.onnx operators only, opset 17, no external data, single files.

Sampling (one step, no guidance)

sigma = 14.614646911621094   # Euler, trailing spacing, sigma_max
x   = randn(1,4,64,64) * sigma
eps = unet(x / sqrt(sigma^2 + 1), t = 999, text_encoder(clip_tokens(prompt)))
x0  = x - sigma * eps
img = vae_decoder(x0) / 2 + 0.5

Checked against the reference diffusers StableDiffusionPipeline (fp32, one step, guidance 0) on three prompts with identical noise: PSNR 41.8 / 50.4 / 47.9 dB. In Chrome (WebGPU, Apple silicon): ~0.5-0.7 s per image after a ~5 s session build.

Licence and use restrictions

Same CreativeML Open RAIL++-M terms as the source model (base DreamShaper 8: CreativeML Open RAIL-M). Their use-based restrictions apply to this export and to its outputs - including no content that exploits or harms minors, no non-consensual imagery of real people, no defamation, harassment or illegal use. Licence texts are linked from the source model cards.