UNet-only ONNX export of SG161222/RealVisXL_V5.0_Lightning for in-browser WebGPU inference (onnxruntime-web) in the Generate AI Images extension.

Only the UNet is hosted here: RealVisXL's text encoders and VAE are within fp16 noise of stock SDXL, so the engine shares the copies from d0gr/sdxl-lightning-onnx-webgpu.

I/O: sample float32 [B,4,H,W] · timestep int64 [steps] · encoder_hidden_states float32 [B,77,2048] · text_embeds float32 [B,1280] · time_ids float32 [B,6] → out_sample float32 [B,4,H,W].

The checkpoint is Lightning-distilled upstream — run it at 4–6 steps, CFG 1–2, Euler with trailing spacing. Exported with tools/export-sdxl-unet.py (LIGHTNING_STEPS=0), retargeted with tools/retarget-unet-profile.py; one UNet pass agrees with the torch fp16 checkpoint at cosine 1.000000.

CreativeML Open RAIL++-M, inherited from the base model — including its use restrictions (Attachment A). All credit for the weights goes to SG161222; this repo only re-packages them for browser inference.