AI images that form as you type.

Real-time diffusion in your browser. Nothing uploads. No login, no account, no credits. Powered by fused WebGPU dispatches — every speedup on this page is measured live on your GPU, and the PyTorch head-to-head ships in the repo (bench-torch.py).

Probe: element-wise chain (mul·add → relu → sub → tanh → mul → sigmoid) over 1M floats. Fused = 1 dispatch, chain lives in registers. Naive = 6 dispatches, intermediates in global memory. Both paths are diffed against each other at boot (max abs Δ reported below); the timing delta is pure dispatch + memory-traffic overhead.

Waiting for probe…

FFN block: (matmul → GELU → matmul → residual) at mid-block shape [2·1024·1280]·[1280·5120]·[5120·1280]. Naive = 4 dispatches with intermediates in global memory. Fused = 2 dispatches with GELU + residual baked into matmul epilogues. Same tiled matmul in both paths — the delta is fusion, not kernel quality.

Waiting for FFN block benchmark…

Self-attention at SD mid-block shape [B=2·H=20·S=256·head=64]. Naive = 3 dispatches (scores → softmax → attn-matmul) with scores + probs matrices materialized in global memory. Fused = 1 dispatch, scores stay in workgroup memory, softmax reduces cooperatively, output emits directly.

Waiting for attention benchmark…

One full SD-Turbo transformer block [B=2·S=256·D=1280·heads=20×64·D_FFN=5120]. Naive = 14 dispatches (LN → Q/K/V → scores → softmax → attn → out-proj → residual → LN → FFN-up → GELU → FFN-down → residual). Fused = 9 dispatches (flash-attn collapses 3 into 1, GELU folds into FFN-up matmul, both residuals fold into their producer matmul epilogues). Runs ~16× per U-Net forward pass.

Waiting for full-block benchmark…

GroupNorm at SD-Turbo mid-block feature-map shape [B=1·C=1280·G=32·16×16]. Naive = 2 dispatches (stats → normalize, with mean/rsigma materialized in scratch). Fused = 1 dispatch, one workgroup per (b,g), 256-thread parallel tree-reduce for mean and variance, stats kept in workgroup scalars. Runs ~30× per U-Net forward pass.

Waiting for GroupNorm benchmark…

Direct 3×3 Conv2d at SD-Turbo mid-block shape [B=1·C=1280·16×16·k=3]. Naive = 3 dispatches (conv → silu → residual-add), with Y re-read + re-written twice. Fused = 1 dispatch, SiLU + residual baked into the conv's per-output-value epilogue. Conv is called ~34× per U-Net forward pass.

Waiting for Conv2d benchmark…

Full SD-Turbo mid-block ResNet [B=1·C=1280·G=32·16×16·k=3]. Naive = 9 dispatches (GN-stats → GN-norm → SiLU → Conv1 → GN-stats → GN-norm → SiLU → Conv2 → skip-add). Fused = 4 dispatches (GN+SiLU collapsed via shared-mem stats, skip-add folded into Conv2's epilogue). First milestone where v2.5.0 and v2.5.1 chain end-to-end. Runs ~20× per U-Net forward pass.

Waiting for ResNet benchmark…

Cross-attn at SD-Turbo mid-block shape [B=2·H=20·S_q=256·S_kv=77·head=64]. Queries from the image feature map, K/V from text cond. Naive = 3 dispatches with scores [S_q × S_kv] in global memory. Fused = 1 dispatch, softmax kept on-chip in workgroup memory. Called ~32× per U-Net forward pass.

Waiting for cross-attention benchmark…

Timestep embed: sinusoidal [320] → Linear [1280] → SiLU → Linear [1280]. Fires once per denoise step. Naive = 3 dispatches. Fused = 1 dispatch with both intermediates (emb 320, hidden 1280) living in workgroup memory.

Waiting for timestep-embed benchmark…

End-to-end test of the hand-rolled ONNX protobuf parser against the real 1.73 GB SD-Turbo U-Net. Opt-in because it's a 1.73 GB download each run (this diagnostic path is deliberately not cached). Proves onnx-parser.js reads every TensorProto from the schmuell bundle without any protobuf library.

Click to fetch + parse the full U-Net and print the tensor-name tree. This is the audit step before v2.5.6 wires wgslUnetForward against real weights.