sd15-controlnet-canny-int8

sd15-controlnet-canny-int8

Stable Diffusion 1.5 with ControlNet (canny), int8 ONNX, for Latent Studio.

Why this exists

A stock SD 1.5 UNet export takes three inputs (sample, timestep, encoder_hidden_states). ControlNet residuals have nowhere to go, so ControlNet cannot run against it — this is true of the public ONNX exports too, including ones published by projects that support ControlNet.

The UNet here was traced through a wrapper whose forward() takes the 13 residual tensors as real arguments, so the graph exposes:

down_block_additional_residuals.0 ... .11
mid_block_additional_residual

16 inputs in total. Verified present after int8 quantisation.

Layout

unet/model.onnx + unet/model.onnx.data                 # 16 inputs
controlnet/model.onnx + controlnet/model.onnx.data     # 4 in, 13 out
controlnet/type.txt                                    # "canny"
text_encoder/, vae_decoder/, vae_encoder/, tokenizer/

Text encoder and VAE are exported from SD 1.5 base, matching the UNet, rather than borrowed from a fine-tune.

Conditioning

controlnet_cond is a NCHW float tensor in [0,1] (not the VAE's [-1,1]). Residual scaling is applied by the caller.