sona-forge/sd15-inpainting-controlnet-canny-fp16

Sona Forge SD 1.5 Inpainting ControlNet Canny FP16 ONNX

This repository hosts the Sona Forge Phase 8 conversion of the SD 1.5 inpainting UNet to ONNX FP16 with additional ControlNet residual inputs.

Intended use

This model is for local outfit restyling beta features in Sona Forge. It accepts an SD 1.5 inpainting sample tensor and the residual tensors produced by the separate Phase 7 ControlNet Canny ONNX.

It is not an exact virtual try-on model and should not be presented as fit, sizing, or purchase guidance.

Inputs

Input	Shape	Dtype	Notes
sample	[batch, 9, 64, 64]	FP16	Inpaint latent: noisy latent + masked image latent + mask channel.
timestep	[batch]	FP16	Scheduler timestep.
encoder_hidden_states	[batch, 77, 768]	FP16	SD 1.5 text embeddings.
down_residual_0..11	12 tensors	FP16	ControlNet down-block residuals.
mid_residual	[batch, 1280, 8, 8]	FP16	ControlNet mid-block residual.

Output

Output	Shape	Dtype	Notes
noise_pred	[batch, 4, 64, 64]	FP16	Predicted latent noise.

Phase 8 spike artifact

File	Size	SHA-256
model.onnx	1,720,220,149 bytes	1f69cf75e7cc5de94774d4419397f86dcec400cdec9e5fd688b6231a1aa01d51

Licence

This conversion inherits the CreativeML OpenRAIL-M use restrictions from the SD 1.5 inpainting base model. The ControlNet Canny source is also tracked in Sona Forge licence documentation.