# ICBINP ONNX Laria Bundle

ONNX image-generation bundle prepared for Laria Android/Python parity testing.

## Layout

- `tokenizer/`
- `scheduler/scheduler_config.json`
- `text_encoder/model.onnx`
- `unet/model.onnx`
- `unet/model.onnx.data`
- `vae_encoder/model.onnx`
- `vae_encoder/config.json`
- `vae_decoder/model.onnx`

## Notes

- This model behaves like a classic Stable Diffusion pipeline and requires
  classifier-free guidance for coherent output.
- Python validation used guidance scale `5.0`.
- Img2img requires the included `vae_encoder`.
- The `vae_encoder/config.json` only carries the VAE scaling factor needed by
  the local pipeline.

## Local Test Scripts

Reference scripts live in the companion ONNX bench:

- `onnx_txt2img_ICBINP.py`
- `onnx_img2img_ICBINP.py`