---
title: InstantRetouch
sdk: gradio
sdk_version: "4.44.1"
python_version: "3.10"
app_file: app.py
pinned: false
license: other
---

# InstantRetouch / IP2P-BiLA Demo

Public Hugging Face ZeroGPU demo for instruction-guided image retouching with the validation-selected IP2P/BiLA checkpoint.

- Model: IP2P/BiLA
- UI: image upload, optional instruction, seed, max side, strength, and selectable examples

This Space is isolated from the research repository. It does not import `agent/`, training scripts, or local experiment paths at runtime. Weights live in a separate Hugging Face model repo and are downloaded lazily through `BILA_WEIGHTS_REPO`.

## Required Space Variables

Set one of these in the Space environment:

- `BILA_WEIGHTS_REPO`: Hugging Face model repo containing the IP2P weight layout below.
- `BILA_MODEL_ROOT`: local path with the same layout, useful only for staging/debugging.

Optional:

- `HF_TOKEN`: required if `BILA_WEIGHTS_REPO` is private.
- `BILA_MODEL_CACHE`: cache location. If unset, the app uses `/data/bila-space-demo/hf-cache` when persistent storage exists, otherwise `/tmp/bila-space-demo/hf-cache`.

## Weight Repo Layout

Do not commit weights into this Space repo. Put them in a separate HF model repo:

```text
ip2p/
  base/
    tokenizer/
    text_encoder/
    vae/
    unet/
  checkpoints/
    <bila-checkpoint>.pth
metrics/
  <metric-summary>.json
```

The app follows the validation-style direct flow: load the IP2P base model, load the BiLA checkpoint named in `model_manifest.json`, generate `bila_output`, then apply strength as `input + strength * (bila_output - input)`.