README.md

metadata

title: ASASR Super-Resolution
emoji: 🎨
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.10.0
app_file: app.py
short_description: Faithful x4 image super-resolution with FLUX.1-dev dual-LoRA
startup_duration_timeout: 1h
license: cc-by-nc-4.0

✦ ASASR · Coloring the Noise

Adversarial Sobolev Alignment for Faithful Image Super-Resolution ICML 2026

Hongbo Wang · Huaibo Huang · Pin Wang · Jinhua Hao · Chao Zhou · Ran He

---

Overview

This Space is an interactive demo for Coloring the Noise: Adversarial Sobolev Alignment for Faithful Image Super-Resolution (ICML 2026).

ASASR performs x4 image super-resolution with a FLUX.1-dev backbone and dual-LoRA inference:

• sr_lora/pytorch_lora_weights_v2.safetensors — base SR LoRA.
• dpo_lora/adapter_model.safetensors — ASASR AS-DPO alignment LoRA.

Both LoRAs are loaded from wafer-bob/ASASR. The base model is black-forest-labs/FLUX.1-dev, which is gated and requires an HF_TOKEN Space secret with model access.

How to use

1. Upload a low-resolution square image (≈128 × 128), or click one of the bundled examples.
2. Click ✦ Super-Resolve to run the 28-step ASASR sampler on ZeroGPU.
3. Drag the ImageSlider to compare the low-resolution input against the 512 × 512 ASASR reconstruction.

The interface is centered on an interactive before/after comparison slider, with a clean hero header, six standard Gradio examples, and a collapsible citation block. A modern light theme with an indigo/violet accent is used throughout.

Demo UI

• Hero header with the ICML 2026 badge, paper title, author list, and quick-link buttons to the paper, code, and model.
• Centerpiece gradio_imageslider comparison view (Input LR | ASASR HR).
• Standard Gradio examples — six diverse 128 × 128 crops wired to the super-resolution function with lazy example caching.
• Footer with a CC-BY-NC-4.0 license notice and a collapsible BibTeX citation block.

Setup notes

• The Space runs on ZeroGPU; the inference function is decorated with @spaces.GPU with a tunable duration (ASASR_GPU_DURATION, default 45 s, clamped to 30–300 s).
• Asset resolution is deferred to the first request unless a valid HF_TOKEN secret is present, in which case a startup prefetch warms the cache.
• Model artifacts are cached under /tmp/.cache/huggingface to stay within the ephemeral disk budget.

License

ASASR code and weights are licensed under CC-BY-NC-4.0 for non-commercial research use only. The demo also depends on the non-commercial terms of FLUX.1-dev.

Citation

BibTeX

@inproceedings{wang2026asasr,
  title     = {Coloring the Noise: Adversarial Sobolev Alignment for Faithful Image Super-Resolution},
  author    = {Wang, Hongbo and Huang, Huaibo and Wang, Pin and Hao, Jinhua and Zhou, Chao and He, Ran},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}

References

• ASASR model weights: https://huggingface.co/wafer-bob/ASASR
• ASASR source code: https://github.com/wafer-bob/ASASR
• Paper (arXiv): https://arxiv.org/abs/2605.23264
• FLUX.1-dev base model: https://huggingface.co/black-forest-labs/FLUX.1-dev