---
title: Foveated Diffusion Demo
emoji: 👁️
colorFrom: gray
colorTo: blue
sdk: gradio
sdk_version: 6.14.0
app_file: app.py
pinned: false
license: apache-2.0
suggested_hardware: l40sx1
hf_oauth: false
# Spaces' default startup timeout (30 min) isn't enough for FLUX.2 + LoRAs to
# download on first boot. Bump to 1 h so cold starts don't fail.
startup_duration_timeout: 5h
tags:
  - lora
  - diffusion
  - foveated-rendering
  - text-to-image
  - flux
github: https://github.com/bchao1/foveated_diffusion
---

# Foveated Diffusion — live demo

Live demo of [**Foveated Diffusion: Efficient Spatially Adaptive Image and
Video Generation**](https://bchao1.github.io/foveated-diffusion/). Place a
circular foveal region you want rendered at high resolution, pick a LoRA
adapter, and run a 50-step inference pass through the foveated FLUX.2 pipeline.

- Project page: https://bchao1.github.io/foveated-diffusion/
- Paper: https://arxiv.org/abs/2603.23491
- Code: https://github.com/bchao1/foveated_diffusion
- Model weights: https://huggingface.co/bchao1/foveated_diffusion

## How to use

1. Type a prompt and pick a seed.
2. Choose a LoRA from the dropdown:
   - `no_fov` — finetuned baseline, no foveation conditioning (always full-HR).
   - `random` — foveation conditioning at random gaze locations.
   - `saliency` — saliency-driven foveation masks.
3. Set the foveal circle: center `cx, cy ∈ [-0.5, 0.5]` (0 = image center) and
   radius `r` (relative to half the image diagonal). With `no_fov` the controls
   are disabled — the whole image is rendered HR.
4. The **Tokenization mask** preview shows exactly which 32×32-px blocks will
   be rendered at HR (white) vs LR (gray) before you commit.
5. Click **Generate**. The denoising progress bar reports steps `k / 50`.

## Fixed pipeline settings

| Setting | Value |
|---|---|
| Resolution | 1024 × 1024 |
| Steps | 50 |
| CFG | 4.0 |
| Decode mode | `merge` |
| Prediction type | `clean` |
| LR downsample factor | 2 |
| Soft foveation blend | on |

## Hardware

FLUX.2-klein-base-4B + LoRA in `bfloat16` needs ~30 GB of VRAM. We recommend
running this Space on an **L40S** (`l40sx1`) or larger GPU tier. The CPU-Basic
free tier cannot host the model.

## Citation

```bibtex
@misc{chao2026foveateddiffusion,
      title={Foveated Diffusion: Efficient Spatially Adaptive Image and Video Generation},
      author={Brian Chao and Lior Yariv and Howard Xiao and Gordon Wetzstein},
      year={2026},
      eprint={2603.23491},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.23491},
}
```