---
title: SPEED
emoji: 🌈
colorFrom: indigo
colorTo: blue
sdk: gradio
sdk_version: 5.34.0
app_file: demo_hf/app.py
pinned: false
short_description: Training-free progressive sampling for FLUX & WAN 2.1
hardware: zero-a10g-large
hf_oauth: false
python_version: "3.11"
---

# Training-free demo of Spectral Progressive Diffusion for Efficient Image and Video Generation

- Project page: https://howardxiao.ca
- Paper: https://arxiv.org/abs/2605.18736

## Inputs

Same arguments as the CLI release:

- `prompt`
- `transform` — `dct` (default) / `dwt` / `fft`
- `scales` — comma-separated, e.g. `0.5,1.0`. Strictly increasing, ends at `1.0`. DWT requires every consecutive ratio = 2.
- `delta` — noise-dominated tolerance (paper Eq. 9), default `0.01`.
- `seed`
- Modality-specific: `n_steps`, `height`, `width`, `guidance` (+ `num_frames`, `shift` for WAN 2.1).