README.md

metadata

title: Loopy
emoji: 🔁
colorFrom: gray
colorTo: red
sdk: gradio
sdk_version: 6.26.0
app_file: app.py
pinned: false
python_version: '3.12'
startup_duration_timeout: 1h
short_description: Seamless looping video generation with Wan2.2 + Loopy
models:
  - htdong/Loopy
  - Wan-AI/Wan2.2-T2V-A14B
  - lightx2v/Wan2.2-Distill-Loras

🔁 Loopy — Seamless Video Loop Generation

Demo of Loopy — Seamless Video Loop Generation via Anchored Looping Shift of Positional Embedding (code · project page).

Loopy turns Wan2.2-T2V-A14B into a looping video generator. Its trick is the anchored looping shift of positional embedding: inside self-attention block i the temporal RoPE frequencies are cyclically rolled by shift = 0 if i == 0 else (i - 1) % (F - 1) + 1 (with F latent frames), which puts the temporal positions on a circle so the last frame becomes a neighbour of the first one and the clip loops without a visible seam. Two small LoRAs (high-noise / low-noise expert) adapt the base model to that shifted geometry.

What this Space runs

Faithful port of the reference implementation (generate_2.2_new.py, Wan2.2/wan2/text2video_roll.py, Wan2.2/wan2/modules/model_roll.py) onto 🧨 diffusers' WanPipeline so it fits on ZeroGPU:

Base	linoyts/Wan2.2-T2V-A14B-Diffusers-BF16 (bf16 mirror of Wan-AI/Wan2.2-T2V-A14B-Diffusers)
LoRAs merged	LightX2V 4-step distillation (rank 64, high + low) + Loopy (rank 4, high + low)
RoPE	anchored looping shift, applied to self-attention only, per the paper
Sampler	UniPC, flow sigmas, shift = 12.0, 4 steps, guidance = 1.0 (CFG-free)
Default output	832×480, 53 frames @ 16 fps (the repo's test.sh setting)
Quantization	fp8 dynamic (both 14B experts), int8 weight-only (UMT5-XXL text encoder)

Prompts in the examples are the ones the authors showcase on the project page.

Prompting tips

Describe motion that is naturally periodic or continuous — falling snow, drifting clouds, flowing water, a walking animal, a rotating object — and state the style and shot type. English and Chinese both work.