AnimateDiff
Plug-and-play motion module that turns Stable Diffusion models into animators.
About
AnimateDiff (ICLR 2024 Spotlight) adds motion to frozen Stable Diffusion checkpoints by injecting a temporal-attention motion module into the UNet — no retraining of the base image model required. That design made it the standard way to animate the thousands of community SD 1.5 and SDXL checkpoints, spawning a large ecosystem of motion LoRAs and ComfyUI workflows.
The official repo (guoyww/AnimateDiff, Apache 2.0) ships v3 motion modules plus an SDXL beta branch, and the method underpins countless short-form animation pipelines. Expect 2–4 second clips with the classic diffusion-animation look; longer or highly coherent motion still needs newer video foundation models.
What it can do
- Motion module injection for SD 1.5 and SDXL
- No base-model retraining required
- Motion LoRA ecosystem for style control
- Hugging Face Spaces demo
Setup
Clone https://github.com/guoyww/AnimateDiff, install requirements, download the v3 motion module safetensors from Hugging Face, run with any SD 1.5 checkpoint.
Prerequisites
Python, CUDA GPU with 12GB+ VRAM, a Stable Diffusion 1.5 or SDXL checkpoint.
Use cases
Works with
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 63; Trending score 74/100 from researcher assessment.
★ 10,000 GitHub stars
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.