AnimateDiffWelcome to the central repository for understanding AnimateDiff's core components: the motion module. Unlike traditional AI video models that are monolithic, AnimateDiff pioneered a modular approach. It separates the appearance of a video (handled by a standard Stable Diffusion checkpoint model) from the motion (handled by a small, specialized motion module). This is the key to its flexibility.
A motion module is a trained neural network that has learned “motion priors” from a vast dataset of video clips. It understands generic concepts of movement—how things flow, bounce, walk, and pan. As explained by Guo et al. in the foundational paper (arXiv:2307.04725), this module is injected into the attention layers of a frozen, pre-trained text-to-image model like Stable Diffusion. The result is that the static image generator suddenly learns to produce coherent sequences of frames. Choosing the right motion module is critical, as it directly impacts animation quality, style, and compatibility with your base model.
O que é um módulo de movimento?
Think of a motion module as a “motion instruction set” for Stable Diffusion. A base Stable Diffusion 1.5 model knows how to draw a picture of a “running wolf,” but it has no concept of how to show the sequence of movements from one frame to the next. The motion module provides that temporal knowledge. It doesn't contain any visual style itself; it only contains the logic of motion. This plug-and-play design means you can apply the same motion module to countless Stable Diffusion checkpoint models, whether it's a model for photorealism like Realistic Vision or an anime model like ToonYou.
This is why most AnimateDiff motion modules are designed for Stable Diffusion 1.5. The classic modules were trained to be inserted into the architecture of SD 1.5 models and are generally not compatible with SD 2.0 or SDXL without specific, newer modules.
Catálogo oficial de módulos de movimento
Here is a breakdown of the official motion modules released by the AnimateDiff team. For most users, starting with mm_sd_v15_v2 is the recommended choice for a balance of quality and compatibility. When you need the absolute best motion quality for SD 1.5, the AnimateDiff v3 model is the top choice.
| Arquivo do módulo de movimento | Versão base do SD | Resolução | Notas |
|---|---|---|---|
| mm_sd_v14.ckpt | Stable Diffusion 1.5 | 256x256 (trained) | The first version. Good for historical context, but superseded by later models. Can produce a distinct, slightly more “jittery” style. Useful for some artistic effects. Download size is smaller. |
| mm_sd_v15.ckpt | Stable Diffusion 1.5 | 256x256 (trained) | An improvement on v1.4 with more stable and coherent motion. Was the standard for a long time. A solid choice, but v1.5_v2 is generally preferred now. |
| mm_sd_v15_v2.ckpt | Stable Diffusion 1.5 | 384x384 (trained) | (Recommended starting point). An improved version of v1.5 with significantly better motion dynamics, trained at a higher resolution. Often called the “improved motion” model. Offers a great balance of performance and quality for general-purpose use. This is the workhorse AnimateDiff 1.5 percent motion module. |
| v3_sd15_mm.ckpt | Stable Diffusion 1.5 | 512x512 (trained) | The version 3 motion module, offering the highest quality motion for SD 1.5. The motion is smoother and more detailed. It's designed to be used with an associated v3_sd15_adapter.ckpt Domain Adapter LoRA for best results, especially for “sparse-control” scenarios. This model represents the cutting edge for SD 1.5 animations. |
| mm_sdxl_v10_beta.ckpt | Stable Diffusion XL | 512x512 (trained) | The experimental beta motion module for the SDXL architecture. It allows AnimateDiff to work with SDXL base models, enabling higher-resolution outputs (e.g., 1024x1024) and leveraging SDXL's improved prompt understanding. As a beta model, results can be more varied, but it's the key to unlocking AnimateDiff for the SDXL ecosystem. |
Compatibilidade com modelos personalizados e checkpoints
The magic of AnimateDiff is its broad compatibility. Because the motion module only influences movement, you can use it with almost any custom Stable Diffusion 1.5 checkpoint model. This includes popular community models on Civitai such as:
- ToonYou: Perfect for creating anime and cartoon-style animations.
- Realistic Vision: Ideal for generating lifelike, photorealistic video clips.
- Counterfeit, MeinaMix, majicMIX: Other popular models for various artistic styles that can all be animated.
The workflow remains the same: you load your chosen checkpoint model for the visual style, and then you load a compatible AnimateDiff motion module to provide the animation. This simple combination of a style model and a motion model unlocks limitless possibilities.
Onde baixar os modelos do AnimateDiff
To ensure you're getting the official, safe files, it's best to download the motion modules from the original sources. There are two primary locations:
- Hugging Face: The official repository from the researchers is on Hugging Face under the guoyww/animatediff profile. This is the most direct and reliable source for the core motion module files. You can find the original models here.
- Civitai: The Civitai community hub also hosts copies of the motion modules, often with user reviews and example outputs. This can be a great place to discover not only the modules but also compatible base models and see what others are creating.
When you download a motion module, make sure you place it in the correct directory for your software, whether it’s the models folder within the AnimateDiff extension for AUTOMATIC1111 or the designated custom node folder in ComfyUI.
O módulo de movimento AnimateDiff v3: um grande salto
The release of the AnimateDiff v3 model marked a significant evolution. Trained at a higher resolution and on a more diverse dataset, the v3_sd15_mm.ckpt motion module provides noticeably smoother and more naturalistic motion compared to its predecessors. It's particularly effective at reducing the “boiling” or “flickering” artifacts that can sometimes appear in AI-generated video.
A key innovation with the v3 release is the introduction of the “Domain Adapter” LoRA (v3_sd15_adapter.ckpt). This LoRA is designed to work alongside the v3 motion module to improve its adaptability to different visual domains and to enable a new feature called “sparse-control.” This allows for more nuanced control over the animation, making the AnimateDiff v3 model a powerful tool for serious artists.