The CircleStone Model is licensed by CircleStone Labs LLC under the CircleStone Non-Commercial License. Copyright CircleStone Labs LLC. IN NO EVENT SHALL CIRCLESTONE LABS LLC BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

Built on NVIDIA Cosmos

An AnimateDiff-style motion module for Anima, with a ComfyUI node and a "freeze or animate anything" workflow.

Anima is an image model. This module adds temporal attention after each of its 28 blocks, so 16 frames can be generated together and stay coherent. Anima itself stays frozen, so every frame keeps full Anima image quality.

What it is good at today

Pick any part of an image with a word (WeVisOne matting: character, face, sky, sea…) and either:

freeze it and let everything else flow: a still character while the sea, moon and sky shift around them, AnimateDiff style;
animate only it: expressions and small gestures over a background that stays pixel-identical.

Keyframe modes

Generate the keyframe with a character sheet (Anima Ref Adapter).
Generate it from text only (T2V); the keyframe becomes the video's character reference.
Use your own image; it keeps its aspect ratio and becomes the character reference.

Output: 16 frames, 2 s at 8 fps. About 30–45 s per clip on an RTX 5060 Ti 16 GB with Anima Turbo.

What's in the files

anima_motion_r2v_v0.1.safetensors: the motion module (117.5M parameters). Put it in ComfyUI/models/anima_motion/.
anima_motion_v0.1_workflow_and_node.zip: the ComfyUI node (custom_nodes/ComfyUI-Anima-Motion), the workflow and INSTALL.md with every other download.

Requirements

Anima: Turbo or Base, the Qwen3 0.6B text encoder and qwen_image_vae.
Anima Ref Adapter nodes and the v2 LoRA.
ComfyUI-WeVisOne-VL for the selection mask (my port of Tencent WeVisOne).

Recommended settings

denoise 0.8; 0.4–0.6 for subtle breathing; 1.0 mutates.
Mask margin 4–8 px with a soft edge when freezing (no halo), 16 px for faces.
When freezing a character, the background pass runs without the charref and with no humans in the prompt (the workflow does this for you).

Experimental: please read

Trained on only 20 clips (MiniMax H3 image-to-video from Anima keyframes, 384×576, 8 fps), 2000 steps.
Motion is small: expressions and gestures. Backgrounds morph rather than move physically.
Without a mask it does not start from your image.

Next plans: about 300 clips (characters plus fixed-camera landscapes with ambient motion), keyframe conditioning in training, a bigger cloud run, and 32-frame 16:9 and 9:16 clips.

Full write-up with result tables, training scripts and the first I2V module: huggingface.co/paralaif/anima-motion-module

License: CircleStone Labs Non-Commercial License, as for Anima.