Basic camera motion LoRAs for Animatediff v2 motion models - version 1
Training
This LoRA was trained on RunPod cloud GPUs.
Model Description
This repository contains a collection of basic camera motion LoRAs finetuned on Animatediff v2 models. The training process utilized Kijai's AnimateDiff Motion Director ComfyUI nodes. These LoRAs have been tested exclusively in ComfyUI with Kosinkadink's AnimateDiff-Evolved nodes, however they may be compatible with other AnimateDiff implementations.
File summary
| Name | Compatible Motion Model | Base AR (w/h) | Frames | input videos |
|---|---|---|---|---|
| 1600_cseti_9192119_camera-zoomout-10vid_mv2.safetensors | v2 | (3:2) | 16 | 10 |
| 1600_cseti_1077723_camera-zoomin-10vid_mv2.safetensors | v2 | (3:2) | 16 | 10 |
| 2800_cseti_1093371_camera-zoomin-32f-10vid_mv2.safetensors | v2 | (3:2) | 32 | 10 |
| 1600_cseti_6664682_camera-lateral_right-10vid_mv2.safetensors | v2 | (3:2) | 16 | 10 |
| 1600_cseti_9213319_camera-lateral_left-10vid_mv2.safetensors | v2 | (3:2) | 16 | 10 |
| 1600_cseti_9406077_camera-crane_down-10vid_mv2.safetensors | v2 | (3:2) | 16 | 10 |
| 1600_cseti_1720230_camera-crane_up-10vid_mv2.safetensors | v2 | (3:2) | 16 | 10 |
How to use
- copy the motion lora files into your "ComfyUI\models\animatediff_motion_lora" folder
- Use the "Load Animatediff LoRA" node
General recommendation
These are finetuned on a v2 model. They work also with AnimateLCM but don't work with v3 models
They work with SD 1.5 checkpoints
Animatediff v3 adapter LoRA is recommended regardless they are v2 models
If you want more motion try increasing the scale multival
I recommend using the above Base AR (Aspect ratio) for inference however they might work with other ARs too
Try playing with the strength of
- motion LoRA
- scale multival
- v3_adapter lora
like increasing the scale multival and lowering the lora strength.
I recommend using IP-Adapter, however it reduces the effect of the mLORAs, so play with the strength. "Ease in out" weight type is a good starting point.
If you chain multiple motion LoRAs, I recommend lowering their strength, otherwise, the quality of the result will be worse
Details and examples
Zooming in (16 frame)
- recommended strength: 0.7-1.0
Zooming in (32 frame)
- this is an experimental motion LoRA trained on 32 frames so it can be used with context length 32.
- recommended strength: 0.7-0.85
Zooming out
- recommended strength: 0.7-1.0
Lateral (horizontal) left
- recommended strength: 0.7-1.0
Lateral (horizontal) right
- recommended strength: 0.7-1.0
Crane (vertical) up
- recommended strength: 0.7-1.0
Crane (vertical) down
- recommended strength: 0.7-1.0
Support
Producing and sharing this kind of open-source work requires renting cloud GPUs, which gets expensive quickly. If you find it useful and would like me to keep contributing, your support is very much appreciated: