Anima-ControlNet-VACE-Depth
A depth ControlNet for Anima, trained with the VACE spaced-block-duplication + zero-conv architecture.
⚠️ This release contains adapter weights only. It is not a standalone model — you must load it together with the Anima base model.
Training
Trained with sd-scripts by kohya-ss
| Setting | Value |
|---|---|
| Base Model | Anima-Base-V1.0 |
| Optimizer | AdamW (fp32 states), weight decay 0.01 |
| Learning rate | 5e-5, cosine, 500-step warmup |
| Effective batch | 16 |
| Steps | 20,000 |
| Resolution | 1024² |
| Precision | bf16 (full_bf16) |
| Flow shift | 5.0 (matches inference) |
| Loss | L2 on flow-matching velocity |
| Weighting | uniform |
Tips
- Depth source: Use DepthAnything V2 for best performance, although this model is trained on a mix of different depth sources, and thus any popular depth preprocessor should in theory work.
Usage
Supported for the moment in ComfyUI through my fork of ComfyUI-Advanced-ControlNet
Credits
- Anima - Circlestone Labs
- sd-scripts — kohya-ss.