khanghy1000/Anima-ControlNet-VACE-Canny

Anima-ControlNet-VACE-Canny

A Canny ControlNet for Anima, trained with Anima VACE ControlNet implementation by TaihoC.

⚠️ Requires the fix/anima-vace-hardening branch of PineCookie/ComfyUI-Advanced-ControlNet. (TaihoC's fork currently has a duplicate control hook bug.)

---

Training Details

Dataset

• Dataset: ~15,000 Danbooru images from RicemanT/booru-essence-2026.
• Canny Edge Maps: Generated using cv2.Canny with min_threshold=100 and max_threshold=200 (no random thresholds).
• Caption Dropout: 10%.

Hyperparameters

• Control Blocks: 4 control blocks, connected to the base model at blocks 0, 7, 14, and 21.
• Batch Size: 16 (GPU: 2, Gradient Accumulation: 8)
• Training Steps: 3000
• Learning Rate: 5e-5, cosine, 150 warmup steps
• Optimizer: AdamW_adv
  • betas: (0.9, 0.99)
  • use_atan2: True
  • orthogonal_gradient: iterative
  • state_precision: bf16_sr
  • stochastic_rounding: True
• Training Precision: full bf16
• Timestep Sampling: shift, discrete_flow_shift=3.0, sigmoid_scale=1.0
• Resolution: 1024x1024

Peak VRAM Usage: ~32GB (no gradient checkpointing)

Compute Cost: ~10 hours on 1x L40S

---

Credits

• Anima VACE ControlNet - TaihoC
• booru-essence-2026 - RicemanT
• sd-scripts - kohya-ss
• Anima - Circlestone Labs