You can deploy this model and start earning money today!

sambhav-pantheon/
vidihand

Name: vidihand
Availability: InStock

Views

No views yet

ViDiHand: Wan2.1-VACE backbone and three Stage-2 decoders

Private research checkpoint bundle for the Pantheon ViDiHand replication. It contains one shared video-diffusion backbone and all three full-size post-Stage-2 decoder heads.

text

backbone/
  wan2.1-vace-1.3b/               Stock Wan-AI/Wan2.1-VACE-1.3B snapshot
  stage1b_vace_step10000.safetensors
decoders/
  model1/decoder_step30000.safetensors
  model2/decoder_step30000.safetensors
  model3/decoder_step30000.safetensors
code/                              Exact decoder/feature-contract source
MODEL_CHECKPOINTS.json             Paths, sizes, SHA-256, architectures
vidihand_config.json               End-to-end inference contract
licenses/WAN_LICENSE.txt           Upstream Apache-2.0 license

All three variants share the stock Wan2.1-VACE-1.3B DiT/VAE/UMT5 weights and the same Stage-1b VACE branch. Only the Stage-2 decoder and captured feature depths differ:

Variant	Feature capture	Decoder	Shape behavior
Model 1	DiT block 15 at diffusion call 34/50	36.9M, width 512	Per-frame MANO beta
Model 2	Blocks 8/15/22 at call 34/50	121.8M, width 768	Per-frame MANO beta
Model 3	Blocks 8/15/22 at call 34/50	121.8M, width 768	One clip-token-mean beta per hand

Model 3 additionally used articulation loss weight 0.5 during training versus 0.1 for Model 2. This changes the trained weights, not the diffusion backbone.

Exact runtime

Source: https://github.com/Pantheon-Industries-Inc/vidihand-pantheon
Source revision: 8231b3ee97017cf91dd9d32984897320d08b8da6
DiffSynth-Studio revision: 3f5fa5286fa6d35b3b12a739b4dde5500b191e5d
Input: 81 RGB frames at 30 fps, short side 480, dimensions divisible by 16
Diffusion: flow-matching schedule, 50 nominal steps, sigma shift 5, seed 0, CFG 1.0; execute only the prefix through one-based call 34
Decoder output: two hands over 81 frames with MANO orientation, articulation, beta, depth, visibility, and 2D joints

Use MODEL_CHECKPOINTS.json as the checksum authority. The Stage-1b file is the trained VACE branch, not a merged replacement for the stock Wan DiT.

MANO dependency

MANO model files are not included. Obtain MANO separately under its license and provide mano_torch_left.npz and mano_torch_right.npz at runtime. Do not redistribute MANO assets from this repository.

Licensing

The bundled upstream Wan2.1 files retain their Apache-2.0 license in licenses/WAN_LICENSE.txt. Licensing for the reconstructed ViDiHand code and trained Stage-1b/Stage-2 deltas has not yet been formally selected, so this repository remains private and uses license: other.

sambhav-pantheon/vidihand

ViDiHand: Wan2.1-VACE backbone and three Stage-2 decoders

Contents

Exact runtime

MANO dependency

Licensing

sambhav-pantheon/
vidihand