Sign In

Motion Boost v1 — Wan 2.2 I2V A14B (high noise)

0

Not recently loaded

Download

1 variant available

fp16 SafeTensor

motion_boost_v1_wan22_i2v_high.safetensors

Half precision, best balance • 585.13 MB

Verified:

Type
LoRA
Stats

13

0

Reviews
Published

Nov 21, 2026

Base Model

Wan Video 2.2 I2V-A14B

Usage Tips
Strength: 0.8
Hash
AutoV29B8F16B9A0

A motion LoRA for the high-noise expert of Wan 2.2 I2V A14B. It was trained only on motion-capture renders (CMU Graphics Lab database, rendered on neutral adult 3D characters) and CC BY real-world clips, so there are no artist styles or identities in it: it only pushes the "how the body moves" part of the model.

  • Load it on the high-noise model only (after your 4-step / lightx2v LoRA if you use one). Weight 0.6–1.0, 0.8 recommended.
  • What it helps: floor moves and handstands in breakdance keep the body intact, turns complete, more happens in 5 seconds. Martial arts and everyday motion change little.
  • Known limits: trained at 640x368 / 49 frames, so the effect is softer at 720p. Later epochs of this run learned the mocap camera and started to move the camera mid-shot, which is why only epoch 2 is released.
  • Training: kohya-ss/musubi-tuner, rank 32, timesteps 900–1000, 2 epochs. Apache-2.0 base.

Credits: CMU Graphics Lab Motion Capture Database (free for all uses, please credit), VideoUFO (CC BY).