Downloads · 30 days
0
zhenyuxie-zhzh/wan22-egohoi-rgb-lora-stage1
wan22-egohoi-rgb-lora-stage1 is a machine learning model from zhenyuxie-zhzh. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Stage-1 export of the EgoHOI RGB LoRA fine-tune: the original Wan2.2-TI2V-5B DiT stays frozen; rank-128 / alpha-128 LoRA adapters on the attention q/k/v/o and FFN linears of all 30 blocks (300 linears, 600 tensors), t…
Downloads · 30 days
0
Access
Public
Updated Sep 19, 2026
Repo size
1.1 GB
Likes
0
Public
Click a slice to open those files.
.safetensors1.1 GB · 100%
From the Hugging Face model README
Stage-1 export of the EgoHOI RGB LoRA fine-tune: the original Wan2.2-TI2V-5B
DiT stays frozen; rank-128 / alpha-128 LoRA adapters on the attention
q/k/v/o and FFN linears of all 30 blocks (300 linears, 600 tensors), the
per-block modulation vectors, the time-embedding MLPs and the head modulation
(37 tensors) and the EgoHOI conditioner (hand-skeleton video + camera
trajectory, 120 tensors) were trained for 2000 optimizer updates on the 1442
HOT3D EgoHOI training clips (81 frames, 480x480, 64 samples per update,
AdamW lr 1e-4, weight decay 0.01, 500-update warmup, EMA 0.999; this file is
the live weights at step 2000).
It is the starting point of RGB LoRA stage 2 on the V4-formal architecture
(docs/RGB_LORA_STAGE2_V4FORMAL_PORTING.md in the code repository) and can
also be evaluated on its own with examples/inference_egohoi_val_rgb_lora_stage1.py
(the loader fuses the LoRA into the frozen DiT).
| File | What | sha256 |
|---|---|---|
final-step-00002000.safetensors | live stage-1 weights (LoRA A/B, modulation/time tensors, conditioner), bf16, 1.09 GB | 93f8e0267074dbc57b93ee55bc69be700692d34515d4de4412a5eb105bdaa1fc |
rgb_lora_stage1_contract.json | immutable training contract of the run (architecture, LoRA scheme, dataset fingerprint, optimizer) | |
training_args.initial.json | full argument list of the stage-1 run (paths are those of the source cluster) | |
SHA256SUMS | checksums |
Tensor naming: pipe.dit.blocks.N.self_attn.q.lora_A.default.weight (and
lora_B), pipe.dit.blocks.N.modulation, pipe.dit.time_embedding.*,
pipe.dit.time_projection.*, pipe.dit.head.modulation,
pipe.egohoi_conditioner.*. Base weights are not included: download
Wan-AI/Wan2.2-TI2V-5B separately.
hf download zhenyuxie-zhzh/wan22-egohoi-rgb-lora-stage1 --local-dir /models/rgb-lora-stage1
sha256sum -c /models/rgb-lora-stage1/SHA256SUMS
export RGB_LORA_STAGE2_STAGE1_CHECKPOINT=/models/rgb-lora-stage1/final-step-00002000.safetensors
Trained on HOT3D-derived data (Meta HOT3D license applies to the data; the adapter weights are released under the base model's Apache-2.0 terms for research use).