hunyuanvideo-1.5-robotwin15-i2v-lora

HunyuanVideo 1.5 RobotWin-15 I2V LoRA checkpoints

This public RoMALab repository contains HunyuanVideo 1.5 four-step I2V LoRA checkpoints post-trained on 15 RobotWin tasks. The comparison runs use stride-16 temporal coverage, rank-16 video attention Q/K/V/O adapters, video FFN adapters, mixed spatial gating, a future-latent MeanFlow objective, and global batch 16 over four GPUs.

August 2026 conditioned and wrist-view runs

These are the latest checkpoints from the four requested RobotWin-15 post-training experiments. All use global batch 16, stride-16 temporal windows, minimum future arm-joint path 0.05, and a configured 15,000-step budget; the published step is the final checkpoint reached by each allocation.

Each folder contains config.yaml, resolved_config.json, metadata.json, conditioning.pt, LoRA safetensors, and the resumable training state. metadata.json records byte sizes and SHA-256 hashes. For the three explicitly conditioned variants, conditioning.pt is required for inference; the LoRA file alone does not contain the task/arm or wrist-fusion parameters.

Completed July 2026 runs

The gate probability applies only to samples with reviewed spatial-gate metadata; other samples remain ungated. The jointmotion005 run restricts examples using a minimum future arm-joint path of 0.05 and was trained from initialization without a resume boundary. It stopped normally at its 22.98-hour allocation limit.

Each completed-run folder contains:

config.yaml
resolved_config.json
metadata.json
lora/pytorch_lora_weights.safetensors
training_state.pt

The LoRA file is sufficient for inference. training_state.pt contains optimizer and learning-rate-scheduler state for exact-topology continuation; PyTorch checkpoint files should only be loaded from trusted sources. SHA-256 hashes for both binary files are recorded in each metadata.json.

Loading

from diffusers import HunyuanVideo15Transformer3DModel

transformer = HunyuanVideo15Transformer3DModel.from_pretrained(
    "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v_step_distilled",
    subfolder="transformer",
)
transformer.load_lora_adapter(
    "RoMALab/hunyuanvideo-1.5-robotwin15-i2v-lora",
    subfolder="checkpoints/ffn-r32-gate75-step-015000/lora",
    adapter_name="robotwin",
)

Earlier checkpoints

The pre-continuation comparison checkpoints remain under:

• checkpoints/ffn-r32-gate75-step-009279/
• checkpoints/ffn-r16-gate75-step-009000/
• checkpoints/ffn-r16-gate20-step-010419/

checkpoint-006000/ is also retained for reproducibility. It is the earlier attention-only rank-16 checkpoint trained with four temporal windows per episode, rather than the stride-16 FFN comparison recipe.

Completed non-distilled MimicVideo transfer

The batch-64, rank-128, 61-frame checkpoint at update 938 includes adapter and optimizer state. It uses the non-distilled hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v base; see its directory for the exact configuration, revision and checksums.