VincentNi/wan22-rollout-blocks-ranking-rgb-step0

Wan2.2 TI2V Step-0 Rollout — RoboTwin blocks_ranking_rgb

160 video rollouts (10 scenes × 16 samples) generated by Wan2.2-TI2V-5B + merged Vidar LoRA on the RoboTwin blocks_ranking_rgb task (arrange R/G/B blocks left-to-right). These are the pre-NFT-training (step-0) baseline samples used to evaluate reward-model behaviour and seed RL fine-tuning.

Generation config

• —Base model : Wan2.2-TI2V-5B
• —LoRA : vidar/merged_vidar_lora.pt (vidar baseline merged into DiT)
• —Sampler : deterministic ODE (Euler flow-matching), eta=0
• —Steps : 50
• —Shift : 5.0
• —CFG scale : 5.0
• —Frames : 121 @ 640×736
• —Seeds : 42–57 per scene (one per rollout index 0–15)

The sampler is byte-aligned with fastvideo/train_nft_wan_2_2_ti2v.py::ode_rollout_batch (same flow_ode_step, same batched-CFG forward, same per-step first-frame mask re-application), so each video reproduces the corresponding step-0 video the NFT trainer would emit for the same (scene, rollout_index).

Layout

robotwin_blocks_ranking_rgb_<scene_id>/
  robotwin_blocks_ranking_rgb_<scene_id>_g000_s42.mp4
  ...
  robotwin_blocks_ranking_rgb_<scene_id>_g015_s57.mp4

• —10 scenes × 16 rollouts = 160 mp4 files (~415 MB total)

Reproducing

Run scripts/inference/rollout_first_round_blocks_ranking_rgb.sh from the EmbodiedVideoRL repo with the default config above.