ActionRoPE checkpoints

Checkpoints of ActionRoPE: Action as Coordinates for Spatially Consistent Control in Planar-Camera Game World Models (arXiv:2610.25861). Code: github.com/TeaWhiteBro/ActionRoPE · project page: teawhite.cn/ActionRoPE.

ActionRoPE writes the action-driven camera displacement of planar-camera (top-down / isometric) games into the rotary positions of a video DiT, so that tokens at the same world location share the same positional address over time. Coordinate-Conditioned Noise Scheduling (CNS) lets the regions that the first frame already shows settle earlier than the newly revealed ones. Each checkpoint is the full method (ActionRoPE + CNS + the known/new input channel) fine-tuned for 5,000 steps on gameplay of a top-down action game; these are the models reported in the paper. The model generates 81 frames (5 s at 16 fps, 832×480) from one first frame and a camera path.

Folder	Backbone	File	Format	Fine-tuning
wan2.2-ti2v-5b/	Wan2.2-TI2V-5B	actionrope_cns.safetensors (10.0 GB)	full DiT, bf16	full, lr 1e-5
skyreels-v2-df-1.3b/	SkyReels-V2-DF-1.3B-540P	actionrope_cns.safetensors (2.9 GB)	full DiT, bf16	full, lr 1e-5
ltx-2.5-22b/	LTX-2.5 (22B)	actionrope_cns.safetensors (38 GB)	full DiT, bf16 (LoRA merged)	LoRA r128, lr 1e-4
minimax-h3-33b/	MiniMax-H3 (33B, FL2VA)	actionrope_cns_lora.safetensors (1.2 GB)	LoRA + input channel, bf16	LoRA r128, lr 1e-4

The Wan2.2 and SkyReels-V2 folders also contain umt5_null_prompt.pt, the UMT5 embedding of the scene-free prompt "In a 2.5D top down view." that all evaluation clips use. With it, sampling with that prompt does not need to load the 11 GB text encoder.

Every file carries its settings in the safetensors metadata (model, format, mask_channel, cns, cns_shift, cns_shift_known, cns_alpha, step, and lora_* for LoRA files). The sampler reads them from there.

Usage

huggingface-cli download teawhite/ActionRoPE --include "wan2.2-ti2v-5b/*" --local-dir checkpoints
python scripts/sample.py --model wan --ckpt checkpoints/wan2.2-ti2v-5b/actionrope_cns.safetensors \
    --model_dir models/Wan2.2-TI2V-5B --first_frame frame.png --actions "right:40,left:40" --out clip.mp4

Each checkpoint also needs the files of its base model (VAE, text encoder, and the DiT for LoRA files). See the code repository for setup and training.

Limitations

The models are trained on one game with a fixed top-down camera that follows the player. Control is defined as planar camera translation. The models do not take other actions (attacks, interactions) or 3D camera motion.

License

Each checkpoint is a fine-tuned derivative of its base model and is distributed under that model's license; see LICENSE.md.

Citation

@misc{deng2026actionropeactioncoordinatesspatially,
      title={ActionRoPE: action as coordinates for spatially consistent control in planar-camera game world models},
      author={Zhiyang Deng and Yuzhe Wu and Zijun Lin and Yeying Jin},
      year={2026},
      eprint={2610.25861},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.25861},
}