ActionSplice

Same-step state editing for interruptible world models.

Paper: ActionSplice: In-Flight Action Editing for Interactive World Models

Project page: pardistaghavi.github.io/actionsplice-website

ActionSplice updates an active world-model rollout when the control input changes before the current video chunk has finished sampling. Waiting delays the response, directly switching the conditioning leaves the intermediate solver state shaped by the previous action, and restarting repeats completed computation. A learned Counterfactual State Transport (CST) corrector moves the interrupted solver state toward the matched counterfactual state at the same solver step, then lets the frozen backbone and sampler finish the normal trajectory without replaying completed evaluations.

Contents

• Method
• Installation
• Training
• Inference and model loading
• Repository structure
• Citation

Method

ActionSplice provides two separately trained correctors:

Variant	Edited region	Intended use
CST-R	Complete active chunk	Retarget a chunk after a control interruption
CST-T	Temporal suffix at boundary m	Preserve the temporal prefix at the intervention step and correct only the suffix

Backend state conventions

Backend	Native state	Corrector input	Corrector target	Same-step reconstruction
minWM Wan Action2V	[B,T,16,H,W]	[B,T,16,H,W]	Clean prediction	Stored transition noise
HY-WM1.5	[B,32,T,H,W]	[B,T,32,H,W]	Direct Euler state	None; deterministic resume

Installation

Core and development tools

git clone https://github.com/PardisTaghavi/ActionSplice.git
cd ActionSplice

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e '.[dev]'
pytest

Optional dependencies are grouped by workflow:

python -m pip install -e '.[train]'
python -m pip install -e '.[inference]'

Upstream backbones

ActionSplice integrates with upstream repositories rather than redistributing their code or weights. Prepare pinned source checkouts with:

bash scripts/prepare_backends.sh

Note

minWM and HY-WorldPlay have independent dependency stacks. Use a separate Python environment for each backend and install that upstream repository's requirements there. ActionSplice's inference extra supplies only the shared client-side dependencies.

Pinned revisions and licensing notes are recorded in THIRD_PARTY.md.

Training

Choose one backend-specific configuration:

Backend	CST-R	CST-T
minWM	configs/minwm/train_cst_r.json	configs/minwm/train_cst_t.json
HY-WM1.5	configs/hyworld15/train_cst_r.json	configs/hyworld15/train_cst_t.json

Replace the <PATH_TO_...> values in a copied configuration, then train one corrector:

actionsplice-train \
  --capture-dir /path/to/captures \
  --output-dir outputs/minwm-cst-r \
  --config configs/minwm/train_cst_r.json

Resume an interrupted run with --resume /path/to/checkpoint.pt.

The training code preserves the research objectives already used by the project: normalized state reconstruction, the minWM clean-prediction delta, suffix-masked CST-T reconstruction, the CST-T intra-chunk boundary term, and optional decoded LPIPS/temporal/boundary terms. Repository cleanup did not introduce new losses.

All loss coefficients are configurable through the loss_weights block. Its defaults reproduce the current HY-WM1.5 objective; the minWM configurations override the residual coefficient. See the training guide for the complete lambda schema.

Capture commands, manifest formats, split checks, and backend-specific examples are documented in docs/training.md.

Inference and model loading

The repository contains no backbone or corrector checkpoints. Use a local checkpoint path with the commands below.

Local corrector checkpoints can be loaded independently of the backbone:

from pathlib import Path

import torch

from cst.backends import get_backend
from cst.core.runtime import load_transport_model

corrector, metadata = load_transport_model(
    Path("checkpoints/hyworld15-cst-r/best.pt"),
    device=torch.device("cuda"),
    dtype=torch.bfloat16,
)
get_backend("hyworld15").validate_checkpoint_config(
    metadata["model_config"], method="cst_r"
)

The loader validates checkpoint structure and role; the backend validation checks the target type, latent-channel count, solver length, and selected method. Load the upstream backbone separately under its original license.

Backend-specific commands run both variants end to end from a small rollout configuration:

actionsplice-infer-minwm \
  --method cst_r \
  --task-config configs/inference/minwm_cst_r.example.json \
  --minwm-root /path/to/minWM \
  --transport-checkpoint /path/to/minwm-cst-r.pt \
  --output outputs/minwm-cst-r.mp4

actionsplice-infer-hyworld15 \
  --method cst_t \
  --task-config configs/inference/hyworld15_cst_t.example.json \
  --reference-image /path/to/reference.png \
  --hyworld-root /path/to/HY-WorldPlay \
  --model-path /path/to/HunyuanVideo-1.5 \
  --action-checkpoint /path/to/action/checkpoint.safetensors \
  --transport-checkpoint /path/to/hyworld15-cst-t.pt \
  --output outputs/hyworld15-cst-t.mp4

• Inference sequence and current hooks
• Checkpoint validation

Repository structure

ActionSplice/
├── src/cst/
│   ├── backends/       minWM/HY adapters and shared backend registry
│   ├── core/           corrector model, state layouts, same-step runtime
│   ├── data/           capture schemas, datasets, partitions, manifests
│   ├── training/       shared corrector training loop and losses
│   └── cli/            capture, manifest, training, and inference commands
├── configs/
│   ├── minwm/          minWM CST-R/CST-T capture and training configs
│   ├── hyworld15/      HY CST-R/CST-T capture and training configs
│   └── inference/      runnable single-rollout CST-R/CST-T examples
├── docs/               training, inference, and model loading guides
├── scripts/            pinned upstream setup
└── tests/               CPU unit tests

The distribution is named actionsplice; the import package remains cst because CST is the framework's core operation. cst.backends is the shared backend boundary and keeps upstream imports lazy.

Checkpoint naming compatibility

Research checkpoints retain legacy role strings for strict validation:

Public method	minWM role	HY-WM1.5 role
CST-R	action_h0	action_h0_state
CST-T	action_hm	action_hm_state

Public commands, files, and documentation use CST-R/CST-T.

Citation

If you use ActionSplice, please cite:

@article{taghavi2026actionsplice,
  title={ActionSplice: In-Flight Action Editing for Interactive World Models},
  author={Taghavi, Pardis and Guo, Tingyu and Lossner, Jonas and Pandey, Gaurav and Langari, Reza},
  journal={arXiv preprint arXiv:2609.08230},
  year={2026}
}

License

ActionSplice is released under the Apache License 2.0.

Contributions are welcome; see CONTRIBUTING.md for the development and pull-request workflow.