Reshoot-Anything is a self-supervised video reshooting model built on top of Wan2.2-I2V-A14B. Given a source video and a target camera trajectory (encoded as an anchor video), it generates a high-fidelity reshoot that faithfully follows the new camera path while preserving original content, complex dynamics, and temporal consistency — trained entirely on in-the-wild monocular videos.
Reshoot-Anything: A Self-Supervised Model for In-the-Wild Video Reshooting
Avinash Paliwal, Adithya Iyer, Shivin Yadav, Muhammad Ali Afridi, Midhun Harikumar
Morphic Inc. · arXiv:2604.21776
Source Video
Reshot Video
Source video
Reshot video
Model Files
This repository contains two LoRA checkpoints (rank-512, applied to attention and feed-forward layers of Wan2.2-I2V-A14B):
File
Role
Notes
jan06_scaling_80k_ckpt1400.safetensors
High-noise expert
Controls early denoising steps. Primarily responsible for camera motion alignment and global scene structure. Trained on ~80k clips with scaling augmentations + 15% synthetic data mixture.
Controls late denoising steps. Responsible for texture fidelity and fine detail. Uses standard black-background anchors, no source reconstruction loss.
At inference, generate an anchor video by converting your source video to a 4D point cloud, applying the target camera trajectory, and forward-warping to produce the geometric anchor. See the repo's anchor_generation/ scripts for details.
Note: Refer to the GitHub README for the authoritative argument names and single-GPU usage.
How It Works
Reshoot-Anything adapts the Wan2.2-14B Mixture-of-Experts (MoE) DiT with two key architectural changes:
Dual-stream token conditioning — Both the anchor video V_a (geometric guide) and source video V_s (texture reference) are VAE-encoded and temporally concatenated as tokens into the model's main self-attention mechanism. This outperforms cross-attention for view synchronization by letting the model directly route textures across spatial and temporal positions.
Offset RoPE — A fixed temporal offset of 50 is added to source video token positional embeddings, strictly decoupling source context from the active denoising trajectory.
The model learns implicit 4D spatiotemporal routing — when a target frame requires content occluded in the corresponding source frame, the model locates and re-projects the missing texture from a different timestep in the source video.
Self-Supervised Training Pipeline
Training requires no paired multi-view data. From a single monocular video:
Two independent smooth random-walk crop trajectories are sampled → source V_s and target V_t
V_s[0] is forward-warped via AllTracker dense flow + crop offset → anchor V_a
The triplet (V_s, V_a, V_t) forms the training signal
A hybrid dataset strategy augments the monocular pipeline with a 15% mixture of paired synthetic data from ReCamMaster, enabling generalization to extreme (120°+) orbital camera trajectories.
Results on 100 five-second videos from OpenSora-Mixkit (16fps, 480p):
Method
CLIP-F ↑
RotErr ↓
TransErr ↓
Mat. Pix ↑
FVD-V ↓
CLIP-V ↑
ReCamMaster
98.49
11.29
19.59
1314.00
732.52
88.91
EX-4D
98.94
3.94
4.21
2188.98
685.63
89.77
TrajectoryCrafter (49f)
98.80
2.26
3.03
1851.80
582.56
92.40
Ours
99.03
2.76
4.23
2720.83
586.24
93.16
Ours (49f)
99.01
2.61
2.73
2737.65
488.22
94.96
Citation
bibtex
1@article{paliwal2026reshootanything,
2 title={Reshoot-Anything: A Self-Supervised Model for In-the-Wild Video Reshooting},
3 author={Paliwal, Avinash and Iyer, Adithya and Yadav, Shivin and Afridi, Muhammad Ali and Harikumar, Midhun},
4 journal={arXiv preprint arXiv:2604.21776},
5 year={2026}
6}
License
Model weights are released under the Apache 2.0 license, consistent with the Wan2.2 base model.
