Vorch-Director

Interactive World Story Model via Noise-Aware Error Rectification

Lisai Zhang^1∗, Yidi Wu^1∗, Qi Liu^1,2∗, Xin Ma¹, Yang Ding¹, Gang Yue¹, Siqian Yang¹, Jingyuan Chen²,
Lin Ma^†, Yaohui Wang^1†

¹ Vorch Team ² Zhejiang University

*Equal contribution †Corresponding author

Paper Code (Coming soon)

Scroll to explore

Selected results

The Antique Shop

Subject references

Across the Snowline

Subject references

Case two subject reference one — Subject 01

Case two subject reference two — Subject 02

The Dark Forest (3 Body Problem)

Subject references

Shot 1

Shot 2

Shot two subject reference one — Miao Wang

Shot two subject reference two — Qiang Shi

Shot 3

Shot three subject reference one — Miao Wang

Shot three subject reference two — Qiang Shi

Shot 4

Shot four subject reference one — Wenjie Ye

Shot four subject reference two — Ji Luo

Shot 5

Shot five subject reference one — Yi Ding

Shot five subject reference two — Human Fleet

Shot five subject reference three — Droplet

More stories are on the way. New generations and baseline comparisons will be added here.

Method overview

Autoregressive generators are trained on clean history, but at inference they must continue from their own imperfect outputs. That gap compounds into identity drift, blur, and audio-visual desynchronization.

Vorch-Director makes training history imperfect in the right way: it reuses the model’s prediction residuals, matches them to the current flow noise level σ, and preserves a clean one-second anchor for stable long-horizon generation.

Vorch-Director method overview showing sigma-aware residual injection, clean sink, and joint audio-video transformer.

01 / σ-aware correction

Match error to the denoising regime.

Each residual is stored with the noise level at which it arose. Training samples from the matching neighborhood, so coarse and fine errors are corrected at the right intensity.

02 / clean anchor

Keep one trustworthy memory.

A clean one-second sink and subject references remain untouched while generated history is corrupted, giving the model a stable cue for identity and appearance.

Long-horizon evaluation

Method	Cross-Shot Consistency				Video Quality		Text Consist.
Method	ViCLIP ↑	Self-CIDS ↑	ARC ↑	Reappear ↑	Aesthetic ↑	Imaging ↑	CLIP-T ↑
StoryMem	0.5645	0.4816	0.4034	0.3940	0.5992	0.6350	0.2815
HoloCine	0.4666	0.4871	0.3329	0.3399	0.5557	0.5178	0.2945
Memento	0.5258	0.4883	0.4003	0.3978	0.5979	0.6291	0.3063
JoyAI-Echo	0.5900	0.4567	0.3961	0.4114	0.5068	0.6696	0.2840
Vorch-Director (ours)	0.7887	0.6348	0.6570	0.6648	0.5969	0.5950	0.2444

ST-Bench comparison (30 cases) under our unified metric suite. All rows are computed by the same evaluator under identical settings; no published numbers are mixed in. ST-Bench has no audio track, so speech metrics are omitted. Best per column in bold.

Method	Cross-Shot Consistency				Video Quality		Text	Speech
Method	ViCLIP ↑	Self-CIDS ↑	ARC ↑	Reappear ↑	Aesthetic ↑	Imaging ↑	CLIP-T ↑	Voice ↑	Acc ↑
Text-to-video (T2V)
StoryMem	0.4997	0.4847	0.3743	0.3742	0.5415	0.5853	0.2818	n/a	n/a
HoloCine	0.7079	0.6211	0.6048	0.5978	0.4941	0.3413	0.2696	n/a	n/a
Memento	0.5356	0.4947	0.4402	0.4125	0.5343	0.5771	0.2789	n/a	n/a
JoyAI-Echo	0.5497	0.4603	0.4020	0.4065	0.3850	0.6436	0.2761	0.6193	0.2376
Vorch-Director (ours)	0.7335	0.6031	0.6544	0.6605	0.5533	0.5479	0.2538	0.6358	0.2403
Reference-to-video (R2V)
DreamID-Omni	0.6770	0.5368	0.5453	0.5581	0.5913	0.7232	0.2671	0.6191	0.1979
Vorch-Director (ours)	0.7422	0.6332	0.6861	0.6756	0.5820	0.6625	0.2551	0.5450	0.2432

UnityShots comparison under the unified metric suite, T2V (top block) and R2V (bottom block), 200 cases each. Visual-only baselines emit no audio track, so their speech metrics are "n/a". Best per column within each block in bold.

Research artifact

Read & cite.

Full method, training recipe, benchmark design, and implementation details are available in the paper.

Open PDF ↗

BibTeX

@article{zhang2026vorchdirector,
  title={Vorch-Director: Noise-Aware Error Correction
    for Long Audio-Visual Video Generation},
  author={Zhang, Lisai and Wu, Yidi and Liu, Qi and
    Ma, Xin and Ding, Yang and Yue, Gang and Yang, Siqian and
    Chen, Jingyuan and Ma, Lin and Wang, Yaohui},
  year={2026}
}