Technical Report | August 2026 | Echo Team @ Joy Future Academy, JD

JoyAI-Echo-1.5: Long-Horizon Audio-Visual Generation for Persistent Stories

A long-video system that preserves character appearance, speaker identity, narrative continuity, and synchronized sound across multi-shot stories.

Paper Code

10 min+ long-horizon generation

A/V Memory cross-shot identity & voice

Abstract

Video generation is moving beyond isolated clips toward persistent long-form narratives, where characters, voices, events, and sound must remain coherent across many shots. We present the long-video variant of JoyAI-Echo-1.5, a unified audio-visual generation system built around composable cross-shot memory. It aggregates visual evidence from multiple prior shots and speaker cues from speech-filtered full-shot audio, preserving appearance and voice identity while supporting flexible text, image, and memory conditioning. A causal few-step generator, trained with progressive audio-visual teacher forcing and short- and long-horizon Self-Gradient Forcing on self-generated rollouts, improves stability over extended generation. On a benchmark of 100 stories and 3,000 shots, JoyAI-Echo-1.5 achieves the best result on six of seven automatic metrics, covering cross-shot consistency, imaging quality, text alignment, and speech fidelity. Human evaluation further favors JoyAI-Echo-1.5 over HappyOyster across all five long-video dimensions, with the largest margin in audio-visual synchronization.

Key Conclusions

Identity persists across distant shots.

Composable visual and speech-filtered audio memory delivers the strongest ViCLIP, Self-CIDS, and speaker-consistency scores among all evaluated systems.

Consistency does not trade away instruction following.

JoyAI-Echo-1.5 reaches the best CLIP score of 0.2868 while preserving changing actions, scenes, camera views, and shot-specific narrative details.

Audio and video remain one coherent timeline.

Joint generation achieves 0.9674 speech recall and a 28.1-point human-preference lead over HappyOyster in audio-visual synchronization.

Few-step generation remains stable over long rollouts.

An 8-step distilled generator combines causal audio-visual teacher forcing with short- and long-horizon Self-Gradient Forcing on self-generated histories.

Results

Evaluation on 100 stories and 3,000 shots measures cross-shot consistency, quality, text alignment, speech fidelity, and human preference.

Long-horizon automatic evaluation

JoyAI-Echo-1.5 ranks first on six of seven metrics and second on aesthetic quality. All metrics are higher-is-better.

Method	Cross-Shot Consistency			Video Quality		Text	Speech
Method	ViCLIP	Self-CIDS	Voice	Aesthetic	Imaging	CLIP	Recall
JavisDiT++	0.6955	0.5621	0.7933	0.5019	0.6084	0.2522	0.0951
LTX-2.3	0.7047	0.6135	0.6847	0.5436	0.4751	0.2559	0.9304
LTX-2.5	0.7775	0.6640	0.7603	0.6019	0.5804	0.2683	0.9545
ShotStream + MMAudio	0.7988	0.7308	0.7945	0.5255	0.6044	0.1998	0.0297
StoryMem + MMAudio	0.7897	0.7491	0.7643	0.5265	0.6320	0.2368	0.0301
HappyOyster (Directing)	0.7535	0.6940	0.7705	0.5606	0.6701	0.2544	0.1085
JoyAI-Echo-1.0	0.8026	0.7793	0.8129	0.5679	0.7058	0.2658	0.9489
JoyAI-Echo-1.5	0.8264	0.7937	0.8524	0.5705	0.7467	0.2868	0.9674

Human evaluation against HappyOyster

Good-Similar-Bad pairwise comparison for long-horizon audio-visual generation.

Aspect	JoyAI-Echo-1.5	Similar	HappyOyster
Story-instruction following	38.6%	43.8%	17.7%
Dialogue/background-audio following	39.2%	37.3%	23.5%
Memory/identity consistency	24.8%	59.5%	15.7%
Audio-visual synchronization	48.4%	31.4%	20.3%
Audio quality	44.4%	22.2%	33.3%

Authors

Echo Team

BibTeX

@techreport{joyai2026echo15,
  title        = {Long-Horizon Audio-Visual Generation for Persistent Stories and Interactive Worlds},
  author       = {{Echo Team @ Joy Future Academy, JD}},
  institution  = {Joy Future Academy, JD},
  year         = {2026},
  month        = {August}
}