EPFL Stable Video Infinity: How Error Recycling Solves AI Video's Biggest Problem

Every AI video model has the same dirty secret. Generate a 4-second clip and the results look stunning. Push past 15 seconds and characters start morphing, physics breaks down, colors shift, and the whole scene drifts into incoherence. EPFL's VITA Lab just published a solution, and it might be the most important paper in video generation this year.

The Drift Problem Nobody Solved

If you have tried generating long-form AI video with any current model, you know the frustration. Sora 2 caps at 15 to 25 seconds depending on your plan. Runway Gen-4.5 maxes out at 10. Kling 3.0 pushes to 15. The reason is not compute or memory. It is temporal drift.

💡

Temporal drift occurs when autoregressive video models accumulate small errors frame by frame. Each generated frame becomes the input for the next, and tiny imperfections compound exponentially. After a few hundred frames, the output barely resembles the original prompt.

This is fundamentally similar to the "telephone game" problem. Whisper a message through enough people and it becomes unrecognizable. Previous approaches tried to fight drift by generating shorter clips and stitching them together, but the seams are visible. Others used hierarchical generation (keyframes first, interpolation second), which helps but does not eliminate the core issue.

Enter Error Recycling

The paper, titled "Stable Video Infinity" and accepted as an ICLR 2026 Oral presentation, introduces a deceptively simple idea: teach the model to handle its own mistakes.

Oral

ICLR 2026 Status

Extra Compute Cost

Minutes

Video Length

Here is the key insight. During training, standard video diffusion models learn from clean ground-truth data. They never see their own generated output. When deployed, they suddenly have to work with their own imperfect predictions as input, and they do not know how to handle the noise.

Error recycling fixes this by modifying the training loop:

Step 1

Standard Forward Pass

The model generates a video segment from clean training data.

Step 2

Error Collection

The model's own predictions (with their imperfections) are captured instead of discarded.

Step 3

Error Recycling

These imperfect outputs are fed back as input for the next training iteration, teaching the model to generate stable output even from noisy, self-generated frames.

Step 4

Iterative Refinement

Over many training cycles, the model learns a robust self-correction mechanism that prevents error accumulation.

The beauty of this approach is its elegance. There are no new model architectures, no additional parameters, no inference-time tricks. The model simply learns during training what real deployment conditions look like.

Why This Matters More Than You Think

The implications extend far beyond generating longer cat videos. Here is what changes:

Before Error Recycling

Video models limited to 4-20 seconds
Scene consistency degrades rapidly
Character identity drifts after a few seconds
Physics becomes increasingly unrealistic
Color and lighting shift unpredictably

After Error Recycling

Multi-minute coherent video generation
Stable character appearance throughout
Consistent physics and spatial relationships
Reliable color grading and lighting
No visible quality degradation over time

The Numbers

The VITA Lab team tested Stable Video Infinity across multiple architectures and benchmarks. The results are striking:

Metric	Without Error Recycling	With Error Recycling	Improvement
FVD (lower is better)	Degrades after 4s	Stable through 2min+	Significant
Character Consistency	Drops below 60% at 15s	Maintains 90%+ at 2min	~50% relative
Temporal Coherence	Visible drift at 8s	No visible drift at 2min	Qualitative leap
Compute Overhead	Baseline	Baseline (0% increase)	Zero cost

💡

The zero-compute-overhead aspect is critical. Error recycling is a training-time technique, not an inference-time one. Once trained, the model runs at exactly the same speed and cost as before.

How Error Recycling Works Under the Hood

For those who want the technical details, here is what happens in the modified training pipeline.

Standard video diffusion training uses a noise schedule epsilon applied to clean ground-truth frames x_0. The model learns to predict the noise and recover the original signal. At inference time, the model autoregressively generates frame t+1 conditioned on its prediction of frame t.

The gap between training (clean inputs) and inference (self-generated inputs) is called exposure bias. Error recycling directly addresses this.

Technical Details: Modified Training Objective▼

During training, the model periodically generates frames using its own current weights rather than using ground-truth data. These self-generated frames, complete with their imperfections, are then used as conditioning inputs for subsequent frames in the training sequence.

The key hyperparameter is the recycling ratio r, which controls how often self-generated frames replace ground-truth frames during training. The paper finds that r = 0.3 (30% recycled frames) achieves optimal performance, balancing stability improvement with training signal quality.

The modified loss function remains the standard diffusion objective. The only difference is the data distribution the model sees during training. This is why there is no computational overhead at inference time.

This is conceptually similar to scheduled sampling in sequence-to-sequence models, but adapted for the continuous diffusion framework. The VITA Lab team credits this connection in their paper while noting that the naive application of scheduled sampling to diffusion models does not work. Their contribution is the specific formulation that makes it stable for video generation.

The Open-Source Advantage

Perhaps the most exciting aspect: the code is fully open-source on GitHub (vita-epfl/Stable-Video-Infinity). This means any research lab or company can apply error recycling to their existing video models.

✓Why Open-Source Matters

Any existing video diffusion model can be retrofitted with error recycling. No architectural changes needed, just a modified training loop. This could improve Sora, Runway, Kling, and every open-source model simultaneously.

✗Practical Limitations

Requires retraining or fine-tuning the model. Companies with proprietary models need to invest compute in retraining. The technique's effectiveness may vary across different architectures and scales.

The open-source release follows a growing trend in the AI video research community. Models like LTX-2 and Wan 2.6 have shown that open-source approaches can compete with proprietary alternatives.

What This Means for the Industry

The timing is remarkable. We are in the middle of an AI video arms race where every major player, from Runway to ByteDance, is pushing for longer, higher-quality output. Error recycling could be the missing piece that unlocks the next generation of capabilities.

🎬

For Filmmakers and Creators

Long-form coherent video generation enables actual narrative content. Not just clips, but scenes, sequences, and eventually short films generated from a single prompt.

🔬

For Researchers

Error recycling provides a generalizable framework. The same principle could apply to audio generation, 3D scene synthesis, and any autoregressive generative model that suffers from drift.

🏢

For Enterprise

Stable long-form generation makes AI video viable for professional use cases: product demos, training videos, marketing campaigns that require consistent quality across minutes of content.

Looking Ahead

The VITA Lab's work represents a fundamental shift in how we think about AI video generation. Instead of building ever-larger models or adding complex inference-time mechanisms, the solution was to simply show the model what its own output looks like.

The paper will be presented at ICLR 2026, and several industry sources suggest that major video model providers are already exploring integration. If world models represent the future of how AI understands physics and space, error recycling represents the future of how AI maintains that understanding over time.

✅

The age of 4-second AI videos may be ending sooner than anyone expected. And it will not require a new model architecture or a billion-dollar compute budget. Just a smarter training loop.

Sources

International Conference on Learning Representations: Oral (ICLR 2026 Status) (International Conference on Learning Representations)
EPFL VITA researchers via arXiv: Stable Video Infinity supports infinite-length video generation with no additional inference… (EPFL VITA researchers via arXiv)