TL;DR
Any Forcing is a new framework for autoregressive video generation that achieves "Any-Step Generality" and "One-Step Stability." By combining a multi-task adversarial discriminator with a smart, training-free memory selection mechanism, it enables high-quality video generation at real-time speeds (up to 24.7 FPS on an H100) while preventing the "content drift" and visual collapse typically seen in long-sequence, one-step generation.
What Problem Does This Paper Address?
Autoregressive video generation is the "streaming" approach to AI video: it predicts the next few frames based on what it has already generated. While efficient, current models face a "quality vs. speed" catch-22.
- Fixed Sampling: Most models are tuned for a specific number of denoising steps. If you try to speed them up by reducing steps, quality falls off a cliff.
- One-Step Instability: The fastest possible mode—generating a frame in a single step—is notoriously unstable. Small errors in one frame get magnified in the next, leading to "content drift" (where the subject transforms into something else) or "visual collapse" (where the video turns into static or a blurry mess) over long sequences.
Methods
The authors introduce two primary innovations to solve these issues:
1. Unified Adversarial Discriminator
Instead of a simple training objective, they use a discriminator with three specialized "heads" to guide the generator:
- Real-Distribution Alignment (RA): Ensures the generated video looks like real footage.
- Inter-Step Consistency Alignment (CA): Forces the model to produce similar results whether it's taking 1 step or 4 steps, enabling "any-step" flexibility.
- First-Frame Enhancement (FE): Specifically polishes the very first frame. Since autoregressive models are a chain, a stronger "link zero" prevents errors from cascading later.
2. Adaptive Long-Term Memory Selection
To keep long videos stable without blowing up the computational budget, the model doesn't just look at the most recent frames. It maintains a "compressed" memory cache using:
- Global Semantic Relevance: Keeps frames that define the overall scene (e.g., the background).
- Local Value Contribution: Keeps frames that help predict the immediate next motion.
- Position Mapping: A clever trick that remaps the timestamps of these "saved" frames so they always stay within the model's trained "horizon," preventing the positional encoding from glitching out during long generations.
Datasets
The model was trained using real video latents and teacher trajectories from Causal Forcing (which is based on the Wan2.1 architecture). For evaluation, the authors used:
- VBench: For 5-second short-video quality.
- VBench++: For 30-second and 60-second long-video consistency.
- Human Preference Study: 48 evaluators performed blind pairwise comparisons against state-of-the-art baselines.
Key Insights and Ideas
- The "First Frame" Multiplier: One of the most significant findings was that enhancing the first frame (FE) has a disproportionately large impact on the quality of the entire 60-second sequence. It provides a "clean slate" for the rest of the autoregressive process.
- Smart Pruning > Raw Context: Simply giving a model a massive "window" of past frames is inefficient. By filtering memory based on both global and local relevance, Any Forcing maintains better subject identity (like keeping a cat's markings consistent) than models that just look at the last 10 frames.
- Training-Free Stability: The memory selection mechanism is "training-free," meaning it can be plugged into existing autoregressive pipelines to improve their long-term stability without retraining the entire core model.
Key Results
- Speed: Any Forcing is incredibly fast, achieving 20.9 FPS for short videos and 24.7 FPS for long videos on a single NVIDIA H100 GPU.
- Visual Quality: In one-step mode, it outperformed major baselines (like ASD and One Forcing) across almost all VBench metrics, particularly in "Object Class" and "Aesthetic Quality."
- Long-Horizon Stability: While other one-step methods collapsed after a few seconds, Any Forcing successfully generated stable 60-second videos.
- Human Preference: In blind tests, users preferred Any Forcing over competitors by significant margins (e.g., an 83.3% win rate against ASD in the one-step setting).
Limitations and Future Work
The authors note that their attention-based memory selection, while powerful, still struggles with extremely complex scene dynamics (e.g., very fast camera cuts or chaotic motion). Because the memory is selected based on historical "clean" states, it may occasionally miss sudden, unpredictable changes in a scene. Future work will likely focus on more robust memory strategies that can better handle high-entropy motion and complex "scene-switching" logic.
Summary updated