[CVPR 2026] MMM: When Mode Seeking Meets Mean Seeking for Fast, High-Fidelity Long Video Generation

TL;DR

Generating minute-long videos is hard because long-form data is rare, leading to a choice between "blurry but coherent" (SFT) or "sharp but drifting" (Autoregressive) models. MMM solves this by decoupling the task: a shared backbone learns long-range context, while two separate "heads" handle global structure (Mean Seeking) and local sharpness (Mode Seeking) simultaneously. The result? Sharp, consistent 60-second videos generated in just 4 steps.

The Fidelity-Horizon Gap: Why Long Video is Not Just "Longer"

In image generation, moving from 256p to 1024p is largely a matter of spatial interpolation—the local statistics remain similar. Authors argue that video is different: a 1-minute video is an extrapolation of a 5-second clip. It adds new causal chains, narrative structures, and events.

Current SFT (Supervised Fine-Tuning) approaches on "mixed-length" data suffer because the model tries to "average" the scarce long-video data, resulting in a Mean Seeking behavior that washes out fine details. Conversely, teacher-distillation methods are "blind" to long-term logic.

Methodology: The Decoupled Diffusion Transformer (DDT)

To resolve the conflict between global coherence and local fidelity, the researchers introduced a dual-head architecture:

1. Shared Long-Context Encoder ($E_{\varphi }$): A Transformer backbone that processes the entire long-video latent. It learns a unified spatiotemporal representation.
2. Flow Matching (FM) Head: This head uses standard Supervised Fine-Tuning (SFT) on real long videos. It is "Mean Seeking," aiming to capture the average narrative flow of the data.
3. Distribution Matching (DM) Head: This is the "secret sauce." It takes sliding windows of the student’s output and aligns them to a frozen expert short-video teacher using a reverse-KL divergence. Because reverse-KL is Mode Seeking, it forces the student to pick the "sharpest" possible local textures that the teacher (which saw web-scale data) knows how to produce.

The Math of Mode Seeking

The gradient for the DM head leverages DMD (Distribution Matching Distillation). Instead of comparing pixels, it compares the "velocities" (scores) of the student and the teacher. If the student starts to produce a blurry "average" frame, the teacher's score pushes it back toward a high-fidelity mode.

Experimental Results: Closing the Gap

The team tested MMM on the Wan-1.3B and 14B models. The quantitative results in VBench-Long confirm that MMM achieves the best balance of subject consistency and imaging quality.

• Speed: By utilizing the DM head as a distilled sampler, MMM performs inference in just 4 steps (NFE), compared to 50 steps for standard SFT models.
• Quality: Unlike "LongSFT" which loses texture, MMM maintains the identity of subjects and background sharpness over 30+ seconds.

Ablation Insights

The ablation study proved that the DDT Dual-Head design is critical. Training a single head with both objectives leads to gradient interference, where the "mean-seeking" SFT signal and the "mode-seeking" DM signal fight each other, resulting in mediocre performance on both fronts.

Critical Analysis & Conclusion

MMM demonstrates that we don't necessarily need "web-scale" long-video data to generate high-quality long videos. We can borrow the local intelligence of short-video models (the local "physics") and combine it with the global "logic" of smaller long-video datasets.

Limitations: The current model is bidirectional. While great for offline generation, future work needs to convert this into a causal (autoregressive) format for interactive world modeling. However, the paradigm of "Mode + Mean Seeking" provides a robust blueprint for any generative task suffering from data scarcity at scale.