💻 computer science

FlowLong: Inference-time Long Video Generation via Manifold-constrained Tweedie Matching

FlowLong is a training-free, architecture-agnostic inference-time method that generates long videos by blending overlapping sliding windows via manifold-constrained Tweedie matching and stochastic early-phase sampling to ensure temporal consistency and visual fidelity without additional model training.

Original authors: Jangho Park, Geon Yeong Park, Gihyun Kwon, Jong Chul Ye

Published 2026-05-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jangho Park, Geon Yeong Park, Gihyun Kwon, Jong Chul Ye

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very talented artist who can draw beautiful, high-definition scenes, but they have a strict rule: they can only work on a single postcard-sized canvas at a time. If you ask them to draw a whole movie, they get tired, the style changes, or the characters start repeating the same movements over and over.

This is the problem with current AI video generators. They are great at making short clips (like a postcard), but when you try to make a long movie, the quality falls apart, the story drifts, or the motion becomes robotic and repetitive.

FlowLong is a new "director" that helps these artists make long movies without needing to retrain them or change their style. It does this using two clever tricks: The Overlap Blend and The Fresh Start.

1. The Overlap Blend (Tweedie Matching)

Imagine you want to paint a long mural, but your artist can only paint a 10-foot section at a time.

The Old Way: You ask the artist to paint the first 10 feet, then move the canvas and paint the next 10 feet. The problem? The end of the first section might look slightly different from the start of the second. The colors might shift, or the character's arm might be in a different position.
The FlowLong Way: You tell the artist to paint the first 10 feet, but also paint the next 10 feet at the same time, making sure the last 2 feet of the first section and the first 2 feet of the second section overlap.
The Magic: FlowLong takes the "cleanest" version of the image from both overlapping sections and blends them together perfectly, like a seamless cross-fade in a movie. This ensures that the transition between the two sections is smooth and consistent. It forces the two separate paintings to agree on what the shared part looks like.

2. The Fresh Start (Stochastic Early-Phase Sampling)

Here is the tricky part: Even if you blend the overlaps perfectly, the artist might still get "stuck" in a rut. If they start the second section with a slightly different idea, their brain might drift back to their original, separate path, causing the movie to look disjointed later on.

The Problem: Think of it like two hikers starting on different trails. Even if they meet at a bridge (the overlap), they might immediately walk back to their own separate paths because of "inertia."
The FlowLong Solution: At the very beginning of the process (when the image is still just a blurry cloud of noise), FlowLong gives the artist a gentle "shake." It injects a little bit of fresh randomness (noise) into the mix.
The Result: This shake breaks the hikers' habit of sticking to their old paths. It forces the two separate sections to mix and mingle while they are still blurry. Once they have agreed on the general direction, FlowLong stops shaking them and lets them walk deterministically (smoothly) to the finish line. This ensures the whole movie stays on the same track without losing the sharp, high-quality details.

Why This Matters

No Retraining Needed: You don't need to teach the artist new tricks. You just give them a new set of instructions on how to organize their work. It works with any video AI model out of the box.
Versatile: It doesn't just work for video. The paper shows it can also:
- Generate video and audio together (like a movie with a soundtrack).
- Turn text into 3D worlds (creating a 3D scene from a description).
Better Quality: Unlike other methods that make videos longer but full of repetitive loops or drifting errors, FlowLong creates long videos that stay consistent, diverse, and high-quality.

In Summary

FlowLong is like a smart production manager for AI artists. Instead of letting them struggle to draw a long movie alone, it breaks the job into overlapping chunks, blends the edges perfectly, and gives them a little nudge at the start to make sure they all stay on the same page. The result is a long, coherent, and high-quality video generated without needing to retrain the underlying AI.

Technical Summary: FlowLong

Problem Statement

Extending the generation horizon of video diffusion models beyond their native training length remains a fundamental challenge. Current video diffusion models, typically trained on short clips due to data scarcity, suffer from severe quality degradation when applied to longer sequences. Existing solutions fall into two categories, both with significant limitations:

Training-free bidirectional extensions: Methods like FIFO-Diffusion, RIFLEx, and UltraViCo extend pre-trained models without additional training but rely on architecture-specific modifications. They suffer from accumulating visual artifacts and consistency degradation as video length increases.
Autoregressive approaches: Methods like CausVid and Self-Forcing generate videos sequentially. While effective, they are prone to exposure bias and temporal drift due to the reuse of KV-caches, often leading to repetitive motion patterns and error accumulation over time. Furthermore, they typically require distillation from a bidirectional teacher model, limiting their applicability to new architectures (e.g., joint audio-video models).

Methodology

The authors propose FlowLong, a novel inference-time framework that generates long videos by harmonizing multiple overlapping short video chunks without any additional training or architectural modifications. The method is grounded in the geometric view of flow-based generative models and treats long video generation as an inverse problem.

The framework operates through two core mechanisms:

1. Tweedie Matching

To ensure temporal consistency across adjacent video chunks, FlowLong enforces a manifold constraint on overlapping regions.

Overlap Constraint: Adjacent chunks $x_k$ and $x_{k+1}$ share an overlap of $O$ frames. The method requires the last $O$ frames of chunk $k$ to coincide with the first $O$ frames of chunk $k+1$ .
Guidance Loss: A guidance loss is defined on the clean data manifold to relax this constraint. It measures the discrepancy between the denoised estimate of the current chunk and the predicted clean sample of the adjacent chunk within the overlap region.
Gradient Correction: During the reverse sampling process, this loss is integrated as a one-step gradient correction on the denoised estimate. This results in a closed-form update called Tweedie matching, which interpolates the predicted clean samples of adjacent chunks in the overlap region. This ensures smooth transitions and enforces manifold constraints without altering the backbone model.

2. Stochastic Early-Phase Sampling

While Tweedie matching aligns chunks, deterministic Ordinary Differential Equation (ODE) sampling can cause trajectories to revert to their original divergent paths due to "inertia" from independent noise initializations.

Hybrid Sampling: To break this inertia, the method introduces stochastic noise during the early, high-noise phase of sampling (where $t$ is large). This perturbs the state after each Tweedie matching correction, effectively remixing trajectories and facilitating cross-chunk mixing.
Transition to ODE: After the early stochastic phase, the method transitions to deterministic ODE sampling. This preserves fine-grained visual fidelity while maintaining the temporal consistency established during the stochastic phase.

Generalizability

The framework is model-agnostic and applies to any flow-based model generating fixed-size windows. It has been demonstrated on:

Text-to-Video: Extending models like Wan 2.1 and LTX-2.
Audio-Video Joint Generation: Extending LTX-2 by applying Tweedie matching and stochastic sampling to both video and audio latents simultaneously.
Text-to-3DGS: Extending VIST3A (which combines Wan 2.1 with AnySplat) to generate longer 3D Gaussian Splat scenes from text.

Key Contributions

FlowLong Framework: A training-free, architecture-agnostic solution that extends pre-trained flow-based diffusion models beyond their native horizon purely at inference time.
Tweedie Matching: A technique that enforces manifold constraints and temporal consistency by blending predicted clean samples across overlapping segments.
Stochastic Early-Phase Sampling: A strategy that breaks per-window trajectory inertia by injecting noise in the high-noise regime before switching to deterministic ODE sampling.
Versatility: The method is validated across text-to-video, audio-video joint generation, and text-to-3DGS, outperforming baselines without fine-tuning.

Experimental Results

The authors evaluated FlowLong on 30-second and 60-second video generation tasks using the VBench benchmark (covering aesthetic quality, motion smoothness, temporal flickering, etc.) and text-to-3DGS generation.

Video Generation: FlowLong consistently outperformed both training-free bidirectional baselines (RIFLEx, UltraViCo) and autoregressive baselines (CausVid, Self-Forcing, LongLive).
- It achieved superior Dynamic Degree and Motion Smoothness, avoiding the repetitive motion patterns common in autoregressive models.
- It eliminated the error drift and pixel saturation observed in other methods as video length increased.
- It supported multi-shot generation with global and local prompts, maintaining semantic coherence across scene transitions.
Text-to-3DGS: When applied to VIST3A, FlowLong generated 1.64 $\times$ more 3D Gaussians per scene than the baseline. The generated 3D scenes exhibited richer viewpoint diversity and higher geometric confidence scores (mean confidence logit increased from 26.27 to 41.52).
Ablation Studies: Experiments confirmed that the combination of Tweedie matching and stochastic early-phase sampling yields the best balance of temporal consistency and visual quality. Full ODE sampling lacked consistency, while full SDE sampling degraded visual quality.

Significance and Claims

The paper claims that FlowLong provides a versatile, plug-and-play framework for extrapolating the generative horizon of any flow-based model. Its primary significance lies in its ability to:

Generate videos several times longer than the native window length without any additional training or architectural modification.
Overcome the exposure bias and motion repetition inherent in autoregressive approaches by avoiding KV-cache reuse.
Seamlessly extend to complex tasks like joint audio-video generation and 3D scene generation, where existing methods struggle due to architectural dependencies.

The authors acknowledge a limitation: the overlap-based consistency constraint is inherently local, which may hinder global semantic coherence in extremely long videos, a challenge they identify for future work.