🤖 AI

BRACE: Taming Sharp Irregularities via Barycentric Rational Forecasting for Fast Diffusion Transformers Inference

The paper introduces BRACE, a novel inference acceleration method for Diffusion Transformers that replaces unstable derivative-based polynomial extrapolation with a numerically stable barycentric rational forecasting approach using Chebyshev weights to effectively handle sharp feature irregularities while maintaining high generation quality.

Original authors: Jinlong Yang, Jinke Wu, Lizilin, Yao Zhou

Published 2026-08-11

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Jinlong Yang, Jinke Wu, Lizilin, Yao Zhou

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 are trying to paint a masterpiece, but instead of using a brush, you are using a robot that has to take thousands of tiny, careful steps to finish the picture. This robot is a "Diffusion Transformer," a type of artificial intelligence that creates stunning images and videos by slowly turning random static noise into clear pictures. The problem is that this robot is incredibly slow; it takes a long time to take all those steps, making it hard to use for real-time fun like chatting with an AI or generating videos on the fly. To speed things up, scientists have tried to tell the robot, "Hey, you don't need to calculate every single step; just guess the next few based on what you did a moment ago." This is called "feature caching." However, the old guessing methods were like trying to predict a rollercoaster's path by drawing a straight line through a few points. If the track suddenly twisted or looped, the straight-line guess would fly off the rails, resulting in blurry or weird-looking art.

This paper introduces a new way to make that robot faster without ruining the picture quality. The authors, a team from Sichuan University, noticed that while the robot's path is usually smooth, it occasionally hits sudden, sharp bumps or twists. The old methods, which relied on calculating "derivatives" (a fancy math way of measuring how fast things are changing), would get confused by these sharp turns and make wild, inaccurate guesses. The team proposes a new method called BRACE (Barycentric Rational Forecasting with Chebyshev Enhancement). Instead of trying to measure the speed of the changes, BRACE looks at the actual history of the robot's steps and uses a special mathematical "rational function" (a type of fraction-based formula) to predict the future. Think of it like a GPS that doesn't just draw a straight line to your destination but uses a flexible, stretchy rubber band to snap perfectly to the winding road, even when the road suddenly curves. By using this flexible approach, BRACE allows the AI to skip many steps at once, speeding up the process significantly while keeping the images sharp and the videos moving smoothly.

The Problem: The "Straight Line" Trap

To understand why this new method is needed, imagine you are driving a car on a road that is mostly straight but has a few sudden, sharp hairpin turns. If you are driving fast and you try to predict where the road will be in ten seconds by drawing a straight line from your current position, you will likely crash into a tree or fly off a cliff. This is exactly what happened with previous AI acceleration methods.

The old way of speeding up Diffusion Transformers was like using a "Taylor expansion," which is a mathematical tool that tries to predict the future by looking at how fast the car is currently moving and how fast that speed is changing (acceleration). For smooth, gentle curves, this works great. But the researchers found that the "road" the AI travels on isn't always smooth. It has "sharp irregularities"—sudden, jarring changes in direction. When the AI tried to use the old "straight line" or "smooth curve" math to predict these sharp turns, the predictions would go wildly wrong. The result? The AI would skip steps, but the image it produced would be distorted, blurry, or have strange artifacts, like a face with too many eyes or a car with extra wheels.

The Solution: BRACE's Flexible Rubber Band

The authors realized that instead of trying to measure the speed of the turn (which is hard to do accurately when the turn is sudden), they should just look at the actual path the car has taken recently and use a smarter way to connect the dots. They call their new method BRACE.

Here is how BRACE works, using a simple analogy:
Imagine you are trying to guess the shape of a wiggly snake by looking at a few of its body segments.

The Old Way (Polynomial): You try to draw a single, rigid stick (a polynomial) that touches all the segments you can see. If the snake suddenly twists, your rigid stick breaks or misses the snake entirely.
The BRACE Way (Rational Forecasting): Instead of a rigid stick, you use a flexible, stretchy rubber band (a rational function). You attach the rubber band to the segments you've seen. Because the rubber band is flexible, it can bend and twist to follow the snake's sharp turns perfectly, even if the turn happens very suddenly.

In technical terms, BRACE uses a "barycentric rational function." This is a fancy math formula that acts like that flexible rubber band. It takes the actual data points (the "features" the AI has already calculated) and combines them in a way that naturally handles sharp changes without getting confused.

Why It's Special: The "Chebyshev" Secret Sauce

To make this rubber band work even better, the authors added something called "Adapted Chebyshev weights." Think of this as a special tension setting on your rubber band. If you pull a rubber band too hard at the ends, it might snap or stretch unevenly. The Chebyshev weights are a mathematical trick that ensures the rubber band stays balanced and stable, no matter how much it has to stretch to predict the future.

The paper shows that this method is incredibly stable. Even when the AI is told to skip a huge number of steps (like jumping from step 1 to step 20 in one go), the "rubber band" doesn't break. It stays on the path, ensuring the final image looks just as good as if the AI had taken every single step slowly.

What the Experiments Showed

The researchers tested BRACE on three different types of AI tasks:

Creating Images from Text: They used a model called FLUX.1-dev. When they asked the AI to generate complex images (like a shark in a desert or a storefront with specific text), the old methods often messed up the text or blurred the details. BRACE, however, kept the text sharp and the details clear, even when speeding up the process by more than 5 times.
Creating Videos: They tested it on a video model called HunyuanVideo. In video generation, things move, and if the prediction is wrong, the movement looks jerky or the objects disappear. BRACE managed to keep the movement smooth and the objects (like a running horse or a swimming person) looking natural, outperforming other fast methods.
Standard Image Generation: On a classic test called ImageNet, BRACE produced images with the highest quality scores (measured by FID, a score where lower is better) compared to other fast methods.

In their tests, BRACE was able to speed up the AI by factors of 3.5x to 5.5x. For example, on the FLUX.1 model, it achieved a speedup of 5.55x while still producing images that looked almost identical to the slow, high-quality version. The researchers noted that while other methods started to fail and produce "ghosting" or "distortions" at these high speeds, BRACE remained stable.

The Bottom Line

The paper doesn't claim to have solved every problem in AI, but it does suggest a very strong new direction. The authors argue that the old way of thinking about speeding up AI—using rigid, derivative-based math—was fundamentally flawed for handling the "sharp turns" in the AI's learning path. By switching to a flexible, rational approach (BRACE), they found a way to make these powerful image and video generators much faster without sacrificing the quality of the art they create.

It's a bit like realizing that to navigate a winding mountain road, you don't need a straighter road; you just need a better suspension system. BRACE provides that suspension, allowing the AI to zoom through the steps while staying firmly on the path to a beautiful result.

Technical Summary: BRACE

Problem Statement

Diffusion Transformers (DiTs) have achieved state-of-the-art performance in high-fidelity image and video generation but suffer from prohibitive computational overhead due to their iterative sampling processes. To address this, feature caching has emerged as a training-free acceleration strategy. While traditional "cache-and-reuse" methods degrade quality over long skip intervals, recent "cache-and-forecast" approaches (e.g., TaylorSeer, HiCache) attempt to predict future features using derivative-driven polynomial extrapolation (Taylor expansions).

However, the authors identify a critical bottleneck in these existing methods: DiT feature trajectories, while globally smooth, frequently exhibit sharp irregularities and local non-smoothness (e.g., rapid structural shifts during specific denoising intervals). Derivative-driven polynomial methods rely on finite-difference estimates of instantaneous derivatives. At sharp inflection points, these estimates become unstable, and rigid polynomials lack the structural adaptability to model abrupt shifts. This leads to severe extrapolation divergence, oscillation, and loss of structural integrity when using aggressive skip intervals.

Methodology: BRACE

The authors propose BRACE (Barycentric Rational Forecasting with Chebyshev Enhancement), a training-free framework that shifts the forecasting paradigm from derivative-driven polynomials to feature-driven rational forecasting.

Core Components

Local Sliding Window (FIFO Queue):
BRACE maintains a sparse local sliding window (capacity $C$ ) that caches historical state tuples $(F_\tau, \tau)$ , where $F_\tau$ represents exact intermediate features and $\tau$ is the timestep. This confines the forecasting horizon to a local window, decoupling temporally decayed features and ensuring the trajectory satisfies local smoothness constraints.
Canonical Domain Mapping:
To ensure numerical stability, raw timesteps are mapped to a canonical interval $[-1, 1]$ via an affine transformation. This aligns the diffusion timesteps with the theoretical foundations of the weighting scheme and ensures scale invariance across different schedulers.
Barycentric Rational Extrapolation:
Instead of estimating derivatives, BRACE directly aggregates raw historical features using the second barycentric form of rational interpolation:
$F_{pred} = \frac{\sum_{j=1}^{C} \frac{w_j}{x(t_{pred}) - x(\tau_j)} F_{\tau_j}}{\sum_{j=1}^{C} \frac{w_j}{x(t_{pred}) - x(\tau_j)}}$
This formulation offers two key advantages:
- Self-Normalization: The denominator inherently bounds predicted features, preventing the unbounded divergence common in polynomial extrapolation.
- Non-Smoothness Absorption: By aggregating states rather than relying on unstable high-order derivatives, the rational function naturally accommodates localized non-smoothness.
Adapted Chebyshev Weights:
The weights $w_j$ are crucial for stability. BRACE employs an adapted scheme inspired by Chebyshev–Gauss–Lobatto nodes:
$w_j = (-1)^{j+1} \cdot \delta_j$
where $\delta_j$ includes a boundary sensitivity coefficient $\gamma$ for the terminal weight. The authors note that for ultra-sparse capacities ( $C \le 3$ ), these optimal nodes are mathematically equivalent to equidistant points, allowing for simple fixed-interval sampling without complex dynamic adjustments.

Theoretical Stability

The authors provide an error propagation analysis showing that the extrapolation error in BRACE is governed by the skip interval $k$ and the Lipschitz constant $L$ (first-order derivative). In contrast, Taylor methods incur errors proportional to $s^{m+1} \|F^{(m+1)}\|_\infty$ , which can explode during sharp transitions. BRACE theoretically bounds the error, ensuring predictable stability even under aggressive skipping.

Key Contributions

The BRACE Framework: A novel, training-free rational forecasting framework that synergizes sparse historical caching with adapted Chebyshev weights within a barycentric rational structure. It ensures numerical stability and tames abrupt transitions where derivative-driven methods fail.
Geometric Analysis of DiT Dynamics: The paper empirically reveals the dual geometric nature of DiT evolution (globally smooth but punctuated by sharp irregularities) and validates that rational bases inherently outperform polynomials in fitting these trajectories.
State-of-the-Art Acceleration: Extensive evaluations demonstrate that BRACE consistently outperforms existing baselines (e.g., TaylorSeer, HiCache, FORA) across diverse architectures and modalities.

Experimental Results

BRACE was evaluated on three tasks:

Class-Conditional Image Generation (DiT-XL/2 on ImageNet): BRACE achieved the lowest FID and sFID scores across various skip intervals. In the 3.56× speedup configuration, the optimal window capacity was found to be $C=3$ (FID 2.46), outperforming $C=2$ and $C=4$ . Qualitative results showed better preservation of sharp details and semantic structure compared to the over-smoothing or ghosting artifacts of baselines.
Text-to-Image Generation (FLUX.1-dev): BRACE set new state-of-the-art results in semantic alignment (ImageReward, CLIP Score) and structural fidelity (SSIM, LPIPS). Notably, at a 5.55× speedup, the experiments utilized a window capacity of $C=2$ (with $\gamma=0.7$ ), preserving legible typography and complex text structures that other methods distorted or blurred.
Text-to-Video Generation (HunyuanVideo): BRACE achieved the highest VBench score (80.14) at high skip intervals (using $C=2$ ), significantly outperforming direct reuse and derivative-driven forecasting. It successfully preserved spatiotemporal consistency, avoiding entity loss and structural artifacts (e.g., extra limbs) seen in competing methods.

Ablation Studies:

Window Capacity ( $C$ ): On DiT-XL/2, $C=3$ provided the optimal approximation of feature curvature. Increasing $C$ to 4 led to performance degradation due to accumulated long-range extrapolation noise, while reducing $C$ to 2 weakened fitting capability. However, for the higher-speedup configurations on FLUX and Video, $C=2$ was the configuration used to balance speed and fidelity.
Boundary Sensitivity ( $\gamma$ ): Perceptual quality peaked at $\gamma=0.7$ , while reconstruction metrics improved monotonically toward 1.0.
Weighting Schemes: The adapted Chebyshev weights outperformed Uniform, Berrut, and Floater-Hormann schemes, confirming their suitability for the asymmetric temporal dynamics of DiT features.

Significance

The paper claims that BRACE addresses a fundamental limitation in current acceleration strategies: the inability of polynomial extrapolation to handle the "stiff" system behavior and sharp irregularities inherent in DiT feature trajectories. By shifting to a feature-driven rational forecasting paradigm, BRACE achieves a superior quality-efficiency trade-off with negligible computational overhead. The authors position this work as a structural shift in the feature forecasting paradigm, offering a robust alternative for fast DiT inference that maintains high fidelity even under aggressive acceleration.