emerging
Rectified-CFG++ for Flow Based Models
Shreshth Saini, Alan C. Bovik · 2025-10 · arXiv:2510.07631
Why it matters
Reworks classifier-free guidance specifically for rectified-flow models, where the original diffusion-era formulation is mismatched to straight-path transport and degrades samples at high guidance scales.
What this paper does
read: full textBefore this
Classifier-free guidance was carried over unchanged from diffusion models to rectified-flow models, linearly extrapolating the unconditional velocity toward the conditional velocity the same way it extrapolates score functions in diffusion.
The problem
That extrapolation is mismatched to deterministic rectified flows. Diffusion's stochastic sampling regularizes the trajectory back toward the data manifold, while a rectified flow's deterministic ODE integration provides no such correction, so standard CFG pushes samples off the manifold and produces color blow-outs, warped geometry and text misalignment at high guidance scales.
The idea
Replace linear extrapolation with a predictor-corrector scheme. First take a conditional rectified-flow step to keep the sample anchored on the learned transport path, then apply a scheduled, decaying interpolation between conditional and unconditional velocities rather than an unbounded extrapolation.
How it works
This is a training-free, inference-time sampling algorithm, not a new base generator. Each step first computes a predictor update using only the conditional velocity field, moving the sample halfway toward the next timestep to anchor it near the transport path. A corrector then evaluates conditional and unconditional velocities at that predicted point and adds a time-scheduled correction term, decaying as sampling approaches the data, built from the difference between those two velocities rather than a full linear mix of them. The authors prove the resulting velocity field is marginally consistent and that trajectories stay within a bounded tubular neighborhood of the data manifold, with the bound scaling with training error and numerical step size. It was applied unmodified, with no retraining, to four already-trained flow models, Flux, Stable Diffusion 3, Stable Diffusion 3.5 and Lumina-Next.
Evidence
On MS-COCO 10K, Rectified-CFG++ improves FID over standard CFG for all three models tested at matched settings, Flux from 37.86 to 32.23, Lumina from 26.93 to 22.49, and SD3 from 23.89 to 23.39, while also improving or holding HPSv2 and CLIP scores. On T2I-CompBench, it improves compositional attribute binding, texture rises from 0.7413 to 0.7627 and shape from 0.5792 to 0.6014. Standard CFG degrades sharply once guidance scale exceeds about 7.5, while Rectified-CFG++ stays stable across a wide range of scales in the paper's sweep. At 512x512 with 20 function evaluations, Rectified-CFG++ reaches FID 74.47 against CFG's 85.82 at 28 function evaluations, matching or beating quality with fewer sampling steps.
Limitations
The authors concede the method is scoped to flow-based models, with applicability to diffusion models left unclear, and that failure cases are only briefly catalogued rather than analyzed in depth at extreme guidance scales. The correction schedule still carries its own hyperparameters, a maximum weight and a decay exponent, even though the paper frames the method as adding no tuning beyond the guidance scale itself.
Why it matters
Guidance strength is the main lever practitioners use to trade fidelity for prompt adherence, and this work removes the specific failure mode, off-manifold drift, that made pushing that lever hard on the rectified-flow models now used by most leading text-to-image systems, without retraining any of them.
Abstract, in the authors' own words
Classifier-free guidance (CFG) is the workhorse for steering large diffusion models toward text-conditioned targets, yet its native application to rectified flow (RF) based models provokes severe off-manifold drift, yielding visual artifacts, text misalignment, and brittle behaviour. We present Rectified-CFG++, an adaptive predictor-corrector guidance that couples the deterministic efficiency of rectified flows with a geometry-aware conditioning rule. Each inference step first executes a conditional RF update that anchors the sample near the learned transport path, then applies a weighted conditional correction that interpolates between conditional and unconditional velocity fields. We prove that the resulting velocity field is marginally consistent and that its trajectories remain within a bounded tubular neighbourhood of the data manifold, ensuring stability across a wide range of guidance strengths. Extensive experiments on large-scale text-to-image models (Flux, Stable Diffusion 3/3.5, Lumina) show that Rectified-CFG++ consistently outperforms standard CFG on benchmark datasets such as MS-COCO, LAION-Aesthetic, and T2I-CompBench. Project page: https://rectified-cfgpp.github.io/
Research line
Design-axis choices
Inference
Builds on
Nothing recorded yet.
Built on by
Nothing recorded yet.
This is a emerging entry — a short-form summary. Full paper-page explanations (before/problem/ core idea/representation/architecture/objective/conditioning/training/inference/results/ ablations/limitations) are written for landmark and core papers first; see PROJECT_STATE.md for the schedule.