Feature request
Add community pipelines that sample rectified-flow models with Rectified-CFG++ (Saini, Gupta, Bovik; NeurIPS 2025): paper, reference implementation.
Rectified-CFG++ is a training-free replacement for the classifier-free guidance extrapolation in flow samplers. Each ODE step becomes a predictor-corrector update: the predictor steps along the conditional velocity, the corrector evaluates the conditional and unconditional velocities at that predicted point, and the effective velocity is the conditional velocity at the current point plus the scaled corrector difference. High guidance scales keep prompt alignment without the over-saturation and structural artifacts of standard CFG. On SD3.5, Flux-dev, and Lumina-Image-2.0 the paper reports better FID, CLIP score, and ImageReward than CFG at the same number of steps.
Why a community pipeline
The method needs a second model evaluation at the predicted point inside the step, so it does not fit the guiders API, which recombines predictions from a single evaluation. A community pipeline per backbone (generated from the upstream pipeline with only the denoising loop replaced) keeps everything else, LoRA, IP-Adapter, skip-layer guidance, callbacks, working unchanged. Happy to move to modular denoise blocks if maintainers prefer that route.
Scope
Backbones from the paper: Stable Diffusion 3 / 3.5, Flux, Lumina-Image-2.0. Beyond the paper: Qwen-Image and Wan 2.1 / 2.2 (video), which use the same flow-matching sampling.
I am the first author of the paper. PR #14726 implements this; opening the issue for coordination as the contribution guide asks.
Feature request
Add community pipelines that sample rectified-flow models with Rectified-CFG++ (Saini, Gupta, Bovik; NeurIPS 2025): paper, reference implementation.
Rectified-CFG++ is a training-free replacement for the classifier-free guidance extrapolation in flow samplers. Each ODE step becomes a predictor-corrector update: the predictor steps along the conditional velocity, the corrector evaluates the conditional and unconditional velocities at that predicted point, and the effective velocity is the conditional velocity at the current point plus the scaled corrector difference. High guidance scales keep prompt alignment without the over-saturation and structural artifacts of standard CFG. On SD3.5, Flux-dev, and Lumina-Image-2.0 the paper reports better FID, CLIP score, and ImageReward than CFG at the same number of steps.
Why a community pipeline
The method needs a second model evaluation at the predicted point inside the step, so it does not fit the
guidersAPI, which recombines predictions from a single evaluation. A community pipeline per backbone (generated from the upstream pipeline with only the denoising loop replaced) keeps everything else, LoRA, IP-Adapter, skip-layer guidance, callbacks, working unchanged. Happy to move to modular denoise blocks if maintainers prefer that route.Scope
Backbones from the paper: Stable Diffusion 3 / 3.5, Flux, Lumina-Image-2.0. Beyond the paper: Qwen-Image and Wan 2.1 / 2.2 (video), which use the same flow-matching sampling.
I am the first author of the paper. PR #14726 implements this; opening the issue for coordination as the contribution guide asks.