Plain-English meaning
Rectified flow describes how a generative model can learn a path from a simple starting distribution, usually noise, to the distribution of its training data. The original method trains a neural ordinary differential equation to follow straight connections between paired samples as closely as possible. Straighter trajectories can be approximated with coarser numerical integration, although a real model's quality and useful step count still depend on its training and release.
The term names a training and generation formulation. It does not name the network that predicts the velocity field. A rectified-flow image model can use a Diffusion Transformer, but rectified flow and DiT are not synonyms.
Training and inference
During training, the model sees interpolated points between samples from the starting and target distributions and learns a velocity that moves those points in the target direction. The original rectified-flow work also describes repeated rectification, or reflow, to make learned trajectories straighter. A specific production model may use related flow-matching choices without publishing every pairing, loss, or reflow detail.
During inference, the model predicts a direction at a sequence of time values. A numerical scheduler or solver updates the latent along that field. Fewer steps can reduce latency, but the architecture, training objective, distillation, solver, and release-specific guidance behavior all affect how few steps remain useful.
Where it appears
The Stable Diffusion 3 paper describes a rectified-flow formulation paired with its MMDiT architecture. Stable Diffusion 3 and 3.5 therefore differ from the U-Net diffusion path used by SDXL in more than backbone size. Black Forest Labs describes FLUX.1 dev as a rectified-flow transformer. These families share a broad formulation but are not checkpoint, adapter, or scheduler compatible by default.
The Diffusion Transformer Systems dossier places flow objectives beside transformer denoisers, text encoders, latent representations, and local inference requirements.
Why it matters in adult AI
Users encounter rectified-flow models through family names, low-step variants, LoRAs, and scheduler choices in tools such as ComfyUI and model repositories such as Hugging Face. The label helps identify a model's generation path, but it does not predict adult-content support, prompt fidelity, memory use, license, or adapter compatibility.
Rectified flow also does not make output deterministic or factually reliable. Character identity, anatomy, and fine detail can still drift. The exact model card remains the source for documented inputs, limits, license, and intended use.
Common confusions
- Rectified flow is not a model family. SD3 and FLUX use related ideas but have different weights and components.
- It is not a sampler name. The learned velocity field and the numerical method that integrates it are separate parts of inference.
- It does not guarantee one-step generation. Straight paths motivate efficient integration, but release-specific training and distillation determine practical step counts.
Technical context
These dossiers show how rectified-flow training is paired with different transformer backbones, encoders, schedulers, and compatibility ecosystems.