Generative Vision Atlas

Research line · ascendant

Guidance and sampling

Change what happens at inference time — guidance formulation, stochasticity, step schedule — without retraining the model.

What defines membership

Substantial quality is available at sampling time, and the guidance machinery inherited from diffusion is mismatched to flow models.

How the line developed

Read top to bottom: what came before, the idea itself, the evidence for it, what improved, and where it breaks.

The idea

Classifier-Free Guidance · 2022-07landmark

Jointly trains conditional and unconditional prediction in one network so sampling can extrapolate between them. Everything else on this page is a refinement of this.

Improvement

Guidance interval · 2024-04core

Guidance is harmful early and pointless late; applying it only in a middle interval improves both sample and distribution quality.

CFG++ · 2024-06core

Diagnoses oversaturation as guidance pushing samples off the data manifold, and reframes the correction as an inverse problem.

Autoguidance · 2024-06landmark

Separates quality from diversity by guiding with a weaker version of the same model rather than an unconditional one.

Adaptive Projected Guidance · 2024-10core

Splits the guidance update into parallel and orthogonal parts and finds only the orthogonal part carries prompt adherence.

GLASS Flows · 2025-09strong-followup

Simulates a flow inside a flow to get stochastic-sampler quality at deterministic-sampler cost, with no retraining.

Rectified-CFG++ · 2025-10emerging

Carries the CFG++ correction to rectified-flow models, where the diffusion-era formulation is mismatched to straight-path transport.

What it gets right

  • No training cost, so improvements apply to already-released models
  • Composes with every other line

Where it is weak

  • Bounded by what the trained model already represents
  • Guidance interacts with representation choice in ways not yet well characterized

Also in this line

Self-Sample Guidance