Pith. sign in

REVIEW 24 cited by

Classifier-Free Guidance is a Predictor-Corrector

Not yet reviewed by Pith; the record is open.

Re-run · record.json Download PDF Read on arXiv ↗

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2408.09000 v2 pith:DYT2I4DZ submitted 2024-08-16 cs.LG cs.AIcs.CV

Arwen Bradley , Preetum Nakkiran This is my paper · ORCID
classification cs.LGcs.AIcs.CV
keywords distributiongammaguidancepredictor-correctorclassifier-freeconditionalddimdiffusion
verification ladder T0 review T1 audit T2 compute T3 formal

Signed reviews

No signed human review yet.

0 comments
abstract

We investigate the theoretical foundations of classifier-free guidance (CFG). CFG is the dominant method of conditional sampling for text-to-image diffusion models, yet unlike other aspects of diffusion, it remains on shaky theoretical footing. In this paper, we disprove common misconceptions, by showing that CFG interacts differently with DDPM (Ho et al., 2020) and DDIM (Song et al., 2021), and neither sampler with CFG generates the gamma-powered distribution $p(x|c)^\gamma p(x)^{1-\gamma}$. Then, we clarify the behavior of CFG by showing that it is a kind of predictor-corrector method (Song et al., 2020) that alternates between denoising and sharpening, which we call predictor-corrector guidance (PCG). We prove that in the SDE limit, CFG is actually equivalent to combining a DDIM predictor for the conditional distribution together with a Langevin dynamics corrector for a gamma-powered distribution (with a carefully chosen gamma). Our work thus provides a lens to theoretically understand CFG by embedding it in a broader design space of principled sampling methods.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 24 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. What Exactly Does Guidance Do in Masked Discrete Diffusion Models

    stat.ML 2025-06 accept novelty 8.0 of 10

    With exact scores and no discretization error, CFG in 1D masked discrete diffusion samples exactly the tilted distribution; in 2D it does not, and the TV convergence rate is double-exponential in guidance strength.

  2. Analytic Distribution of Classifier-Free Guidance for Schedule Design

    cs.LG 2026-07 conditional novelty 7.0 of 10

    Deterministic CFG samples from p_t0 times an exponential path integral of the score discrepancy, and the resulting schedule DG-CFG reduces sampling steps at high guidance.

  3. Guidance Breaks the Fitted Operator: A Terminal-Fitted Repair for Classifier-Free Guidance

    cs.LG 2026-07 conditional novelty 7.0 of 10

    Replacing CFG's w(r-1) coefficient with r^(1+w)-r removes a sigma_min-divergent residual blow-up on a Gaussian calibration model and stabilizes high-guidance diffusion sampling at zero extra NFE.

  4. Conditional Diffusion Models with Classifier-Free Gibbs-like Guidance

    cs.LG 2025-05 conditional novelty 7.0 of 10

    CFG is missing a Rényi divergence repulsion term, and the new CFG IG sampler uses iterative noising and denoising to preserve diversity while improving conditional generation quality.

  5. Classifier-Free Guidance: From High-Dimensional Analysis to Generalized Guidance Forms

    cs.LG 2025-02 conditional novelty 7.0 of 10

    CFG's distortion of the target distribution vanishes as data dimension grows, and a power-law generalization improves fidelity and diversity in high-dimensional generative models.

  6. Beyond and Free from Diffusion: Invertible Guided Consistency Training

    cs.CV 2025-02 conditional novelty 7.0 of 10

    iGCT trains guided consistency models from scratch by mixing the original noise with a direction to a random target-class image, and reports better FID and precision than classifier-free guidance at high guidance on CIFAR-10.

  7. A Decomposable Probe for Few-Step Diffusion Models: Prompt, Latent, and Score Selectivity across Backbone Families and Distillation Paradigms

    cs.CV 2026-07 conditional novelty 6.5 of 10

    A three-layer perturbation probe shows latent selectivity is a near-binary rectified-flow fingerprint that survives ADD distillation, while score selectivity tracks distillation objective across 23 T2I models.

  8. CODA: Algorithm-Hardware Co-design for Edge Video Diffusion via NMP-Enabled Compute-Cache Operator Disaggregation

    cs.AR 2026-07 conditional novelty 6.0 of 10

    Disaggregating cache operators from compute and overlapping them across the two classifier-free-guidance branches turns cross-timestep caching into up to 1.80x real end-to-end speedup on edge GPUs when the cache overf...

  9. Angle Domain Guidance: Latent Diffusion Requires Rotation Rather Than Extrapolation

    cs.LG 2025-05 conditional novelty 6.0 of 10

    ADG replaces linear extrapolation in classifier-free guidance with angular rotation, reducing color distortion at high guidance weights while preserving text alignment.

  10. Provable Efficiency of Guidance in Diffusion Models for General Data Distribution

    stat.ML 2025-05 conditional novelty 6.0 of 10

    For arbitrary data distributions, classifier-free guidance provably decreases the expected reciprocal classifier probability along the reverse diffusion process.

  11. Debiasing Guidance for Discrete Diffusion with Sequential Monte Carlo

    cs.LG 2025-02 conditional novelty 6.0 of 10

    An SMC importance-sampling algorithm debiases discrete diffusion guidance, asymptotically sampling from the target tempered distribution p0(x0)p(ζ|x0)^α.

  12. UniDB: A Unified Diffusion Bridge Framework via Stochastic Optimal Control

    cs.CV 2025-02 conditional novelty 6.0 of 10

    A stochastic optimal control formulation of diffusion bridges, where Doob's h-transform is the infinite-penalty limit and a finite penalty yields a tunable detail-preserving bridge.

  13. REG: Rectified Gradient Guidance for Conditional Diffusion Models

    cs.CV 2025-01 conditional novelty 6.0 of 10

    Guidance in diffusion models is best understood as approximating an intractable joint-scaling optimum, and the proposed REG correction improves this approximation in practice.

  14. Guiding a diffusion model using sliding windows

    cs.CV 2024-11 conditional novelty 6.0 of 10

    Masked sliding window guidance improves diffusion sample quality by guiding the model with its own crop-based predictions, without training.

  15. On the Separability of Information in Diffusion Models

    cs.LG 2025-09 conditional novelty 5.0 of 10

    Diffusion models devote most of their information budget to class-agnostic texture, and the small class-relevant slice is what classifier-free guidance amplifies.

  16. Solving Inverse Problems via Diffusion-Based Priors: An Approximation-Free Ensemble Sampling Approach

    cs.LG 2025-06 conditional novelty 5.0 of 10

    A weighted-particle sampler evolves the posterior through the diffusion model's reverse dynamics, with theoretical error bounds and improved image reconstructions.

  17. Self-Reflective Reinforcement Learning for Diffusion-based Image Reasoning Generation

    cs.CV 2025-05 reject novelty 5.0 of 10

    SRRL trains diffusion models to reason about images by repeatedly denoising, evaluating, and re-noising their own samples across several reflective rounds.

  18. Towards Flow-Matching-based TTS without Classifier-Free Guidance

    eess.AS 2025-04 reject novelty 5.0 of 10

    Modifying the flow-matching training target lets F5-TTS synthesize speech without classifier-free guidance at inference, halving per-step cost and improving measured WER, SIM-O, and MOS.

  19. ContextMRI: Enhancing Compressed Sensing MRI through Metadata Conditioning

    cs.CV 2025-01 conditional novelty 5.0 of 10

    Using clinical metadata as text prompts in a diffusion prior yields 0.2 to 0.5 dB PSNR gains for compressed sensing MRI reconstruction, but the gains are inconsistent at some acceleration factors.

  20. A solvable generative model with a linear, one-step denoiser

    cs.LG 2024-11 reject novelty 5.0 of 10

    The paper derives a closed-form KL divergence for a one-step linear diffusion model on Gaussian data, reports a sample-size threshold at n=d, and gives a heuristic argument that more diffusion steps improve quality.

  21. Comparing SGLD and a fixed-noise Predictor-Corrector adaptation in canonical Joint Energy-Based Models on CIFAR-10

    cs.LG 2026-08 conditional novelty 4.0 of 10

    On canonical JEM for CIFAR-10, Predictor-Corrector and SGLD are practically indistinguishable across generation, OOD detection, and training, and every run crashes late in training via the outlier-buffer mechanism.

  22. Connections between reinforcement learning with feedback,test-time scaling, and diffusion guidance: An anthology

    stat.ML 2025-09 conditional novelty 4.0 of 10

    RLHF, RLIF, and soft best-of-N sampling reduce to the same exponential-tilting objective under parameter matching, and test-time scaling can asymptotically implement classifier-free diffusion guidance.

  23. Inference-Time Alignment Control for Diffusion Models with Reinforcement Learning Guidance

    cs.LG 2025-08 conditional novelty 4.0 of 10

    Blending a base diffusion model with its RL-finetuned version at sampling time lets users dial alignment strength, with the blend weight corresponding to the KL-regularization coefficient beta/w.

  24. DiffIER: Optimizing Diffusion Models with Iterative Error Reduction

    cs.CV 2025-08 reject novelty 4.0 of 10

    DiffIER claims that iteratively minimizing the distance between a diffusion model's predicted noise and a random Gaussian sample at each inference step improves generation quality.

Pith tools