Research map / Image generation / Fast diffusion sampling
CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CFG++, a one-line modification of classifier-free guidance that reuses the unconditional noise in the renoising step and restricts guidance to an interpolation scale in [0,1], improving T2I quality, DDIM inversion/editing, and text-conditioned inverse problems.
Problem
Classifier-free guidance (CFG) with high guidance scales causes off-manifold artifacts, mode collapse, and breaks DDIM inversion and editing, and it is hard to combine with diffusion inverse solvers.
Approach
Reformulate text guidance as an inverse problem with a text-conditioned score-distillation loss, solved via decomposed diffusion sampling; this yields CFG++, which uses the conditional denoised estimate with an interpolation factor lambda in [0,1] while keeping the renoising noise term unconditional.
Key results
- On COCO 10k with SD v1.5 at 50 NFE DDIM, CFG++ improves FID at every matched scale (e.g., 12.75 vs 13.84 at lambda=0.2/omega=2.0; 20.88 vs 21.23 at lambda=1.0/omega=12.5).
- On SDXL-Lightning at 6 NFE, CFG++ raises ImageReward from 0.777 to 0.968 and CLIP from 0.320 to 0.325.
- In PSLD inverse problems on FFHQ, PSLD+CFG++ reduces SR (x8) FID from 46.24 (PSLD) and 41.24 (PSLD+CFG) to 36.58, and motion deblur FID from 97.51 to 65.67.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace CFG renoising with unconditional renoising and use interpolating guidance | SD v1.5 DDIM sampling · COCO 10k, ω=2.0/λ=0.2 | FID: 13.84 -> 12.75 | Helped |
| Replace CFG renoising with unconditional renoising and use interpolating guidance | SD v1.5 DDIM sampling · COCO 10k, ω=5.0/λ=0.4 | FID: 15.08 -> 14.95 | Helped |
| Replace CFG renoising with unconditional renoising and use interpolating guidance | SD v1.5 DDIM sampling · COCO 10k, ω=7.5/λ=0.6 | FID: 17.71 -> 17.47 | Helped |
| Replace CFG renoising with unconditional renoising and use interpolating guidance | SD v1.5 DDIM sampling · COCO 10k, ω=9.0/λ=0.8 | FID: 20.01 -> 19.34 | Helped |
| Replace CFG renoising with unconditional renoising and use interpolating guidance | SD v1.5 DDIM sampling · COCO 10k, ω=12.5/λ=1.0 | FID: 21.23 -> 20.88 | Helped |
| Replace CFG renoising with unconditional renoising and use interpolating guidance | SD v1.5 DDIM sampling · COCO 10k, matched scales | CLIP: 0.298 -> 0.303; 0.310 -> 0.310; 0.312 -> 0.312; 0.312 -> 0.313; 0.313 -> 0.313 | Mixed |
| Add CFG++ guidance | PSLD · FFHQ, SR (x8) | FID: 41.24 -> 36.58 | Helped |
| Add CFG++ guidance | PSLD · FFHQ, Deblur (motion) | FID: 91.90 -> 65.67 | Helped |
| Add CFG++ guidance | PSLD · FFHQ, Deblur (gauss) | FID: 41.52 -> 39.85 | Helped |
| Add CFG++ guidance | PSLD · FFHQ, Inpaint | FID: 9.36 -> 9.78 | Hurt |
| Add CFG++ guidance | PSLD · FFHQ, SR (x8) | LPIPS: 0.394 -> 0.385 | Helped |
| Add CFG++ guidance | PSLD · FFHQ, SR (x8) | PSNR: 24.91 -> 24.87 | Hurt |
Surprising findings
- CFG++ with lambda up to 1.0 matches CFG at omega around 12.5, so guidance scales in [0,1] suffice, contrary to the common moderately high scale practice.
- The CFG++ update corresponds to a specific time-varying CFG schedule, which is shown to be hard to derive heuristically.
Limitations
- Applications explored are limited, as the authors acknowledge.
- Inverse-problem gains are mixed; CFG++ is not uniformly best across all tasks and metrics (e.g., Gaussian deblur FID is worse than PSLD+CFG on LPIPS/PSNR in some cases).
- Matched CFG and CFG++ scales were chosen via LPIPS proximity to the same seed, which is a heuristic matching procedure.
Experimental setup
Datasets: COCO COCO-2014 FFHQ
Benchmarks: COCO 10k COCO-2014 validation FFHQ 512x512
Metrics: FID CLIP score ImageReward PSNR RMSE LPIPS
Baselines: CFG PSLD PSLD + CFG PCG
Builds on: classifier-free guidance DDIM decomposed diffusion sampling diffusion posterior sampling PSLD DPM-solver++
Closest papers by meaning
- Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.93)
- Rectified Diffusion Guidance for Conditional Generation (CVPR 2025 · similarity 0.92)
- Rethinking the Spatial Inconsistency in Classifier-Free Diffusion Guidance (CVPR 2024 · similarity 0.92)
- Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.92)
- TCFG: Tangential Damping Classifier-free Guidance (CVPR 2025 · similarity 0.92)
- Diffusion Models without Classifier-free Guidance (NeurIPS 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.