Research map / Image generation / Fast diffusion sampling

TCFG: Tangential Damping Classifier-free Guidance

CVPR 2025 · Image generation › Fast diffusion sampling · approach: Guidance for conditional diffusion · 2 citations (OpenAlex)

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 TCFG, a training-free classifier-free guidance variant that uses SVD over conditional and unconditional scores to drop the tangential component of the unconditional score, improving FID on SD1.5, SDXL, SD3 and DiT.

Problem

Classifier-free guidance combines conditional and unconditional scores, but the unconditional score's tangential components can misalign with the conditional score, pushing samples off the data manifold and causing artifacts and overexposure.

Approach

Concatenate unconditional and conditional scores into a matrix, take its SVD, and keep only the top singular-vector projection of the unconditional score (dropping low-singular-value tangential components) before applying the usual CFG update, at no extra inference cost.

Key results

• On MS-COCO 30k zero-shot, FID for SD3 drops from 16.66 to 13.74 with TCFG at unchanged CLIPScore (0.32); SDXL 13.36 to 12.65.
• On ImageNet 50k with DiT, FID improves from 32.67 to 29.5 and sFID from 17.92 to 13.27, with recall rising from 0.13 to 0.19.
• Combined with CFG++ on SDXL, FID falls from 13.97 to 13.44 and with SAG from 13.53 to 11.48.

Measured effects reported in the paper

Surprising findings

• The largest FID gains occur for the more capable SD3 model, which the authors speculate has a clearer manifold.
• Using only the modified unconditional score (w=0) on SDXL produces images that reveal the null condition removing unwanted structures, interpreted as reduced misalignment.

Limitations

• Occasionally fails to repair severely wrong regions in baseline samples, and structure can break down when original images have large abnormalities.
• Theoretical justification of the intermediate manifold is absent; only empirical observations are provided.
• Applicability to diffusion distillation with CFG-scale input is not yet explored.

Experimental setup

Datasets: MS-COCO 2014 validation (30k images) ImageNet 50k (evaluation) two moons (toy)

Benchmarks: MS-COCO 2014 30k ImageNet 50k

Metrics: FID CLIPScore sFID Precision Recall Inception Score

Baselines: classifier-free guidance SAG PAG CFG++

Builds on: classifier-free guidance CFG++ score-based manifold estimation (Stanczuk et al.) Stable Diffusion DiT

Closest papers by meaning

• Rectified Diffusion Guidance for Conditional Generation (CVPR 2025 · similarity 0.92)
• CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.92)
• Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.91)
• Manifold Preserving Guided Diffusion (ICLR 2024 · similarity 0.89)
• Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.89)
• Rethinking the Spatial Inconsistency in Classifier-Free Diffusion Guidance (CVPR 2024 · similarity 0.89)