Research map / Image generation / Fast diffusion sampling

Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization

CVPR 2025 · Image generation › Fast diffusion sampling · approach: Diffusion theory and design space · 4 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 an inference-time method that delays classifier-free guidance until a transition point where the diffusion trajectory exits an "attraction basin" (plus opposite guidance) to mitigate memorization in diffusion models.

Problem

Diffusion models verbatim-reproduce (memorize) training images, raising copyright and privacy concerns; existing mitigations require retraining or prompt/embedding perturbation that generalize poorly across memorization causes.

Approach

Memorization arises from an attraction basin in the denoising trajectory under CFG; zero guidance until a transition point (detected via a drop in the magnitude of conditional-minus-unconditional noise prediction, static or dynamic) steers samples out of the basin, and opposite guidance (negative CFG) pushes trajectories out sooner.

Key results

• On Scenario 1, OG+STP reduces 95pc SSCD similarity from 0.6504 (no mitigation) to 0.3811 while improving FID from 16.84 to 15.67 and CLIP score from 0.3027 to 0.3020.
• On Scenario 3, OG+DTP yields 95pc similarity 0.6915 vs 0.7977 unmitigated, with FID 140.05 vs 106.49 (worse) and CLIP 0.2844 vs 0.5513.
• Inference cost is about 1.26 s per image on one A100 for SDv2.1 standard CFG, versus 2.86 s for Wen et al. and 2.08 s for Ren et al.

Measured effects reported in the paper

Surprising findings

• Applying CFG only after the transition point yields non-memorized outputs even when applying CFG from earlier steps always produces the memorized image.
• Previous prompt-level mitigations (Ren et al., Somepalli et al.) fail to generalize across memorization scenarios.

Limitations

• Transition points do not exist for non-memorized samples, so memorization must be detected first (requires denoising twice or a magnitude threshold).
• Applying CFG too late degrades alignment and image quality; the transition point must be found per prompt or per model.
• Opposite guidance worsened FID in Scenario 3 (140.05 vs 106.49 unmitigated).

Experimental setup

Datasets: LAION-10k LAION-100k Imagenette

Benchmarks: LAION-10k fine-tuning memorization scenario Imagenette fine-tuning scenario duplicated-prompt fine-tuning scenario

Metrics: SSCD similarity (95th percentile) CLIP score FID

Baselines: Wen et al. [28] Ren et al. [20] Add rand numb [24] Add rand word [24]

Builds on: classifier-free guidance Wen et al. memorization detection Ren et al. cross-attention mitigation Somepalli et al. diffusion memorization

Closest papers by meaning

• Adjusting Initial Noise to Mitigate Memorization in Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.94)
• How Diffusion Models Memorize (NeurIPS 2026 · similarity 0.91)
• Towards Memorization-Free Diffusion Models (CVPR 2024 · similarity 0.90)
• On the Edge of Memorization in Diffusion Models (NeurIPS 2025 · similarity 0.90)
• Cyclic Denoising Reveals Ultrastable Memories in Diffusion Models (NeurIPS 2026 · similarity 0.89)
• Does Generation Require Memorization? Creative Diffusion Models using Ambient Diffusion (ICML 2025 · similarity 0.89)