Research map / Image generation / Fast diffusion sampling

Diffusion Models without Classifier-free Guidance

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Guidance for conditional diffusion · 0 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 Model-guidance (MG), a training objective that bakes the CFG posterior tilt into diffusion/flow training so sampling needs only one forward pass and no CFG, reaching FID 1.34 on ImageNet 256 with SiT-XL/2.

Problem

Classifier-free guidance (CFG) is applied only at inference, requires training a separate unconditional model and doubles the number of network evaluations per sampling step, and the guided sampling distribution is never explicitly learned during training.

Approach

Model-guidance trains the conditional model to regress a target that includes the posterior-tilt term w*(eps(x_t,c) - eps(x_t,empty)) (stop-gradient, EMA model), so the learned score matches the CFG-tilted distribution without a second forward pass at inference; it also adds a scale-aware input w, optional removal of the empty class via the law of total probability, and an adaptive w schedule.

Key results

• SiT-XL/2+MG reaches FID 1.34 on ImageNet 256 without CFG, versus 8.61 for SiT-XL/2 without CFG and 2.06 with CFG.
• DiT-XL/2+MG improves FID from 9.62 to 2.03 without CFG on ImageNet 256, compared with 2.27 for DiT-XL/2 with CFG.
• On ImageNet 512, SiT-XL/2+MG reaches FID 2.24 vs 2.62 for SiT-XL/2 baseline.

Surprising findings

• Removing the empty (unconditional) class entirely still performs on par with vanilla CFG when the posterior is estimated via the law of total probability.
• MG-trained DiT-XL/2 (FID 3.37) and SiT-XL/2 (FID 2.89) outperform their CFG-guided counterparts in the scaling table.

Limitations

• The guidance scale w is sensitive and needs tuning or an adaptive schedule.
• Removing the empty class degrades posterior estimation (FID 9.66 vs 7.24 on DiT-B/2 in the ablation).
• Only class-conditional ImageNet experiments are reported in the provided text.

Experimental setup

Datasets: ImageNet

Benchmarks: ImageNet 256x256 ImageNet 512x512

Metrics: FID-50K sFID Inception Score Precision Recall throughput (img/s)

Baselines: DiT-XL/2 SiT-XL/2 ADM VDM++ LDM U-ViT MDTv2 REPA LightningDiT VAR RAR MAR

Builds on: classifier-free guidance DiT SiT classifier guidance

Closest papers by meaning

• Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.93)
• Feedback Guidance of Diffusion Models (NeurIPS 2025 · similarity 0.92)
• Rectified Diffusion Guidance for Conditional Generation (CVPR 2025 · similarity 0.91)
• No Training, No Problem: Rethinking Classifier-Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.91)
• CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.91)
• Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.90)