Research map / Image generation / Fast diffusion sampling
Improving Guidance-Free Visual Generation via Self-Contrastive Alignment for Likelihood Estimation
Jiwan Hur, DongJae Lee, Dongyeun Lee, Gyojin Han, Junmo Kim
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 self-contrastive guidance (SCG) for guidance-free visual generation, contrasting the model's own predictions to approximate CFG at single-inference cost.
Problem
Classifier-free guidance improves visual generation quality but doubles sampling cost; existing guidance-free methods that approximate CFG with one inference are bounded by CFG's performance and inherit its limitations.
Approach
Self-contrastive guidance contrasts the model's prediction on its own inputs (the abstract is truncated, so the exact contrast is not stated) to get guidance without a separate unconditional pass, aiming to exceed the CFG-approximation ceiling.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add self-contrastive guidance (SCALE) on top of GFT for diffusion | DiT-XL/2 guidance-free · ImageNet 256x256 | FID-50k: 1.99 -> 1.87 | Improved |
| Add SCALE on top of GFT for autoregressive model | LlamaGen-L guidance-free · ImageNet 256x256 | FID-50k: 2.88 -> 2.58 | Improved |
| Use SCALE instead of CCA on LlamaGen-L | LlamaGen-L · ImageNet 256x256 | IS: 288.2 -> 253.0 | Hurt |
| Replace two-pass CFG with guidance-free SCALE | DiT-XL/2 · ImageNet 256x256 | FID: 2.17 -> 1.87 | Improved |
| Lower alpha from 0.8 to 0.25 at sampling (ablation) | SCALE · SD 1.5 text-to-image | FID-30k: 12.83 -> 16.52 | Hurt |
| Lower alpha from 0.8 to 0.25 at sampling (ablation) | SCALE · SD 1.5 text-to-image | HPSv2.1: 0.253 -> 0.261 | Improved |
| Apply DPO to guidance-free base SD 1.5 | SD 1.5 guidance-free · SD 1.5 text-to-image | preference metrics | No clear effect |
| Add DPO on top of SCALE | SCALE · SD 1.5 text-to-image | preference metrics | Improved |
| Train SCALE from scratch instead of GFT | DiT-B/2 · ImageNet 256x256 (from scratch) | FID: 9.04 -> 8.75 | Improved |
| Use patch-wise mix instead of linear interpolation (ablation) | SCALE · DiT-B/2 ImageNet | FID: 8.67 -> 8.60 | Improved |
| Lower PatchMix ratio lambda from 1.0 to 0.2 (ablation) | SCALE · LlamaGen-L ImageNet | IS: 245.9 -> 253.0 | Improved |
| Lower PatchMix ratio lambda from 1.0 to 0.2 (ablation) | SCALE · LlamaGen-L ImageNet | FID: 2.59 -> 2.58 | No clear effect |
Limitations
- Only the abstract fragment is available; experimental details are unknown
Experimental setup
Baselines: classifier-free guidance guidance-free approaches that approximate CFG
Builds on: classifier-free guidance
Closest papers by meaning
- Toward Guidance-Free AR Visual Generation via Condition Contrastive Alignment (ICLR 2025 · similarity 0.92)
- Visual Generation Without Guidance (ICML 2025 · similarity 0.91)
- Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.89)
- Increasing the Utility of Synthetic Images through Chamfer Guidance (NeurIPS 2025 · similarity 0.87)
- Feedback Guidance of Diffusion Models (NeurIPS 2025 · similarity 0.87)
- Smoothed Energy Guidance: Guiding Diffusion Models with Reduced Energy Curvature of Attention (NeurIPS 2024 · similarity 0.87)
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.
