Research map / Image generation / Text-to-image diffusion
Zigzag Diffusion Sampling: Diffusion Models Can Self-Improve via Self-Reflection
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 Z-Sampling, a training-free diffusion sampler that alternates denoising and inversion (diffusion self-reflection) along the sampling path to accumulate prompt-related semantic information via the guidance gap between denoising and inversion.
Problem
Text-to-image diffusion models often fail to maintain high image quality and prompt-image alignment for challenging prompts when strong classifier-free guidance is used.
Approach
Perform zigzag self-reflection: at each timestep denoise with strong guidance (gamma1) and invert with weak guidance (gamma2), replacing the latent with the inverted one so that the guidance gap injects semantic information step by step, rather than once end-to-end as in vanilla inversion.
Key results
- On Pick-a-Pic, SDXL with Z-Sampling improves HPS v2 from 29.89 to 31.28 and ImageReward from 58.65 to 79.22 over standard sampling.
- On GenEval with SDXL, overall score rises from 52.52% to 57.04%, with counting improving from 33.75% to 46.25%.
- DreamShaper-xl-v2-turbo with Z-Sampling reaches a 94% HPS v2 winning rate over standard sampling per the abstract; it also combines with Diffusion-DPO to improve results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace standard sampling with Z-Sampling | SDXL · GenEval | Overall: 52.52% -> 57.04% | Helped |
| Replace standard sampling with Z-Sampling | DreamShaper-xl-v2-turbo · Pick-a-Pic | HPS v2: 30.04 -> 32.38 | Helped |
| Replace standard sampling with Z-Sampling | Hunyuan-DiT · DrawBench | HPS v2: 30.72 -> 30.53 | Hurt |
| Combine Z-Sampling with AYS | DreamShaper-xl-v2-turbo · Pick-a-Pic | HPS v2: 32.78 -> 33.57 | Helped |
| Combine Z-Sampling with Diffusion-DPO | SDXL · GenEval | Overall: 59.18% -> 59.57% | Helped |
| Use the zigzag operation for more denoising steps (ablation) | SDXL with Z-Sampling · Pick-a-Pic | HPS v2 winning rate: 50% to 75%; 75% to 80% | Helped |
| Set the inversion and denoising guidance scales equal, eliminating the guidance gap (ablation) | Z-Sampling · Pick-a-Pic | HPS v2 winning rate | No clear effect |
| Set inversion guidance above denoising guidance, making the guidance gap negative (ablation) | Z-Sampling · Generation quality | generation quality | Hurt |
| Add Z-Sampling to standard sampling | DreamShaper-xl-turbo-v2 at guidance 3.5 · Pick-a-Pic | HPS v2: 30.04 -> 32.38 | Helped |
| Increase backtracking stepsize from k=1 to k=2 (ablation) | Z-Sampling on SDXL · Pick-a-Pic | HPS v2: 31.28 -> 31.11 | Hurt |
| Increase backtracking stepsize from k=1 to k=3 (ablation) | Z-Sampling on SDXL · Pick-a-Pic | HPS v2: 31.28 -> 30.75 | Hurt |
| Increase backtracking stepsize from k=1 to k=4 (ablation) | Z-Sampling on SDXL · Pick-a-Pic | HPS v2: 31.28 -> 30.59 | Hurt |
Surprising findings
- Z-Sampling with 36% less computation time can still surpass the best performance of standard sampling.
- Standard sampling degrades at high CFG scales (oversaturation/artifacts) while Z-Sampling still yields gains.
Limitations
- Relies on deterministic inversion (e.g., DDIM), limiting applicability to SDE-based samplers.
- Adds extra computation on zigzag steps.
- Only text-to-image is studied in depth; video results are limited.
Experimental setup
Datasets: Pick-a-Pic DrawBench GenEval PartiPrompts MS-COCO 2014 ChronoMagic-Bench-150
Benchmarks: Pick-a-Pic DrawBench GenEval PartiPrompts MS-COCO 2014 ChronoMagic-Bench-150
Metrics: HPS v2 AES PickScore ImageReward winning rate FID IS CLIP score UMT-FVD UMT-SCORE
Baselines: Standard sampling Resampling Semantic-aware CFG AYS Diffusion-DPO SEG CFG++ FreeInit
Builds on: DDIM classifier-free guidance Resampling Diffusion-DPO AYS
Closest papers by meaning
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.89)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.89)
- OneActor: Consistent Subject Generation via Cluster-Conditioned Guidance (NeurIPS 2024 · similarity 0.89)
- Diffusion Self-Distillation for Zero-Shot Customized Image Generation (CVPR 2025 · similarity 0.89)
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.88)
- Fast constrained sampling in pre-trained diffusion models (NeurIPS 2025 · similarity 0.88)
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.