Research map / Image generation / Text-to-image diffusion
Consistent Story Generation: Unlocking the Potential of Zigzag Sampling
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 Asymmetric Zigzag Sampling (AZS), a training-free sampling method for subject-consistent visual story generation that injects cached subject visual tokens only in the zig step and uses a null prompt in the zag step, applied to SDXL and FLUX.
Problem
Maintaining subject consistency across multiple generated images for visual storytelling while preserving faithful text-prompt following, without per-subject fine-tuning or story-specific training.
Approach
Extends zigzag sampling into an asymmetric scheme: visual token injection (Zig Visual Sharing) only in the zig step, a null prompt in the zag step (Asymmetric Prompt Zigzag Inference), and full text guidance in the generation step, combined with one-prompt-one-story prompt fusion and IPCA identity prompts.
Key results
- On SDXL, the method reaches CLIP-T 0.8946, CLIP-I 0.923 and DreamSim 0.1798, beating the training-free 1Prompt1Story (0.8942/0.9117/0.1993).
- On FLUX, it improves over 1Prompt1Story on all three metrics (CLIP-T 0.8949 vs 0.8716, CLIP-I 0.9216 vs 0.9118, DreamSim 0.1843 vs 0.1957).
- In a 20-participant user study, the method received 35.02% of preferences versus 24.50% for 1Prompt1Story.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use asymmetric visual injection rather than symmetric visual injection (ablation) | Zigzag Sampling · Visual story generation (SDXL) | DreamSim: 0.2251 -> 0.1798 | Helped |
| Use asymmetric visual injection rather than symmetric visual injection (ablation) | Zigzag Sampling · Visual story generation (SDXL) | CLIP-T: 0.9098 -> 0.923 | Helped |
| Use asymmetric visual injection rather than symmetric visual injection (ablation) | Zigzag Sampling · Visual story generation (SDXL) | CLIP-I: 0.8841 -> 0.8946 | Helped |
| Set the visual token top-k ratio to 0.4 instead of 0.2 (ablation) | Zig Visual Sharing · Visual story generation (SDXL) | CLIP-T: 0.8946 -> 0.8931 | Hurt |
| Set the visual token top-k ratio to 0.4 instead of 0.2 (ablation) | Zig Visual Sharing · Visual story generation (SDXL) | CLIP-I: 0.923 -> 0.9245 | Helped |
| Set the visual token top-k ratio to 0.6 instead of 0.2 (ablation) | Zig Visual Sharing · Visual story generation (SDXL) | CLIP-T: 0.8946 -> 0.8921 | Hurt |
| Set the visual token top-k ratio to 0.6 instead of 0.2 (ablation) | Zig Visual Sharing · Visual story generation (SDXL) | CLIP-I: 0.923 -> 0.9258 | Helped |
| Set the visual token top-k ratio to 0.8 instead of 0.2 (ablation) | Zig Visual Sharing · Visual story generation (SDXL) | CLIP-T: 0.8946 -> 0.8924 | Hurt |
| Set the visual token top-k ratio to 0.8 instead of 0.2 (ablation) | Zig Visual Sharing · Visual story generation (SDXL) | CLIP-I: 0.923 -> 0.9287 | Helped |
| Apply Asymmetric Zigzag Sampling instead of 1Prompt1Story | SDXL visual storytelling · Visual story generation (SDXL) | CLIP-T: 0.8942 -> 0.8946 | Helped |
| Apply Asymmetric Zigzag Sampling instead of 1Prompt1Story | SDXL visual storytelling · Visual story generation (SDXL) | CLIP-I: 0.9117 -> 0.923 | Helped |
| Apply Asymmetric Zigzag Sampling instead of 1Prompt1Story | FLUX visual storytelling · Visual story generation (FLUX) | CLIP-T: 0.8716 -> 0.8949 | Helped |
Surprising findings
- Fully symmetric visual injection across all three sub-steps performs worst, attributed to over-conditioning and visual artifacts.
- IP-Adapter achieves the highest CLIP-I and DreamSim scores but much lower CLIP-T, indicating overemphasis on the subject and repetitive layouts.
Limitations
- Zigzag sampling adds three sub-steps per step, increasing inference time roughly 3-4x.
- Integrates more seamlessly with FLUX; occasional failures or reduced performance on SDXL.
- Evaluation relies on automatic metrics that depend on pretrained encoders, plus a relatively small user study of 30 prompts.
Experimental setup
Datasets: ConsiStory+
Benchmarks: ConsiStory+
Metrics: CLIP-T CLIP-I DreamSim human preference
Baselines: The Chosen One PhotoMaker IP-Adapter ConsiStory StoryDiffusion 1Prompt1Story
Builds on: zigzag sampling 1Prompt1Story StoryDiffusion ConsiStory IPCA
Closest papers by meaning
- Storybooth: Training-Free Multi-Subject Consistency for Improved Visual Storytelling (ICLR 2025 · similarity 0.89)
- SceneDecorator: Towards Scene-Oriented Story Generation with Scene Planning and Scene Consistency (NeurIPS 2025 · similarity 0.88)
- One-Prompt-One-Story: Free-Lunch Consistent Text-to-Image Generation Using a Single Prompt (ICLR 2025 · similarity 0.88)
- Boosting Consistency in Story Visualization with Rich-Contextual Conditional Diffusion Models (AAAI 2025 · similarity 0.88)
- VinaBench: Benchmark for Faithful and Consistent Visual Narratives (CVPR 2025 · similarity 0.88)
- StoryGPT-V: Large Language Models as Consistent Story Visualizers (CVPR 2025 · 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.