Research map / Vision-language and multimodal / Human and embodied modeling
Visual Programming for Step-by-Step Text-to-Image Generation and Evaluation
Jaemin Cho, Abhay Zala, Mohit Bansal
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train layout LM on Flickr30K+COCO+PaintSkills instead of only Flickr30K (ablation) | VPGEN · skill-based prompts | average skill score: 43.9 -> 51.0 | Improved |
| add COCO and PaintSkills to layout training (ablation) | VPGEN · open-ended prompts | score: 71.0 -> 68.3 | Hurt |
| add COCO and PaintSkills to layout training (ablation) | VPGEN · COCO 30K | FID: 15.9 -> 20.1 | Hurt |
| step-by-step layout generation instead of direct generation | VPGEN vs Stable Diffusion v1.4 · Count skill | VPEval accuracy: 47.4 -> 72.2 | Improved |
| step-by-step layout generation instead of direct generation | VPGEN vs Stable Diffusion v1.4 · Spatial skill | VPEval accuracy: 22.9 -> 56.1 | Improved |
| step-by-step layout generation instead of direct generation | VPGEN vs Stable Diffusion v1.4 · Text Rendering skill | VPEval accuracy: 8.9 -> 3.7 | Hurt |
| train on synthetic PaintSkills only instead of real Flickr30K (ablation) | VPGEN · skill-based / open-ended | average score: 43.9 -> 46.0 (skill); 71.0 -> 65.6 (open-ended) | Mixed |
| use 1000 bins instead of 100 for box quantization (ablation) | VPGEN · skill-based prompts | average skill score: 51.0 -> 49.8 | No clear effect |
| use 1000 bins instead of 100 (ablation) | VPGEN · open-ended prompts | score: 68.3 -> 68.2 | No clear effect |
| replace finetuned Vicuna 13B with GPT-3.5-Turbo (36 in-context examples) (ablation) | VPGEN · skill-based prompts | average skill score: 51.0 -> 40.7 | Hurt |
| replace finetuned Vicuna 13B with GPT-3.5-Turbo (ablation) | VPGEN · open-ended prompts | score: 68.3 -> 65.5 | Hurt |
| use BLIP-2 VQA for object/spatial/scale modules in VPEval (ablation) | VPEval · skill-based human correlation | Spearman rho overall: 73.5 -> 79.0 | Improved |
Closest papers by meaning
- Rethinking The Training And Evaluation of Rich-Context Layout-to-Image Generation (NeurIPS 2024 · similarity 0.91)
- LLM Blueprint: Enabling Text-to-Image Generation with Complex and Detailed Prompts (ICLR 2024 · similarity 0.90)
- Elucidating the design space of language models for image generation (ICML 2025 · similarity 0.89)
- OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation (NeurIPS 2025 · similarity 0.88)
- Imagine Before You Draw: Visual Prompt Engineering for Image Generation (NeurIPS 2026 · similarity 0.88)
- VisCodex: Unified Multimodal Code Generation via Merging Vision and Coding Models (ICLR 2026 · 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.
