FLUX-Reason-6M & PRISM-Bench: A Million-Scale Text-to-Image Reasoning Dataset and Comprehensive Benchmark

Rongyao Fang, Aldrich Yu, Chengqi Duan, Linjiang Huang, Shuai Bai, Yuxuan Cai, Kun Wang, Si Liu, Xihui Liu, Hongsheng Li

Introduction

The rapid evolution of open-source text-to-image (T2I) models has significantly narrowed the performance gap with proprietary systems. However, a persistent "reasoning gap" remains. While current models like Stable Diffusion or FLUX can generate visually stunning images, they often falter when faced with complex, multi-layered prompts requiring nuanced spatial logic, emotional depth, or precise text rendering. This paper introduces FLUX-Reason-6M, a massive dataset of 6 million high-quality images designed to bridge this gap, alongside PRISM-Bench, a comprehensive evaluation framework for measuring T2I reasoning capabilities.

Addressing the Reasoning Gap

The authors identify two primary obstacles to advanced T2I reasoning in existing open-source ecosystems:

  1. Data Quality: Most training data consists of "flat" web-crawled image-text pairs that describe what is in an image but fail to explain the underlying compositional logic or artistic intent.
  2. Evaluation Limitations: Standard metrics like CLIP scores or simple object detection are insufficient for capturing high-level reasoning, such as imaginative synthesis or complex emotional states.

FLUX-Reason-6M and Generation Chain-of-Thought (GCoT)

To address these issues, the researchers developed a massive data curation pipeline (costing approximately 15,000 A100 GPU days) to create a "reasoning-oriented" dataset. The cornerstone of this dataset is the Generation Chain-of-Thought (GCoT). Unlike traditional captions, GCoT provides a detailed step-by-step breakdown of the image's semantic and compositional logic, deconstructing how elements should be arranged.

Examples of GCoT and multi-label reasoning tags.

The dataset categorizes images into six key reasoning tracks, often using a multi-label design to force models to fuse different types of logic:

  • Imagination: Surreal or fantastical concepts.
  • Entity: High-fidelity, real-world objects.
  • Text Rendering: Legible English and Chinese text integration.
  • Style: Specific artistic movements and photographic techniques.
  • Affection: Abstract emotional concepts and moods.
  • Composition: Precise spatial arrangement and object interaction.

PRISM-Bench: A New Standard for Evaluation

Alongside the dataset, the authors present PRISM-Bench, a benchmark consisting of 700 carefully curated prompts (100 for each of the six characteristics plus a "Long Text" track). Evaluation is conducted via a "VLM-as-a-judge" approach, utilizing state-of-the-art vision-language models like GPT-4.1 and Qwen2.5-VL-72B to provide human-aligned scores for prompt adherence and aesthetic quality.

Key Results and Analysis

The benchmark reveals significant insights into the current state of T2I models. Closed-source models like GPT-Image-1 and Gemini 2.5-Flash-Image currently lead the field, particularly in complex spatial composition and overall reasoning.

Model performance across PRISM-Bench reasoning tracks.

Key findings from the evaluation include:

  • The Text Rendering Barrier: Across almost all models, text rendering remains the most difficult task, consistently receiving the lowest overall scores. This highlights a critical area for future architectural improvement.
  • Compositional Strength: GPT-Image-1 leads significantly in the Composition track (92.8), showing a superior ability to handle complex spatial instructions compared to its peers.
  • Open-Source Progress: Models like Qwen-Image and SEEDream 3.0 are becoming highly competitive, particularly in Chinese language tasks and complex composition, nearly matching closed-source performance in the PRISM-Bench-ZH track.
  • The Long Text Challenge: The "Long Text" track proved the most difficult for all systems, highlighting that even state-of-the-art models struggle to follow dense, multi-sentence instructions perfectly.

Conclusion

The release of FLUX-Reason-6M and PRISM-Bench represents a significant step toward democratizing high-quality T2I training data. While closed-source models maintain an edge in complex reasoning and spatial composition, the detailed "recipe" provided by GCoT offers a roadmap for open-source models to close the gap. Future research should focus on improving text rendering and handling increasingly complex, long-form instructions to further enhance model utility in professional design and creative workflows.

Summary updated

BibTeX
@inproceedings{fang2026flux,
  author = {Rongyao Fang and Aldrich Yu and Chengqi Duan and Linjiang Huang and Shuai Bai and Yuxuan Cai and Kun Wang and Si Liu and Xihui Liu and Hongsheng Li},
  title = {{FLUX-Reason-6M \& PRISM-Bench: A Million-Scale Text-to-Image Reasoning Dataset and Comprehensive Benchmark}},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
  url = {https://openreview.net/forum?id=cPzgZnpVbN},
}
Figure