FLUX-Reason-6M: T2I Reasoning Dataset

• FLUX-Reason-6M is a large-scale, bilingual text-to-image reasoning dataset featuring 6M images and 20M captions annotated across six reasoning dimensions.
• It employs a sophisticated computational pipeline, including 15,000 A100 GPU days and multi-label VLM filtering, to ensure high-quality semantic and compositional alignment.
• The accompanying PRISM-Bench framework rigorously evaluates T2I models on multiple tracks, highlighting challenges in text rendering and long-form reasoning.

FLUX-Reason-6M is a million-scale text-to-image reasoning dataset paired with a comprehensive benchmarking suite intended to accelerate open-source text-to-image synthesis research by providing data and evaluation tools previously available only to large industry labs. The corpus consists of 6 million FLUX-generated images and 20 million bilingual (English/Chinese) captions, annotated across six core reasoning characteristics and featuring explicit Generation Chain-of-Thought (GCoT) descriptions to guide models in complex scene synthesis and semantic alignment. Accompanied by PRISM-Bench, a multi-track human-aligned evaluation framework, FLUX-Reason-6M constitutes a cornerstone resource for rigorous performance tracking and targeted model improvements in reasoning-oriented T2I generation (Fang et al., 11 Sep 2025).

1. Dataset Structure and Reasoning Characteristics

FLUX-Reason-6M is centered around 6 million images synthesized via the FLUX.1-dev engine, each accompanied by at least three separate captions. Caption types include legacy captions (sourced from Laion-Aesthetics), dense category-specific descriptions, and Generation Chain-of-Thought (GCoT) annotations. Captions are provided in both English and Chinese, forming a 20 million sample bilingual corpus.

Six reasoning dimensions organize the dataset:

Reasoning Characteristic	Description	Example
Imagination	Surreal/abstract	"A floating castle above a cloud shaped like a tiger"
Entity	Detailed real-world	"The Eiffel Tower at night"
Text Rendering	Legible text/labels	"A billboard displaying 'Open Science Now'"
Style	Artistic/aesthetic	"Portrait in Cubist style"
Affection	Emotion/mood cues	"A tranquil sunrise evoking peace"
Composition	Spatial relations	"Two cars, one behind the other, facing left"

Images may have multi-label assignments. For instance, "The Eiffel Tower in Van Gogh’s style" is classified under both Entity and Style.

2. Generation Chain-of-Thought (GCoT) Annotation

The Generation Chain-of-Thought provides structured reasoning annotations that break down image synthesis steps. Unlike traditional captioning, GCoT details the semantic and compositional logic underlying an image's construction, enabling models to learn fine-grained correspondences and multi-step synthesis procedures. Each GCoT is formed by fusing all dense category captions with the complete image context via a vision-LLM. Examples include stepwise allocations of entities, background, layout, style, and affinity between visual elements.

This explicit chain-of-thought supervision addresses shortcomings of prior datasets that lack detailed reasoning context for complex prompt-image alignment, particularly in compositional and multi-attribute scenes.

3. Data Curation and Computational Pipeline

The curation process required 15,000 A100 GPU days, indicative of the computational intensity involved. The curation pipeline operates as follows:

• An initial pool of over 8 million images is synthesized by FLUX.1-dev.
• Automated filtering with VLMs such as Qwen-VL removes artifacts, verifies typographic integrity, and screens for legibility and composition.
• Each image is scored 1–10 on each of the six reasoning axes by a multi-label VLM system.
• Dense captions and GCoT annotations are generated and merged; bilingual translations unify the corpus for multilingual training.
• Text rendering undergoes a specialized workflow to ensure precision (correct spelling and context).

The data pipeline is depicted in Figure \ref{fig:data_pipe} (see original).

4. PRISM-Bench Benchmarking Framework

PRISM-Bench is a seven-track evaluation suite designed for nuanced model assessment aligned with human standards. Six tracks correspond to dataset characteristics; the seventh, Long Text, leverages GCoT prompts to test extended compositional reasoning. For each track, 100 prompts are selected—50 via k-means semantic clustering, 50 curated for category-specific difficulty.

Evaluation involves advanced vision-LLMs (GPT-4.1, Qwen2.5-VL-72B), scoring each image for prompt-image alignment (scale 1–10) and image aesthetics (scale 1–10). Scores are averaged and mapped to a 0–100 scale. Figure \ref{fig:bench_pipe} in the source shows the protocol in detail.

The Long Text track is notable for exposing model limitations in following complex, multi-paragraph instructions, a key motivation for GCoT supervision.

5. Cross-Model Performance and Identified Deficiencies

A total of 19 models are evaluated on PRISM-Bench, spanning closed-source (e.g., Gemini2.5-Flash-Image, GPT-Image-1) and open-source (Qwen-Image, FLUX.1-Krea-dev) systems. Assessment is conducted on English and Chinese versions (PRISM-Bench-ZH), with numeric scores tabulated in the original (Tables \ref{tab:gpt_en} and \ref{tab:qwen_en}).

Findings include:

• Closed-source models generally outperform open-source alternatives in Imagination, Entity, and Composition tracks.
• Text Rendering remains the weakest area, with lower accuracy and concordance in both prompt alignment and legibility, even for state-of-the-art systems.
• Long Text presents critical challenges, with most evaluated models struggling to synthesize coherent, multi-step instructions into visually consistent outputs.

Floored performance in text rendering and multi-layer reasoning suggests algorithmic deficiencies in current T2I systems’ capacity for compositional control and extended reasoning.

6. Impact, Accessibility, and Future Applications

FLUX-Reason-6M, PRISM-Bench, and evaluation code are released for community use. This resource lowers technical and financial entry barriers, making large-scale, reasoning-centric T2I training and evaluation accessible beyond large industrial labs.

Distinctive attributes include:

• Bilingual annotation corpus supports global research and cross-lingual generalization.
• Explicit multi-label and chain-of-thought reasoning annotations encourage development of next-generation T2I models with advanced compositional and semantic understanding.
• Fine-grained evaluation diagnostics guide model development for targeted improvements (e.g., text rendering, multi-attribute synthesis).

A plausible implication is acceleration in open-source T2I research, with benchmarks serving as clear standard for progress in multimodal reasoning.

7. Summary and Significance

FLUX-Reason-6M provides large-scale, reasoning-specific text-to-image training data, with robust annotation and benchmarking support through PRISM-Bench. It introduces chain-of-thought supervision across six reasoning axes, supports bilingual capabilities, and defines a high bar for evaluation via human-aligned VLMs. Empirical results highlight text rendering and long-form reasoning as chief deficiencies in current models, underlining critical research avenues. Public release of all resources is designed to catalyze more sophisticated, human-aligned T2I model development in academic and open-source communities.