FLUX-Reason-6M: T2I Reasoning Dataset

• FLUX-Reason-6M is a large-scale, open-source text-to-image dataset that integrates chain-of-thought annotations to enhance compositional and multilingual reasoning.
• It employs systematic multi-label annotations across six axes—Imagination, Entity, Text Rendering, Style, Affection, and Composition—to support detailed model evaluation.
• The accompanying PRISM-Bench protocol rigorously benchmarks T2I models, highlighting gaps in logical composition, text rendering, and long instruction following.

FLUX-Reason-6M is a million-scale, open-source text-to-image (T2I) reasoning dataset designed to advance research in multi-dimensional visual reasoning, generation chain-of-thought supervision, and robust T2I model benchmarking. Comprising 6 million FLUX-generated images paired with 20 million bilingual (English and Chinese) captions, the dataset explicitly targets core limitations in existing T2I corpora by offering both high image quality and chain-of-thought style annotations for complex prompt-to-image reasoning across six targeted dimensions: Imagination, Entity, Text Rendering, Style, Affection, and Composition. The accompanying PRISM-Bench benchmark further establishes a comprehensive, multidimensional standard for evaluating T2I model output in terms of both reasoning alignment and aesthetics, leveraging advanced vision-LLMs for scalable, human-aligned assessment (Fang et al., 11 Sep 2025).

1. Dataset Structure and Objectives

FLUX-Reason-6M was created to address critical bottlenecks in open-source T2I research that have traditionally lagged behind leading closed-source systems due to the lack of large-scale, high-resolution, reasoning-oriented datasets and standard evaluation protocols. The dataset consists of 6 million images synthesized using the FLUX.1-dev engine, each accompanied by dense, multi-label bilingual captions totaling 20 million descriptions.

Images and captions are systematically annotated according to six interrelated "reasoning axes" (Imagination, Entity, Text Rendering, Style, Affection, Composition), enabling explicit multi-labeling (i.e., an image can belong to several categories simultaneously). This structure is intended to enhance a model's ability to learn both single and compositional reasoning cues relevant to complex T2I tasks.

The dataset’s explicit goal is robustly formulated: to enable not merely visually plausible generation, but also detailed reasoning, logical composition, complex instruction following, multilingual synthesis, and precise text rendering—dimensions persistently challenging for autoregressive and diffusion-based T2I architectures.

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

A principal innovation of FLUX-Reason-6M is its implementation of "Generation Chain-of-Thought" (GCoT), which replaces or augments traditional flat captions with stepwise, rationale-rich breakdowns of the image generation logic. Instead of simply describing visible content, a GCoT annotation narrates the reasoning and decision-making process behind compositional choices, such as spatial layout, style selection, object relationships, emotional tone, and color harmony.

The GCoT annotations are generated using advanced vision-LLMs (Qwen-VL, Qwen3-32B, Gemini2.5-Pro), which process the input image alongside its category-specific caption to output a natural-language chain of reasoning. This process is designed to serve as an explicit supervisory signal, with the overarching generative function decomposed into a sequence of reasoning components:

$F(\text{image}) \approx f(R_1, R_2, \ldots, R_n)$

where each $R_i$ encodes a particular rationale or decision point that informs the final image.

By integrating GCoT, the dataset aims to train models that can not only generate a target image, but also internalize and replicate intermediate logical and creative steps—a critical capability for reasoning-centric T2I synthesis.

3. Data Curation Methodology

The data curation pipeline is multi-faceted and resource-intensive, with the entire process requiring approximately 15,000 A100 GPU days (128 A100 GPUs over four months) (Fang et al., 11 Sep 2025). The synthesis and annotation involve the following stages:

• Image Synthesis: FLUX.1-dev generates images guided by category-specific, high-quality prompts.
• Augmentation: For imagination, progressive prompt sampling uses Qwen3-32B and Gemini2.5-Pro at higher temperatures to create diverse, challenging scenes. For text rendering, a mining-generation-synthesis system extracts candidate images from Laion-2B, generates captions, and re-renders images to maximize legibility and contextual accuracy.
• Filtering and Scoring: Automated artifact removal is followed by vision-LLM (VLM) scoring. Qwen-VL assigns category relevance scores (1–10) for multi-labeling, and typographic quality scoring is applied to text-rendering samples.
• Dense Captioning and GCoT Annotation: For each image, dense category-specific captions and GCoT narratives are produced and cross-validated. External legacy captions are re-integrated when aligned to dataset quality standards.

As a result, FLUX-Reason-6M provides unprecedented label quality, multi-task coverage, bilingual annotations, and reasoning supervision at scale.

4. Reasoning Axes: Multi-Label Annotation

Each image and caption pair is annotated along six axes, constituting the dataset's core feature space:

Axis	Focus	Example Criteria
Imagination	Surrealism, abstraction	"A city made of glass with rivers of light"
Entity	Knowledge-grounded objects/persons	"A detailed depiction of a famous athlete"
Text Rendering	Legible, contextually rich text	"A sign with stylized, readable English font"
Style	Artistic/photographic modes	"Rendered in the style of Cubism or long exposure photo"
Affection	Emotions and mood	"Evocative lighting and facial expressions"
Composition	Spatial and relational arrangement	"The object is placed next to the landmark"

This multi-labeling structure supports both basic and higher-order compositional reasoning during model training. Notably, images and captions frequently embed several reasoning axes simultaneously (e.g., Entity + Style).

5. PRISM-Bench: Comprehensive Evaluation Protocol

PRISM-Bench is tightly coupled with FLUX-Reason-6M as a reference evaluation suite. The benchmark includes seven tracks: one for each axis (Imagination, Entity, Text Rendering, Style, Affection, Composition) and a "Long Text" challenge using extended GCoT prompts.

Each track features 100 prompts: 50 are semantically-clustered examples sampled from the dataset for even topic coverage, and 50 are LLM-generated to maximize challenge diversity. PRISM-Bench-ZH adapts all prompts into Chinese, adjusting for cultural specificity, especially in text rendering.

Assessment is delegated to advanced VLMs (e.g., GPT-4.1, Qwen2.5-VL-72B) as proxies for human judgment. Two standardized scores are reported per output image:

• Alignment: 1–10 scale, measuring prompt-image content congruence.
• Aesthetics: 1–10 scale, assessing technical image quality. Scores are aggregated to yield per-track and overall benchmark metrics, mapped onto a 0–100 scale for comparison.

6. Model Performance Insights and Limitations

Benchmarking over 19 leading T2I systems, including both open- and closed-source models, revealed clear gaps between current open-source models and top proprietary systems (e.g., GPT-Image-1, Gemini2.5-Flash-Image), particularly in imagination, composition, and long instruction following.

• High performance was observed on Entity and Style tracks, indicating that object fidelity and stylistic cues are more tractable to current models.
• The Text Rendering axis remains a significant challenge; most models, especially those using autoregressive architectures, are unable to reliably synthesize legible or contextually accurate text.
• The Long Text (GCoT-based) track exposes bottlenecks in models' abilities to integrate multifaceted, reasoning-heavy instruction sets.

These findings highlight the persistent need for improved multi-step reasoning, text handling, and compositional logic in T2I systems when measured against human-aligned, reasoning-rich prompts.

7. Significance and Prospective Developments

The FLUX-Reason-6M dataset and PRISM-Bench protocol fundamentally shift the landscape of open T2I research by providing a scale and annotation richness previously limited to large industrial labs. The explicit integration of GCoT reasoning supervision is poised to catalyze methodological advances in T2I model architectures, data curation, benchmarking, and reasoning capacity evaluation.

Immediate impacts include:

• Enhanced representation and compositional learning for T2I models.
• Broad international accessibility due to systematic bilingual labeling.
• A robust testbed for the community to explore new architectures, particularly in GCoT-augmented generation and cross-lingual synthesis.

Future directions suggested by performance gaps include further focus on text rendering, long-form reasoning, and more explicit modeling of intermediate compositional logic in generation pipelines. The combination of high-fidelity data, GCoT annotations, comprehensive benchmarking, and open access is expected to guide both incremental and structural improvements in T2I research and reasoning-centric generative modeling (Fang et al., 11 Sep 2025).