FLUX-Reason-6M Bilingual Text-to-Image Dataset

## Overview FLUX-Reason-6M is a large-scale text-to-image dataset containing 6 million reasoning-focused images synthesized by the FLUX.1-dev model, paired with 20 million bilingual (English and Chinese) descriptions. The dataset is distributed as Parquet shards with image and text modalities, designed to bridge the performance gap between open-source and leading closed-source text-to-image systems by emphasizing complex compositional and reasoning prompts. ## Schema - image — binary/image — synthesized image generated by FLUX.1-dev - prompt_en — string — English prompt/description - prompt_zh — string — Chinese prompt/description - additional caption/metadata fields — string — extended descriptions and reasoning annotations - (See HF dataset page for full per-shard schema) ## Sources - HuggingFace: https://huggingface.co/datasets/LucasFang/FLUX-Reason-6M — Apache-2.0 - Associated paper: arXiv:2509.09680 ## Methodology Images were synthesized using the state-of-the-art FLUX.1-dev text-to-image model from carefully constructed reasoning-focused prompts. Each image is paired with multiple bilingual descriptions (English and Chinese) totaling ~20M captions across the 6M images. The dataset targets compositional, multi-step, and reasoning-heavy generation scenarios that open-source T2I models historically struggle with. ## Known gaps & limitations All images are synthetic outputs of FLUX.1-dev and inherit any biases, artifacts, or failure modes of that model — they are not photographs or human-curated art. Bilingual coverage is limited to English and Chinese. The dataset emphasizes reasoning prompts which may not represent the full distribution of real-world T2I use cases. Source does not exhaustively document gaps; buyers should validate empirically for their target use case. ## Intended use & out-of-scope - IS for: training and fine-tuning text-to-image generative models with reasoning capabilities, distillation from FLUX.1-dev, bilingual T2I research, prompt-image alignment studies. - NOT for: photorealism benchmarks against real-world images, training models intended for non-synthetic image evaluation, or use cases requiring languages beyond English/Chinese. _Federated dataset: 1,180 parquet shards, 821.78 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: FLUX-Reason-6M: 6M Reasoning-Focused Text-to-Image Dataset 6 million high-quality images synthesized by FLUX.1-dev with 20M bilingual (English/Chinese) descriptions, engineered to instill complex reasoning capabilities in text-to-image generative models.