--- license: cc-by-4.0 task_categories: - image-classification - visual-question-answering language: - en tags: - human-preference - text-to-image - dpo - diffusion - pairwise-comparison - image-generation - preference-learning - rlhf - identity-preserving - flux pretty_name: "Text-2-Image DPO Human Preferences" size_categories: - 1K # Text-2-Image DPO Human Preferences A large-scale, quality-controlled human preference dataset for text-to-image generation. 80,000 trust-weighted pairwise judgments from calibrated annotators comparing AI-generated images across two evaluation dimensions: **prompt alignment** and **overall preference**. Built on the [Datapoint](https://trydatapoint.com) annotation platform — purpose-built infrastructure for collecting high-quality human preference data at scale. ## Overview | Metric | Value | |--------|-------| | Total judgments | 80,000 | | Unique calibrated annotators | ~12,500 | | Image pairs evaluated | 5,000 | | Judgments per pair per dimension | 8 | | Evaluation dimensions | 2 (prompt alignment, overall preference) | | Source dataset | [primecai/dsd_dpo_win_lose](https://huggingface.co/datasets/primecai/dsd_dpo_win_lose) | A smaller, higher-annotator-quality variant is available at [datapointai/text-2-image-dpo-human-preferences-small](https://huggingface.co/datasets/datapointai/text-2-image-dpo-human-preferences-small) (2,500 pairs, 40,000 judgments). ## Why This Dataset ### Multi-dimensional preference signals for DPO Most preference datasets collapse human judgment into a single "which is better?" score. But research shows that prompt faithfulness and aesthetic quality are **distinct, often conflicting** evaluation axes ([Rich Human Feedback, CVPR 2024](https://openaccess.thecvf.com/content/CVPR2024/papers/Liang_Rich_Human_Feedback_for_Text-to-Image_Generation_CVPR_2024_paper.pdf); [MPS, CVPR 2024](https://openaccess.thecvf.com/content/CVPR2024/papers/Zhang_Learning_Multi-Dimensional_Human_Preference_for_Text-to-Image_Generation_CVPR_2024_paper.pdf)). This dataset evaluates each image pair on two independent dimensions: | Dimension | Question | |-----------|----------| | **Prompt Alignment** | "Which image better matches the reference image and the description?" | | **Overall Preference** | "Which image do you prefer overall?" | Over a third of image pairs have **different winners** on these two dimensions — an image can be more faithful to the prompt but less visually appealing, or vice versa. This makes multi-dimensional labels critical for training models that balance both objectives. ### Human re-annotation of VLM-judged labels The source [DSD dataset](https://arxiv.org/abs/2411.18616) (CVPR 2025) used VLM scoring to determine win/lose pairs. Our dataset replaces these with calibrated human judgments, enabling: - **Direct comparison** of human vs. VLM preference alignment - **Higher-fidelity DPO training** on subjective dimensions where VLMs and humans may diverge - **Study of VLM-human agreement** in the image generation domain ### Calibrated annotator quality Unlike crowdsourced datasets where all votes count equally, every annotation in this dataset comes from an annotator whose reliability has been measured through gold-standard calibration tasks. Votes are weighted by each annotator's demonstrated accuracy, and only the top annotators by calibration quality are selected for each comparison. ## Source Data This dataset provides human re-annotations of the [primecai/dsd_dpo_win_lose](https://huggingface.co/datasets/primecai/dsd_dpo_win_lose) dataset from the **Diffusion Self-Distillation (DSD)** project ([Cai et al., CVPR 2025](https://arxiv.org/abs/2411.18616)). The source contains 30,400 image triplets (reference image, winning image, losing image) generated by a fine-tuned **FLUX.1-dev** model (Black Forest Labs) for zero-shot identity-preserving image generation. ## Dataset Structure | Column | Type | Description | |--------|------|-------------| | `prompt` | string | Text description / caption for the image generation task | | `image_a` | Image | Candidate image A | | `image_b` | Image | Candidate image B | | `reference_image` | Image | Reference / source image | | `prompt_id` | int | Unique identifier for the source datapoint | | `weighted_results_a_Prompt_Alignment` | float | Trust-weighted vote fraction for A (prompt alignment) | | `weighted_results_b_Prompt_Alignment` | float | Trust-weighted vote fraction for B (prompt alignment) | | `detailedResults_Prompt_Alignment` | list | Per-annotator results: winner, weight, trust_score, calibration_responses, time_taken_ms | | `num_annotations_Prompt_Alignment` | int | Number of annotations (always 8) | | `weighted_results_a_Overall_Preference` | float | Trust-weighted vote fraction for A (overall preference) | | `weighted_results_b_Overall_Preference` | float | Trust-weighted vote fraction for B (overall preference) | | `detailedResults_Overall_Preference` | list | Per-annotator results for overall preference | | `num_annotations_Overall_Preference` | int | Number of annotations (always 8) | ## Usage ```python from datasets import load_dataset ds = load_dataset("datapointai/text-2-image-dpo-human-preferences") # Determine the preferred image per dimension for row in ds["train"]: pa_winner = "A" if row["weighted_results_a_Prompt_Alignment"] > row["weighted_results_b_Prompt_Alignment"] else "B" op_winner = "A" if row["weighted_results_a_Overall_Preference"] > row["weighted_results_b_Overall_Preference"] else "B" ``` ### Use for DPO Training ```python # Convert to chosen/rejected pairs using overall preference dpo_pairs = [] for row in ds["train"]: wa = row["weighted_results_a_Overall_Preference"] wb = row["weighted_results_b_Overall_Preference"] if wa > wb: dpo_pairs.append({"prompt": row["prompt"], "chosen": row["image_a"], "rejected": row["image_b"], "confidence": wa}) elif wb > wa: dpo_pairs.append({"prompt": row["prompt"], "chosen": row["image_b"], "rejected": row["image_a"], "confidence": wb}) ``` ### Custom Aggregation The `detailedResults` columns provide per-annotator data, enabling custom re-aggregation: ```python row = ds["train"][0] details = row["detailedResults_Prompt_Alignment"] # Only count annotators with trust_score >= 0.8 high_trust = [d for d in details if d["trust_score"] >= 0.8] a_votes = sum(1 for d in high_trust if d["winner"] == "a") b_votes = sum(1 for d in high_trust if d["winner"] == "b") ``` ## Comparison to Related Work | | This dataset | [Pick-a-Pic](https://arxiv.org/abs/2305.01569) | [ImageReward](https://arxiv.org/abs/2304.05977) | [HPD v2](https://arxiv.org/abs/2306.09341) | DSD (original) | |---|---|---|---|---|---| | Size | 5,000 pairs | 500K+ pairs | 137K comparisons | 798K preferences | 30,400 pairs | | Annotation source | Calibrated humans | Crowdsourced (unfiltered) | Expert annotators | Curated annotators | VLM-judged | | Quality control | Calibration-based trust scoring | None | Expert training | Curation | Automated | | Dimensions | 2 (alignment + preference) | 1 (preference) | 3 (alignment, fidelity, harmlessness) | 1 (preference) | 1 (VLM score) | | Annotations per pair | 8 (trust-weighted) | 1-2 | 3-5 | ~3 | 1 (VLM) | | Position randomization | Yes | Yes | N/A | N/A | N/A | | Reference image | Yes | No | No | No | Yes | | Base model | FLUX.1-dev | Multiple | Multiple | Multiple | FLUX.1-dev | | Per-annotator detail | Yes | No | No | No | No | ### Key Differentiators - **Calibration-based quality control at scale**: Unlike Pick-a-Pic (unfiltered crowdsourcing) or ImageReward (small expert panel), Datapoint's approach continuously measures annotator reliability through gold-standard calibration tasks, enabling quality control that scales with the annotator pool. - **Multi-dimensional evaluation**: Separating prompt alignment from overall preference captures the documented tension between faithfulness and aesthetic quality, providing richer training signal for DPO. - **Reference-conditioned comparison**: Each pair includes a reference image, enabling evaluation of identity-preserving generation — a capability most preference datasets lack. - **Human re-annotation of VLM labels**: Provides calibrated human preferences for data originally labeled by VLMs, enabling study of human-VLM alignment. - **Per-annotator transparency**: Raw per-annotator results (including trust scores and response times) are included, enabling researchers to apply their own aggregation strategies. ## License CC-BY-4.0 ## Citation ```bibtex @dataset{t2i_dpo_human_preferences_2026, title={Text-2-Image DPO Human Preferences: Trust-Weighted Pairwise Judgments for Image Generation}, author={Datapoint AI}, year={2026}, url={https://huggingface.co/datasets/datapointai/text-2-image-dpo-human-preferences}, note={80,000 calibrated pairwise judgments across 5,000 image pairs and 2 evaluation dimensions} } ``` ### Source Dataset Citation ```bibtex @inproceedings{cai2025dsd, title={Diffusion Self-Distillation for Zero-Shot Customized Image Generation}, author={Cai, Shengqu and Chan, Eric Ryan and Zhang, Yunzhi and Guibas, Leonidas and Wu, Jiajun and Wetzstein, Gordon}, booktitle={CVPR}, year={2025} } ``` ## Built with Datapoint This dataset was collected using [Datapoint](https://trydatapoint.com) — a data labelling platform designed for high-quality human preference data. Datapoint's calibration-based trust scoring system enables quality-controlled annotation at scale, making it possible to build preference datasets where every label is backed by measured annotator reliability. Questions or feedback: sales@trydatapoint.com