The dataset viewer is not available for this split.
Error code: TooBigContentError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
PhotoMaster-Data
PhotoMaster: Learning Holistic Capture-Time Decisions for Guided Photography Reshooting
Learn what to change before taking the next shot.
Zian Qian · Cheuk Him Chau · Siyue Xie · Chufeng Xiao · Qifeng Chen · Rui Liu · Chenyang Lei
HKUST · Huawei Research
Paper Figure 1: an initial photo, PhotoMaster guidance, and the novice's actual guided reshoot. A generated reference and a final camera capture are distinct outputs.
PhotoMaster-Data supports research on capture-time photographic decisions: diagnosing a photo, suggesting practical actions, and generating a visual target for a new shot. Its supervision follows the Photography Decision Taxonomy (PDT), covering 12 dimensions across subject, viewpoint, framing and optics, time, light and color, and scene geometry.
The paper describes training triplets consisting of a novice-side image, PDT supervision, and a professional or professional-style reference. One image endpoint may be synthesized. The downloadable release is stored as Parquet rows; its documented layout is described below.
Downloadable release
| Property | Published value |
|---|---|
| Repository | qwmqza/PhotoMaster |
| Configuration / split | default / train |
| Rows | 186,405 |
| Files | 111 Parquet shards |
| Original file size | 85,047,026,604 bytes (approximately 85.05 GB) |
| Revision | 94646f5fddcc5b9ba5e3c2c9c4d5cb0a84c0fdbd |
The release is publicly accessible without an access request. Files are named photomaster-00001-of-00111.parquet through photomaster-00111-of-00111.parquet. These counts and types follow the Hugging Face dataset metadata and repository inventory, checked on October 30, 2026.
The paper reports 236,904 training triplets. The relationship between that total and the 186,405 uploaded rows has not yet been documented; an uploaded row should not be assumed to correspond to one complete paper triplet.
Load a sample without downloading the full dataset
Install the Hugging Face Datasets library:
python -m pip install datasets
Use streaming with a pinned revision:
from datasets import load_dataset
dataset = load_dataset(
"qwmqza/PhotoMaster",
name="default",
split="train",
revision="94646f5fddcc5b9ba5e3c2c9c4d5cb0a84c0fdbd",
streaming=True,
)
sample = next(iter(dataset))
print("Columns:", list(sample))
print("Sample ID:", sample["sample_id"])
print("Frame field type:", type(sample["frame_webp"]).__name__)
print("Annotation field type:", type(sample["frame_json"]).__name__)
Streaming reads data on demand and avoids downloading the entire release. This example inspects the published fields without assuming an image-pair or annotation layout.
Data fields
| Column | Arrow type |
|---|---|
sample_id |
string |
split |
string |
unsplash_photo_id |
string |
unsplash_url |
string |
series_idx |
string |
label_idx |
string |
frame_webp |
binary |
frame_json |
string |
frame_webp is a binary field, not a Hugging Face Image feature that automatically returns a decoded image. frame_json is stored as a serialized string, not as a typed, nested annotation object. Its internal keys and their relationship to image pairs and PDT labels are not yet documented for this release.
The row-level split field is separate from the exposed Hugging Face train split. Its values and distribution have not been verified. No validation or test split is currently exposed by the Dataset Viewer.
How the data was constructed
The paper introduces two complementary pipelines:
- Forward Decision Transfer (FDT) starts from real novice photos and transfers PDT target states from scene-compatible professional references. It synthesizes a professional-style endpoint, emphasizing decisions such as lighting, color, pose, and expression.
- Decision Inversion (DI) starts from professional-quality photos and synthesizes novice-like counterparts through inverse capture decisions. A camera-controllable world model varies viewpoint, framing, and scene timing, strengthening supervision for spatial and temporal decisions.
For FDT, the paper describes a gallery of 187 photographs from established photographers and photography educators, spanning 129 scene categories. Gemini 3.1 Pro extracts and transfers target decisions, and Gemini 3 Pro Image synthesizes the reference endpoint.
For DI, the paper uses high-quality images from the Unsplash dataset as anchors, filtered by an aesthetic scoring model. Lingbot-World generates camera-controlled sequences, from which candidate frames are selected. Filtering removes weak inversions and semantic failures; Gemini 3 Pro Image repairs local artifacts.
A unified pairwise annotation stage compares the novice-side and reference images using Gemini 3.1 Pro. The conceptual supervision includes active PDT dimensions, diagnoses, executable reshooting actions, and a reference-rendering prompt. These describe the paper's data design; a complete mapping to the hosted row format is still being prepared.
Intended use and training
The dataset is intended for research on photography guidance, decision-level visual reasoning, and instruction-conditioned reference generation. Reshooting actions concern operations before capture, such as moving the camera, changing framing, adjusting exposure, or guiding a subject's pose.
The code repository includes VLM and DiT training implementations. Its VLM trainer expects a LLaMA-Factory ShareGPT registry, while the DiT trainer expects image archives and Parquet annotations. A verified conversion from this hosted schema into both training formats is not yet provided. The GitHub repository is currently private and requires access permission.
Benchmark scope
The paper's PhotoMaster-Bench contains 377 human-curated cases. The five images bundled with the code are inference examples with no populated ground-truth diagnosis labels; they are not the full benchmark. This dataset release does not establish the complete benchmark annotations, evaluation protocol, or training–benchmark overlap checks needed to reproduce the reported diagnosis scores.
Sources, limitations, and data terms
The data engine combines real photographs, generated image endpoints, and model-produced annotations. Generation can introduce artifacts or change subject identity, and annotations can miss photographic issues. The paper's training collection is portrait-heavy (over 75%); the category distribution of the hosted rows has not been independently established. PDT reflects a particular operational taxonomy and a limited expert pool, rather than every photographic style or context.
The release includes Unsplash identifier and URL fields. Full per-example provenance, source permissions, redistribution scope, and the correspondence between synthetic frames and their anchors still need documentation. Public availability of these files does not by itself define permission for every intended use.
A dataset-specific license has not yet been published. The cited paper's license does not automatically apply to the dataset, source photographs, model weights, or code. Applicable source terms and dataset permissions should be confirmed before reuse or redistribution.
Citation
@article{qian2026photomaster,
title = {PhotoMaster: Learning Holistic Capture-Time Decisions for Guided Photography Reshooting},
author = {Qian, Zian and Chau, Cheuk Him and Xie, Siyue and Xiao, Chufeng and Chen, Qifeng and Liu, Rui and Lei, Chenyang},
journal = {ACM Transactions on Graphics},
volume = {45},
number = {6},
articleno = {220},
numpages = {14},
year = {2026},
doi = {10.1145/3842543}
}
- Downloads last month
- 382
