AbstractPhil/deepfashion-multimodal-recaptioned

DeepFashion-MultiModal, recaptioned

New captions, booru tags and 3D body angles for the 42,544 described photos of DeepFashion-MultiModal (from Text2Human, Jiang et al., 2022: 44,096 photos of 750x1101 with parsing, keypoints, DensePose, shape and texture labels and a description for each), whose photos come from the DeepFashion database of the CUHK Multimedia Laboratory (Liu et al., 2016). Each row is keyed by the dataset's own file name; this repository holds none of its photos, descriptions, parsing maps, keypoints or label files. Get the photos from the dataset's official page, under its agreement.

data/recaptioned.parquet (the default config): one row a photo, with view (front, side, back, full, additional, flat: the view in the file name), caption (cleaned of absence statements, below), caption_raw (the caption as written), caption_pose (the clean caption with the body fit's pose sentence appended), pose_line, tags_kept, and tags and pose as JSON strings. raw/ keeps the runs as they were written:

folder	rows	what
raw/20261105_001419_deepfashion_mm_captions	42,544	a caption from the stock Qwen3.5-4B and booru tags from the animetimm EVA-giant tagger (missing.jsonl: the first pass's 58 out-of-memory misses, all captioned by the resume)
raw/20261105_030856_deepfashion_mm_pose	42,544	a 3D body fit for each photo (SAM 3D Body, ViT-H): angles, rotation frames and 70 3D keypoints; 42,530 fitted, 14 flat lays skipped
raw/20261105_035546_deepfashion_mm_captions_pose	42,544	the two joined (the source of the parquet)
raw/20261105_030333_deepfashion_mm_pose_smoke100	100	the fit's first test
raw/20261105_030753_deepfashion_mm_captions_pose_ask256	256	a prompt test: captions whose ask carried the pose line (in the subject's terms; the captioner then named the wrong side of the picture in 17 of 22 turned photos, the reason the pose sentence is appended instead)
raw/20261105_075031_deepfashion_mm_lighting_trial100	100	a lighting trial (qwen-symbolic-thinking 0.1.48, lighting.trial): ShotVL-7B answers its benchmark's two lighting questions, open and as a choice, and CinemaCLIP scores its lighting heads (cast, contrast, direction, edge, silhouette); on these studio photos both read soft, low-contrast light from the front almost everywhere

Captions. Qwen3.5-4B, greedy, thinking off, 160 new tokens at most, asked with captions.FASHION_ASK (each garment with its color and any print or text on it, the pose and the setting) and the photo's own description as a helper line. copy5, the share of a caption's five-word runs found word for word in that description: mean 0.001, the largest 0.098. Tags: animetimm/eva_giant_patch14_560.dbv4-full, every tag with its score, category and published precision; tags_kept applies a reviewer's leave-out list and a precision floor of 0.4 (clothing exempt).

Absences. Asked for "any print or text", the captioner also says when there is none ("with no visible print or text"): caption drops those statements, and those of absent eyewear, headwear, jewelry or accessories, with absences.drop from qwen-symbolic-thinking 0.1.51 (12,684 of 42,544 captions changed, 132 still name an absent item; every other negative stays); caption_raw keeps the caption as the model wrote it.

Body angles. SAM 3D Body fits a body to the photo; the hips, the chest (the shoulder line) and the head become frames read from the fit's named 3D keypoints, each with a turn (+ toward the subject's left, which is the right of the picture for a person facing the camera), a nod (+ up) and a tilt (+ toward the subject's left shoulder) against the camera; twist is the shoulders against the hips, head_on_chest the head against the chest, lean the torso against the camera's vertical (picture: + its top toward the right of the picture). The pose sentence names where the body, shoulders and head face in the picture's own terms ("The body is in profile, facing the right of the picture; the head is turned 30 degrees toward the left of the picture."), turns rounded to 5 degrees. The pose run's own pose_line keeps an earlier wording in the subject's terms; use the joined folder's.

The fit against the views in the file names (absolute turn of the hips from the camera, degrees):

view	photos	0-30	30-60	60-120	120-150	150-180
front	12,278	10,390	1,037	557	238	56
full	6,455	5,494	708	146	92	15
side	8,685	238	976	7,430	26	15
back	8,227	78	10	128	1,434	6,577
additional	6,885	2,729	513	1,413	866	1,364

The in-between counts are mostly three-quarter poses the views do not name; the outright contradictions (a front or full photo read as facing away, a back photo read as facing the camera: 1-2.5%) that were read by eye were label errors, the fit right.

field	meaning
file	the photo's DeepFashion-MultiModal file name
caption, caption_raw	the caption without its absence statements; the caption as written
new_tokens, truncated, helped, copy5	the raw caption's length in tokens, whether it hit the cap, whether a description helped, its copy share
tags, tags_kept	[name, score, category, precision] for every tag; the kept tags as one line
pose	hips, chest, head, head_on_chest: {turn, nod, tilt}; twist; lean: {forward, left, picture}; frames (3x3, columns left / up / forward in the camera's frame); focal_length; keypoints_3d (70, the pose folder only)
pose_line, caption_pose	the pose sentence; the clean caption with it appended (the joined folder's caption_pose appends it to the raw caption)

Code: github AbstractEyes/qwen-symbolic-thinking (captions.run v0.1.41, pose v0.1.43, captions.append_pose v0.1.45, absences.drop v0.1.51; notebooks/deepfashion_captions.ipynb). The photos were read from tilak1114/deepfashion, a full-size copy of the described photos, because the dataset's own download links refuse cloud machines.

Uses

Fashion captioning and caption evaluation, tag prediction against a booru vocabulary, pose-aware captioning, and body-orientation labels (front, three-quarter, profile, back, from the hips' turn) for photos of dressed people. The captions are plain descriptions, not reasoning traces.

License

The captions, tags and angles here are released under CC BY-NC 4.0, for non-commercial research use only: DeepFashion-MultiModal's agreement extends its non-commercial condition to data derived from it, and its photos and their use stay under that agreement. The body angles were made with SAM 3D Body (Meta, facebookresearch/sam-3d-body, the SAM License; its body rig MHR is Apache-2.0), whose use is acknowledged here; the captions with Qwen3.5-4B (Apache-2.0); the tags with animetimm/eva_giant_patch14_560.dbv4-full (GPL-3.0).

Citation

@article{jiang2022text2human,
  title={Text2Human: Text-Driven Controllable Human Image Generation},
  author={Jiang, Yuming and Yang, Shuai and Qiu, Haonan and Wu, Wayne and Loy, Chen Change and Liu, Ziwei},
  journal={ACM Transactions on Graphics (TOG)},
  volume={41},
  number={4},
  articleno={162},
  pages={1--11},
  year={2022},
  publisher={ACM New York, NY, USA},
  doi={10.1145/3528223.3530104},
}

@inproceedings{liuLQWTcvpr16DeepFashion,
 author = {Liu, Ziwei and Luo, Ping and Qiu, Shi and Wang, Xiaogang and Tang, Xiaoou},
 title = {DeepFashion: Powering Robust Clothes Recognition and Retrieval with Rich Annotations},
 booktitle = {Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
 month = {June},
 year = {2016}
 }

@article{yang2026sam3dbody,
  title={SAM 3D Body: Robust Full-Body Human Mesh Recovery},
  author={Yang, Xitong and Kukreja, Devansh and Pinkus, Don and Sagar, Anushka and Fan, Taosha and Park, Jinhyung and Shin, Soyong and Cao, Jinkun and Liu, Jiawei and Ugrinovic, Nicolas and Feiszli, Matt and Malik, Jitendra and Dollar, Piotr and Kitani, Kris},
  journal={arXiv preprint arXiv:2602.15989},
  year={2026}
}