AbstractPhil/ffhq-recaptioned

FFHQ, recaptioned

New captions, booru tags and lighting readings for the 70,000 faces of FFHQ (Flickr-Faces-HQ, NVIDIA: Karras, Laine and Aila, 2019; aligned and cropped photos of 1024x1024 from Flickr). Each row is keyed by FFHQ's own image number; this repository holds none of its images or metadata. Get the images from FFHQ's page, under its terms.

data/ffhq_recaptioned.parquet (the default config): one row a face, 70,000 rows.

field	meaning
image, file	FFHQ's image number (00000-69999) and the file name it was read under
caption	the caption with its statements of absent items removed (below)
caption_raw	the caption as the model wrote it
new_tokens, truncated	its length in tokens; whether it reached the 160-token cap
tags_kept	the kept booru tags as one line
tags	[name, score, category, precision] for every tag, as a JSON string
shot_light	the lighting values the readings support, comma-separated (empty when none does)
luminance	the photo's mean luminance (Rec. 709, 0-1)
light	CinemaCLIP's lighting heads (cast, contrast, direction, edge, silhouette): every label's score and the labels over the model's own thresholds, as a JSON string

raw/ keeps the runs as they were written:

folder	rows	what
raw/20261105_085021_ffhq_captions	70,000	every face: the caption and the tags
raw/20261105_130209_ffhq_lighting	70,000	every face: CinemaCLIP's lighting heads, the luminance and the shot light values under the first rule (below)
raw/20261105_083559_ffhq_captions_smoke100	100	the first test, faces 00000-00099
raw/20261105_083756_ffhq_lighting_smoke100	100	the same faces' lighting
raw/20261105_083805_ffhq_shotvl_light_smoke100	100	a second lighting reader on the same faces: ShotVL-7B's open answer to its benchmark's question "What's the lighting of this shot?" (not used: it said "soft, from the side" of frontal flash)
raw/20261105_084828_ffhq_captions_askB_smoke100, ..._askC_..., ..._askD_...	100 each	three reworded asks on the same faces (below)

Captions. Qwen3.5-4B, greedy, thinking off, 160 new tokens at most, asked:

Describe this picture in one or two sentences: who is in it (apparent age, expression, hair, and any eyewear, headwear or jewelry), each visible garment with its color and any print or text on it, the background and the light. Answer with the visible description ONLY. Do not describe print or text if none is present.

Asked for "any eyewear, headwear or jewelry" and "any print or text", the model also names what is absent ("no eyewear, headwear, or jewelry", "with no visible print or text"): 68 and 36 of the first 100 captions. Three reworded asks on the same faces did not remove it (the best still wrote one in 55 of 100, with longer captions and fewer accessories named), so caption drops those statements with absences.drop from qwen-symbolic-thinking 0.1.52 and caption_raw keeps the original: 58,032 captions changed, 1,178 still name an absent item, most of them exceptions that name what is there ("no jewelry beyond the earrings"), kept on purpose. Other negatives stay ("no visible smile"). Median length 65 words; 290 captions reached the 160-token cap (truncated).

Tags. animetimm/eva_giant_patch14_560.dbv4-full, every tag with its score, category and published precision; tags_kept applies a leave-out list (short hair, see-through, see-through sleeves, see-through shirt, see-through clothes, hair bun, single hair bun, double bun, parted lips, lips, realistic, photorealistic, asian, old, cosplay) and a precision floor of 0.4. The tags are a drawing vocabulary read from photographs: 1girl / 1boy count people, and some tags fire on ordinary photos: cosplay on everyday photos (19 of the first 100 faces) and old on young adults (10 of 100), so both are left out of tags_kept and stay in tags.

Lighting. CinemaCLIP's released encoder (its label schema 1.0.0) reads five lighting heads; shot_light takes, in order: silhouette; the key light's side (from the left, from the right, from above, backlit; left and right are the picture's; front and ambient name none); rim light (edge light, unless backlit); low key (high contrast, mean luminance under 0.35) or high key (low contrast, over 0.65). The head's under label names nothing: of the four faces among the first 100 read as lit from below, the three checked by eye on a sheet were frontal flash shots (the raw/ lighting run still carries "from below" from the first rule). By eye on that sheet, high key (6) and low key (1) were right, and from above, from the left and from the right were right or generic (daylight from above). Counts over every face: from above 15,436, from the right 3,341, high key 2,381, rim light 1,353, low key 1,101, from the left 1,033, backlit 63; none 47,303. The reader names the right of the picture about three times as often as the left (3,341 against 1,033); the faces checked by eye were lit from the side it named, but the imbalance itself is unexplained and may be the reader's own.

FFHQ holds people of every age, children and babies included; the captions give the apparent age as seen.

Code: github AbstractEyes/qwen-symbolic-thinking (captions.run, lighting.heads_run v0.1.49; absences.drop v0.1.52, the light rule v0.1.50; notebooks/ffhq_captions.ipynb). The faces were read from gaunernst/ffhq-1024-wds, FFHQ's images1024x1024 transcoded losslessly to WebP under the original numbers.

Uses

Portrait captioning and caption evaluation, tag prediction against a booru vocabulary, lighting-aware captioning, and conditioning research for image models. Not for the development or improvement of facial recognition technologies (FFHQ's terms), and nothing here names or identifies anyone.

License

CC BY-NC-SA 4.0, as FFHQ itself: non-commercial use, credit by citing the paper below, changes indicated, and works built on this shared under the same license. FFHQ's images are not here; each keeps its own Flickr license (Creative Commons BY 2.0, BY-NC 2.0, Public Domain Mark 1.0, CC0 1.0 or U.S. Government Works), which FFHQ's metadata names per image. Made with Qwen3.5-4B (Apache-2.0), animetimm/eva_giant_patch14_560.dbv4-full (GPL-3.0) and CinemaCLIP (OpenRAIL-M).

Privacy

FFHQ lets photo owners leave it (the no_cv tag on Flickr, a license change, removal, or a request to NVIDIA), and NVIDIA updates the set. The rows here are keyed by image number so that a face removed from FFHQ can be removed here too: open a discussion on this repository with the image number.

Citation

@article{karras2018stylegan,
  title   = {A Style-Based Generator Architecture for Generative Adversarial Networks},
  author  = {Karras, Tero and Laine, Samuli and Aila, Timo},
  journal = {arXiv preprint arXiv:1812.04948},
  year    = {2018}
}