Theia and ArchiveAgent benchmarks
Collection
Theia, a fashion model, and its benchmarks. Request-access datasets: the Theia test set, and local image-model tests. A personal project. • 2 items • Updated
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This repository holds results, protocols, reports and code from our local image-model benchmarks. It has no photos. It is shared for research and for checking our results. Requests are approved by hand.
Log in or Sign Up to review the conditions and access this dataset content.
Local image models on fashion photos
A personal project, run on one desktop GPU on a small budget. This is the first test we ran in the ArchiveAgent project, before Theia: how well do four open image models, run locally, find the right fashion house in runway photos, and how long does it take to run them?
Part of the Theia project: the Theia test set and results are in kalanpeace/theia-bench. Both of our datasets are together in this Hugging Face collection. Everything here ran on one desktop computer with an NVIDIA RTX 5060 (8 GB of VRAM).
There are no images, vectors or per-photo files in this repository. It holds the written-down protocols (each with its SHA-256, written before the runs they cover), the result files, the reports and the code, exactly as we used them.
The data (counts)
- 86,874 distinct runway photos from 15 fashion houses and 1,298 shows, 1991 to 2026. Labels are the show page each photo came from (house, show, season, year). The photos are Vogue Runway coverage and are not shared.
- 8,693 product photos from 908 resale, brand-site and auction listings (used for the product test below).
- All photos were embedded from 768-pixel copies, same recipe for every model.
The four models
| Model | Vector size |
|---|---|
| CLIP ViT-B/16 (LAION-2B) | 512 |
SigLIP 2 So400m (google/siglip2-so400m-patch14-384) |
1152 |
| DINOv3-L | 1024 |
| PE-Core-L14-336 (Perception Encoder) | 1024 |
How we tested (runway)
- Gallery: all 86,874 runway photos. Queries: 5 random photos per show, 6,480 queries from 1,296 shows (seed 20261002). Exact cosine search (not approximate). The query never sees itself, and near-duplicates (perceptual-hash distance 2 or less) never count.
- Easy test: is the top result from the same show? Hard test: the query's own show is removed from the results; is the top result the right house? Is it the right house and year?
- 95% intervals come from resampling whole shows (photos of one show are not independent). A shuffled-answer-key control must collapse to chance on every run; it did.
- All protocols were written and hashed before the runs they cover (
protocols/). We found and fixed one scorer bug on the first day (show names repeat across houses); all runs were redone and the effect is written up in the reports.
Results
| Model | Easy: same show, top 1 | Hard: right house, top 1 | Hard: right house and year, top 1 |
|---|---|---|---|
| CLIP ViT-B/16 (LAION-2B) | 91.6% (90.7 to 92.4) | 52.2% (50.4 to 53.9) | 9.4% (8.3 to 10.4) |
| SigLIP 2 So400m | 93.7% (93.0 to 94.4) | 74.4% (72.7 to 76.0) | 12.8% (11.5 to 13.9) |
| DINOv3-L | 89.4% (88.3 to 90.4) | 48.7% (46.7 to 50.6) | 8.2% (7.2 to 9.3) |
| PE-Core-L14-336 | 96.4% (95.8 to 96.9) | 67.9% (66.1 to 69.6) | 12.2% (11.0 to 13.3) |
| Random guess | 0.07% | 8.7% | 0.3% |
Read this carefully: the winner changes with the test. PE-Core is best at finding the same show; SigLIP 2 is best at naming the house once the show is hidden. Every model drops a long way from easy to hard (for example SigLIP 2 names the right house 98.9% of the time when the show is not hidden and 74.4% when it is), because the easy test partly rewards matching the same photo session. The year is weak for everyone. DINOv3-L was last on both tests, and the follow-up checks (a trained probe, other vectors from the same model, hard negatives) did not close the gap. See reports/ and results/ for the paired differences and the checks.
How long it took (all 86,874 runway photos, per model)
| Model | Vector size | Precision | Time | Photos per second | Peak GPU memory | Hottest |
|---|---|---|---|---|---|---|
| CLIP ViT-B/16 (LAION-2B) | 512 | fp16 | 46.0 min | 31.5 | 0.44 GB | 59 C |
| SigLIP 2 So400m | 1152 | fp16 | 52.7 min | 27.4 | 1.02 GB | 67 C |
| DINOv3-L | 1024 | fp32 | 57.0 min | 25.4 | 1.38 GB | 74 C |
| PE-Core-L14-336 | 1024 | fp16 | 46.0 min | 31.5 | 0.83 GB | 64 C |
Total embedding compute for the four models: about 202 minutes (3.4 hours). Searching all 86,874 vectors exactly takes a median of 18 ms per query for SigLIP 2. The first three models ran at almost the same speed (about 31 photos per second, although one is small and one is large). That points to reading and decoding the image files as the limit, not the model; we did not prove it. DINOv3-L could not run in 16-bit (it produced non-finite values), so it ran in 32-bit. The GPU model is not recorded in the result files; we ran them on an NVIDIA RTX 5060 with 8 GB of VRAM.
Product photos (same four models)
Same-listing retrieval, 8,666 queries from 902 listings (easy by design):
| Model | Recall at 1 (95% interval) | Recall at 5 | mAP at 20 |
|---|---|---|---|
| CLIP ViT-B/16 (LAION-2B) | 65.3% (63.73 to 66.87) | 78.7% | 30.4 |
| SigLIP 2 So400m | 77.9% (76.5 to 79.25) | 88.8% | 46.9 |
| DINOv3-L | 51.5% (49.74 to 53.27) | 65.5% | 17.9 |
| PE-Core-L | 77.6% (76.27 to 78.99) | 88.1% | 45.9 |
All 8,693 product photos were embedded by all four models in 12.2 minutes. Brand on the verified products scored 99.5% to 100% but only two houses were involved (Rick Owens 66 items, Louis Vuitton 6) and a majority-class guess already scores 90%, so it says little. The harder product tests (same product across listings, product to runway) were not run. Details and limits are in reports/REPORT_product_2026-10-03.md.
Other checks in this folder
Method A (a trained probe on the frozen vectors), B (hard negatives), C (other vectors from the same model, plus a small trained adapter), D (report metrics), a validity audit, a gallery-growth experiment, a collection-retrieval and exact-image test, and the Stanford Online Products test split (60,502 eBay images). Each has a summary file under results/ and a frozen protocol under protocols/.
Limits, said plainly
- Runway labels are show-level only, and this is exploratory retrieval, not validated fashion identification: hiding the own show does not remove every shortcut (same photographer, same era look).
- One seed and 5 queries per show. DINOv3-L ran in 32-bit and the others in 16-bit.
- Product brand results are on a small set (2 houses, mostly one), so those scores are saturated.
- The photos and vectors are not shared, so you cannot re-run it from this folder alone. The code and protocols show exactly how it was run. Researchers: email kalanpeace@gmail.com.
Access
The card is visible to everyone; the files are behind a request button and I approve requests by hand. You can also email kalanpeace@gmail.com. Results and analysis: CC BY 4.0. Code: Apache-2.0. No photos are included.
Files
protocols/ frozen protocols with hashes; reports/ the comparison, the Genesis rerun and product reports, proof manifests; results/ summaries, proof checks and manifests per run (no per-query files, no vectors); code/ the scripts; MANIFEST.sha256 SHA-256 of every file here.
- Downloads last month
- 25