README.md

metadata

title: MegaStyle Image Style Comparison
emoji: 🎨
colorFrom: purple
colorTo: pink
sdk: gradio
sdk_version: 5.50.0
python_version: '3.10'
app_file: app.py
pinned: false
short_description: Compare image style similarity with MegaStyle-Encoder
license: mit

Deploying this Space: select ZeroGPU hardware in the Space's Settings → Hardware panel after creating it. ZeroGPU is not configured via frontmatter.

MegaStyle Image Style Comparison

Upload a test image and 1–8 reference images, hit Compare styles, and get a style-similarity score (0–100) plus a human-readable verdict. Powered by MegaStyle-Encoder, a SigLIP-based style encoder trained on the 1.4M-image MegaStyle dataset with style-supervised contrastive learning — see paper MegaStyle (arXiv:2604.08364).

How it works

1. Each image is embedded with MegaStyle-Encoder into a unit-length style vector.
2. Cosine similarity between the test vector and each reference vector gives a per-reference score.
3. The headline score is the mean of those per-reference scores, shown as a percentage for readability. A per-reference table is shown below for transparency.

Verdict labels

The verdict is a heuristic bucketing of the cosine-similarity score:

Score range	Label
≥ 0.75	🟢 Strong style match
0.65 – 0.75	🟢 Good style match
0.55 – 0.65	🟡 Moderate style match
0.45 – 0.55	🟠 Weak style match
< 0.45	🔴 Minimal style match

These thresholds are not calibrated against ground-truth style labels — they are rule-of-thumb bands tuned for the typical cosine-similarity range of SigLIP-family encoders (where even unrelated images can sit around 0.4–0.6). Treat the raw number as the source of truth.

Credits

• Paper: Gao et al., MegaStyle: Constructing Diverse and Scalable Style Dataset via Consistent Text-to-Image Style Mapping, arXiv:2604.08364, 2026.
• Upstream code: Tencent/MegaStyle
• Model weights: Gaojunyao/MegaStyle (MIT)
• Backbone: google/siglip-so400m-patch14-384 (Apache-2.0)