moda-fashion-matryoshka

Highlights

• 256-d ties 768-d on Fine R@1 (67.42 vs 67.21) — 3× index reduction for free
• 128-d still beats FashionSigLIP-768 by +2.39 R@1 at 6× smaller index
• 64-d matches FashionSigLIP-768 (+0.21) at 12× smaller index
• INT8 quantization at 256-d is lossless — 12× total compression over FashionSigLIP fp32

LookBench Results by Dimension

Quantization at 256-d (full LookBench protocol)

Model Spec

Inference — Quick Start

A standalone inference.py is included in this directory. It supports dimension selection and a full sweep.

# Default: encode at 256-d (recommended)
python inference.py --image query.jpg

# Specify dimension
python inference.py --image query.jpg --dim 128

# Two images + cosine similarity at 256-d
python inference.py --image img1.jpg img2.jpg --dim 256 --similarity

# Sweep all dimensions (64 → 768) with index cost
python inference.py --image img1.jpg --sweep

# Run on GPU/MPS
python inference.py --image query.jpg --device cuda

Python API — Truncate to Any Dimension

import open_clip
import torch
import torch.nn.functional as F
from PIL import Image

model, _, preprocess = open_clip.create_model_and_transforms(
    "ViT-B-16-SigLIP",
    pretrained="path/to/moda-fashion-matryoshka/open_clip_model.safetensors",
)
model.eval()

image = preprocess(Image.open("query.jpg")).unsqueeze(0)
with torch.no_grad():
    full_emb = model.encode_image(image)      # [1, 768]

dim = 256
emb_256 = F.normalize(full_emb[:, :dim], dim=-1)  # [1, 256]

INT8 Quantization

import numpy as np

emb = model.encode_image(images)[:, :256]
emb = F.normalize(emb, dim=-1).numpy()

emb_min = emb.min(axis=1, keepdims=True)
emb_max = emb.max(axis=1, keepdims=True)
emb_int8 = np.round((emb - emb_min) / (emb_max - emb_min + 1e-8) * 255).astype(np.uint8)

Requirements

open_clip_torch>=2.20.0
torch>=2.0
Pillow
safetensors

Training Details

• Base model: MODA-Fashion-Distilled (the 768-d distilled student)
• Method: Matryoshka Representation Learning with 6 nested slices {64, 128, 256, 384, 512, 768}
• Loss: Per-slice RKD-Distance + similarity mimicry against the 2048-d ensemble teacher
• Optimizer: AdamW, LR=5e-6, batch=128
• Epochs: 2 (best at step 1000)
• Wall time: 255 minutes on Apple M-series (MPS)

Related Models

License

MIT

Citation

If you use this model, please cite:

@software{moda2026,
  title  = {MODA: Open-source benchmark and models for fashion search},
  author = {Hopit AI},
  year   = {2026},
  url    = {https://github.com/hopit-ai/Moda}
}