---
title: Trendyol Vision Flash
emoji: ⚡
colorFrom: green
colorTo: gray
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
short_description: E-commerce catalog VLM for moderation and enrichment
python_version: "3.12"
startup_duration_timeout: 30m
pinned: false
license: cc-by-4.0
---

# Trendyol-Vision-Flash

Demo of [Trendyol/Trendyol-Vision-Flash](https://huggingface.co/Trendyol/Trendyol-Vision-Flash),
a 1.06B-parameter vision-language model (InternViT-300M + Qwen3-0.6B, fine-tuned from
[OpenGVLab/InternVL3_5-1B-Instruct](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Instruct))
built for e-commerce catalog quality and moderation workflows in Turkish and English.

Supported catalog tasks, each with its official prompt template preloaded:

- **Brand detection** — infer the brand from a product photo, with category context.
- **Attribute extraction** — structured JSON attributes from image + title/description.
- **Title generation** — a clean, moderation-compliant catalog title.
- **Product caption** — grounded English product description.
- **Content safety classification** — `0` Forbidden / `1` Fantasy (+18) / `2` Safe.
- **Per-pack quantity extraction** — `{"amount": …, "unit": "KG"|"L"|"PIECE"}`.
- **Product similarity** — are two listings the same SKU? `1` / `0`.

Inference follows the model card's reference path exactly: `AutoModel.chat()` with
`trust_remote_code=True`, a single bicubic-resized 448×448 tile per image, bf16, and greedy
decoding.

## Credits

Prompt templates and example product photos are taken from the official
[Trendyol-Vision-Flash](https://huggingface.co/Trendyol/Trendyol-Vision-Flash) and
[Trendyol-Vision-Master](https://huggingface.co/Trendyol/Trendyol-Vision-Master) model cards,
released by the Trendyol Data Science Team under
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).