README.md

---
title: Caption Space
emoji: 🎨
colorFrom: indigo
colorTo: gray
sdk: gradio
sdk_version: 5.0.0
app_file: app.py
pinned: false
hardware: zero-a10g
---

# Caption Space — LLaVA-1.5-7B

Vision-language descriptions of inspiration images for the [La Compagnie
d'Intérieur](https://github.com/sciencellama) pipeline.

## What it does

Takes one or more inspiration images (mood boards / Pinterest references) and
returns a natural-language interior-design description: style/era, color
palette, materials, key furniture, patterns, vibe.

The main pipeline uses this output as a starting point for the **style prompt**
that drives FAISS retrieval — the user reads the description, edits it ("but
with more orange tones"), then submits the redesign job with the edited prompt.

## API contract

`api_name="/describe"`

| Input | Type | Notes |
|---|---|---|
| `images_b64`  | str | Single base64 image OR JSON-array string of base64 images |
| `instruction` | str | Optional override; empty string = use the default design prompt |

Returns `{"description": "<prose>"}`.

## Multi-image inputs

LLaVA-1.5 takes one image per call. For N inspiration images, the Space calls
the model N times and joins the outputs with `" Also: "`. The user can then
edit the merged result manually. (Future: swap to LLaVA-OneVision / Qwen2-VL
for native multi-image, when revenue justifies the larger model and slower
cold-start.)

## Why LLaVA-1.5-7B

- Strong style-vocabulary captioning out of the box
- Open-source, no commercial licensing required
- Mature `transformers` integration
- ~14 GB at fp16 → fits ZeroGPU A10G with room to spare
- Faster cold-start than larger alternatives (LLaVA-NeXT, Qwen2-VL-7B)