# Studio 2 image→text caption (Qwen2.5-VL NSFW Caption V4 GGUF)

Caption jobs run on the Windows GPU host via **llama-server** (llama.cpp vision / mmproj). Not Comfy, not Klein, not JoyCaption.

## Weights (~7GB — two files only)

```powershell
powershell -ExecutionPolicy Bypass -File scripts\setup-caption.ps1
```

Installs into the Shared models tree:

`%LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Shared\models\caption\qwen25vl-7b-nsfw-v4\`

| File | Role |
| --- | --- |
| `Qwen2.5-VL-7B-NSFW-Caption-V4.Q5_K_M.gguf` | Language model |
| `Qwen2.5-VL-7B-NSFW-Caption-V4.mmproj-f16.gguf` | Vision projector |

Do **not** download the full 16.6GB safetensors repo.

Runtime needs ~8–10GB VRAM. Exclusive GPU: host stops Comfy (and waits if YuE2/upscale is busy), loads the GGUF for one shot, then kills llama-server (`keep_alive 0`).

## Host agent

Requires `llama-server` on PATH (`winget install ggml.llamacpp`) or `CAPTION_LLAMA_SERVER`.

Restart the Comfy host agent after setup. Confirm:

```text
GET http://127.0.0.1:8199/caption/status  →  { configured, busy, backend: "llama.cpp" }
POST /caption  { imagePath, style }       →  load → { text } → unload
POST /caption/jobs + PUT …/input          →  async job used by Studio 2
```

Styles: `descriptive` | `klein_prompt` | `delta` | `tags`.

## App

`POST /api/studio-2/caption` with `{ folderId, stillId|sourcePath, captionStyle }` queues a Studio job. Bench **Describe** button and the **caption** task pill both enqueue. Result stores `resultText` on the studio2 job and optionally writes `{still}.txt` beside the library file.