---
title: Whyx-PROmpTea
canonical_url: "https://www.modelscope.ai/studios/Whyxeo/Whyx-PROmpTea"
md_url: "https://www.modelscope.ai/studios/Whyxeo/Whyx-PROmpTea.md"
repository: Whyxeo/Whyx-PROmpTea
last_updated: 2026-10-10
sdk_type: gradio
sdk_version: 6.17.3
downloads: 0
stars: 1
---

# Whyx-PROmpTea

> Whyx-PROmpTea - An interactive demo by Whyxeo on ModelScope. Prompt Enhancement Studio — Anima & Illustrious XL

Whyxeo/Whyx-PROmpTea is an interactive demo (Studio) on ModelScope. built with gradio 6.17.3.

- **Repository**: Whyxeo/Whyx-PROmpTea
- **SDK**: gradio
- **SDK version**: 6.17.3
- **Downloads**: 0
- **Stars**: 1
- **Last updated**: 2026-10-10

Source: https://www.modelscope.ai/studios/Whyxeo/Whyx-PROmpTea

---

# ART-PROMPtimizer

Prompt Enhancement Studio for Anima, Illustrious XL & NoobAI-XL models.

Turn your raw prompts into beautiful, detailed variations using curated Danbooru tag pools.

## Features

- **Smart Prompt Parsing** — extracts subject, character, series, artist from your input
- **LEGO-like Enhancement** — choose tag categories (quality, lighting, effects, etc.)
- **5–10 Variations** — slider-controlled with creativity level
- **Model-Optimized** — formats correctly for Anima, Illustrious XL and NoobAI-XL
- **Content Rating** — PG → R+ with appropriate filtering
- **6 Presets** — one-click cinematic, dreamy, vibrant styles
- **Trilingual** — English / Русский / 中文
- **Five themes** — Indigo (default), Emerald, Sunset, Mist, Cream — the same set the browser extension ships
- **Image Tagger** — smart analysis (pose, anatomy, expression, background, depth, palette) + WD14 ensemble (EVA02-Large + ConvNeXt + JoyTag)
- **Cloud VLM, your choice of who holds the key** — *private* mode calls the provider from your browser, so the API key and the image never reach this Space; only the finished caption comes back. *Server* mode proxies through the Space. Verified browser-capable: Gemini, OpenRouter (NVIDIA and local runtimes cannot be called from a page — no CORS).
- **Mature MOD (18+)** — one switch that lets the Cloud VLM describe adult content explicitly instead of in euphemisms (so NSFW tags can actually be lifted from the caption) and keeps those tags when you press Apply. Cloud VLM only, off by default.
- **Tools** — prompt analyzer with live suggestions, artist recommender, web tag search

## Browser extension

The Chrome/Edge/Firefox extension (hover any image → 🏷) is a frontend to this same backend: this Space serves `/api/v1/tag` and the rest of the routes it calls.

The two surfaces are one product and are meant to look alike:

- the palette is the same in both surfaces: `theme/space.css` for the Space, `extension/theme.css` for the extension, same five themes and same colour values. Both are hand-owned, and `tests/test_ui_assets.py` compares them so a palette edit cannot reach one product and miss the other;
- the Tagger names the same things the same way (Mode, Format, Smart analysis, NL caption, Control preview, Blacklist) — the key mapping table lives in `AGENTS.md`.

Build both flavors with `.venv\Scripts\python.exe scripts\build_extension.py`.

## Models Supported
- [Anima](https://huggingface.co/circlestone-labs/Anima) — circlestone-labs/Anima
- [Illustrious XL v1.0](https://huggingface.co/OnomaAIResearch/Illustrious-XL-v1.0) — OnomaAIResearch/Illustrious-XL-v1.0
- [NoobAI-XL](https://huggingface.co/Laxhar/noobai-XL-1.0) — Laxhar/noobai-XL-1.0

## Configuration

Optional behaviour is controlled via environment variables (all default to a safe/off state on CPU):

| Variable | Default | Purpose |
| --- | --- | --- |
| `ART_ENABLE_TAGGER` | `1` | Master switch for the image tagger (torch/timm required). |
| `ART_ENABLE_JOYTAG` | `1` | Load JoyTag alongside WD14 in ensemble mode. |
| `ART_POSE_TTA` | `0` | Test-time augmentation (flip) for pose estimation. `1` = more accurate, slower. |
| `ART_WB_DET_REPO` / `ART_WB_DET_FILE` | `yzd-v/DWPose` / `yolox_l.onnx` | Whole-body detector source. |
| `ART_WB_POSE_REPO` / `ART_WB_POSE_FILE` | `yzd-v/DWPose` / `dw-ll_ucoco_384.onnx` | Whole-body pose model source. |
| `ART_WB_SCORE_RESCALE` | `1.5` | DWPose confidence calibration (see `rescale_pose_score`); `1.0` disables. |
| `ART_DEPTH_REPO` / `ART_DEPTH_FALLBACK_REPO` | `onnx-community/depth-anything-v2-small` / `Xenova/depth-anything-small-hf` | Depth estimator (lazy, only when `depth` control is used). |
| `ART_SEG_REPO` | `nvidia/segformer-b0-finetuned-ade20k-512-512` | Segmentation model (lazy). |
| `ART_CAPTION_MODEL` | `florence` | Caption backbone: only `florence` (Florence-2-base) remains — all retired keys (`smolvlm*`, `moondream`, `qwen*`, `krea*`, `joy*`, `florence_large`, ...) auto-map to it. |
| `ART_FLORENCE_MAX_TOKENS` / `ART_FLORENCE_LONG_EDGE` | `128` / `768` | Florence decode budget / input resolution (floors 64 / 256). Per-request NL depth overrides: brief 96/640, standard 128/768, detailed 256/1024. |
| `ART_CLOUD_KEY` | _(empty)_ | Optional server-side key for the Cloud VLM caption mode (overridden per-request by the extension's key). Image + key leave the host — opt-in only. Not used by the Space's **private** cloud mode, where the call is made by the browser and the key never reaches this host. |
| `ART_CLOUD_BASE` / `ART_CLOUD_MODEL` | Gemini OpenAI-compat endpoint / `gemini-3.6-flash` | Cloud VLM endpoint (any OpenAI-compatible vision backend) and model. |
| `ART_CLOUD_TIMEOUT` / `ART_CLOUD_LONG_EDGE` | `120` / `1536` | Cloud request timeout (s) and upload longest edge (px, JPEG q85). |
| `ART_DISABLE_CAPTION` | `0` | `1` disables VLM captioning entirely. |
| `ART_TAGGER_DEVICE` | `cpu` | Inference device for WD14 taggers: `cpu` \| `cuda` \| `auto`. |
| `ART_DEBUG_POSE` | `0` | `1` enables verbose pose debug logging. |

See `CONTRIBUTING.md` for development setup and the full flag reference.

### Caption models & attribution

The Tagger's optional caption pass runs Florence-2-base (PyTorch) with
depth-dependent budgets; NL styles (general/detailed/flux/midjourney/sd/
structured/json) are shaped deterministically in `src/krea_composer.py`
from the factual caption + smart-analysis tags. Removed backends
(SmolVLM2, moondream2, Qwen-VL, InternVL3, PromptGen, JoyCaption) were
deleted with their weights and env knobs — their kinds auto-map to
`florence` for backward compatibility.