datdang-dev/studio/services/erotic-image-captioner

Erotic Image Captioner Pipeline (Unified Caption Engine)

Overview

The Erotic Image Captioner (EC) has been refactored from a multi-agent manual prompting sequence and complex MCP tool structure into a unified, high-performance local Caption Engine (studio/core/caption_engine/caption_engine.py).

Instead of prompting multiple text-only agents (Kana, Luna, Suki) in-session with large system files or invoking fragmented MCP tools, LND Studio now leverages a single optimized multimodal execution pass using the Qwen2-VL local vision engine. Prompts reside as independent templates in studio/prompts/ and are loaded dynamically on-demand based on the specified --type.

This architecture ensures:

1. Low VRAM footprint: Visual inference runs in clean, dedicated local processes without multiple client-server context duplications.
2. Beautiful Observability: Color-coded, dashboard-style execution logging with real-time terminal output and log files (_lnd-output/caption_engine.log) suitable for tail -f.
3. Toggleable Streaming: CLI option (--no-stream) to switch between instant single-shot generation and real-time chunk streaming.

Unified Architecture:

[User Image] ──> [CaptionEngine (caption_engine.py)]
                        │
                        ├──► Loads: studio/prompts/{type}.txt (e.g., long_thoughts_v2, manga, etc.)
                        ├──► Resolves base64/local files
                        └──► Executes local InProcessQwen2VLAdapter

---

🛠️ CAPTION ENGINE CLI SET

The engine can be invoked directly from the CLI or via automated workflows using the following syntax:

python3 studio/core/caption_engine/caption_engine.py \
    --image "/path/to/image.png" \
    --type "long_thoughts_v2" \
    --mood "AUTO" \
    --temp 0.5 \
    --max-tokens 4096

Parameter Map:

• --image: Absolute path to the target image file (mandatory).
• --type: Prompt style from studio/prompts/ (e.g., long_thoughts_v2, manga, json, md_comic).
• --mood: Mood setting seeds (e.g., AUTO, MANIC, COLD, BRATTY, BROKEN, MASO).
• --context: Optional backstory or descriptive user context to inject.
• --temp: Generation temperature (default: 0.5).
• --max-tokens: Max output generation length (default: 4096).
• --no-stream: Disables real-time streaming, running in clean single-shot mode.

---

📋 PIPELINE EXECUTION STEPS

When executing the pipeline, the active agent (typically Nova) or the orchestrator invokes the Caption Engine directly:

Step 1 — Ingest & Validate

1. Verify the input image path exists.
2. Read the mood_seed parameter (default: AUTO).
3. Capture optional user_context (backstory/constraints).

Step 2 — Direct Multimodal Generation

1. The engine dynamically compiles the prompt using StudioPromptLoader from the target template in studio/prompts/{type}.txt.
2. The engine initializes the local in-process model wrapper (InProcessQwen2VLAdapter) referencing models/ToriiGate-0.5_Q4_K_L.gguf and mmproj_Q8_0.gguf.
3. Inference with Streaming/Single-Shot: Depending on --no-stream, tokens are either streamed back in real-time with an ANSI colored dashboard or processed in a single fast block.
4. Output is verified and persisted in _lnd-output/_captions/ as well as color-logged in _lnd-output/caption_engine.log.

---

📦 OUTPUT DIRECTORY STRUCTURE

For each processed image, the pipeline writes:

_lnd-output/_captions/
└── {image_name}_{timestamp}_{type}.md   # Output markdown/JSON formatted report

And appends live execution monitoring data to:

_lnd-output/caption_engine.log           # Structured color logs (viewable with tail -f)

---

🚀 ADVANTAGES OF THE NEW UNIFIED CONCEPT

1. Deterministic Environment: Bypasses unstable fastmcp/client-server connections. The execution is in-process and fast.
2. Flexible Templates: Switch prompt engineering templates on-the-fly by dropping new txt/md files into studio/prompts/ and calling --type {filename}.
3. Production Observability: Full execution telemetry (token count, generation speed, temperature, and visual parameters) are beautiful, colored, and persisted.