SuperMate-Harness-System

deepseek-eyes

Theme: Image analysis for text AI models — Give Text Models Eyes in DeepSeek Harness

DeepSeek Harness (DSH) is the open-source agent harness by DeepSeek AI — everything is a plugin. The text model inside DSH — DeepSeek by default — can only read text. It cannot see images. That is its biggest capability gap.

This skill is a DSH Skill that fills exactly that gap: it gives DeepSeek inside DSH a pair of eyes.

Image → vision model converts it to text → DeepSeek "sees"

With this skill installed, whenever the DSH agent encounters an image (screenshot, photo, illustration, character sheet, AI-generated image...), it automatically "opens its eyes": a local vision model translates the image into structured text description + retrieval tags, which are injected into DeepSeek's context — DeepSeek can then understand, reason, and answer based on the image content.

The text model in DSH is not limited to DeepSeek — it can be any other text model (including locally deployed ones). The "eyes" are universal: whoever uses them gains vision.

Install in DeepSeek Harness

Put the deepseek-eyes folder into DSH's skill directory:

~/.dsh/skills/deepseek-eyes/

Restart the DSH session and the skill is picked up and enabled automatically.

Use in DeepSeek Harness

Automatic mode (recommended): the DSH agent "opens its eyes" automatically in these tasks — no manual steps:

• Asset entry (turn images into text description + tags for archiving)
• Prompt management (view images to write prompts / manage image assets)
• Screenshot understanding (analyze images you paste into the DSH conversation)
• Image comparison (let DeepSeek compare the content of two images)

Manual mode: run the script in your DSH environment and paste the output into the conversation:

# Local "eyes" (Ollama + vision model, default llava:13b)
.\deepseek-eyes.ps1 -ImagePath "D:\pics\photo.webp"

# Cloud "eyes" (OpenAI-compatible vision API, e.g. gpt-4o / qwen-vl)
$env:VISION_API_KEY = "sk-..."
.\deepseek-eyes.ps1 -ImagePath "D:\pics\photo.png" -Provider openai -Model "gpt-4o-mini"

Example output (injected directly into DeepSeek's context):

This is a character design sheet... (structured description)
Tags: HARUMIN, character-sheet, orange-striped top, silver band-aid, short-haired girl, voice description

The script preprocesses automatically: converts any format (WebP, etc.) to PNG and downscales the long edge to 1080px — feeding a huge image raw makes the model slow and the output blurry (token explosion); after downscaling, the "eyes" read clearer, faster, and cheaper.

Configuration (effective in the DSH environment)

Environment variables (VISION_MODEL / VISION_ENDPOINT / VISION_API_KEY / VISION_PROVIDER) can be set once in the DSH launch environment.

How good are these "eyes"? (DSH local vision partner · RTX 5080 16GB, measured)

• ~5-10s per image, producing 300-700 tokens of description;
• Local vision front-end Qwen3.6-35B-IQ3_S: 100% GPU, 118 tok/s (the only full-speed option);
• Preprocessing gain: large image tokens drop from 3647 to 713 (-80%), 24x faster, no perceptible quality loss.

Prerequisites (either)

1. Local vision model: install Ollama and ollama pull llava:13b (or any vision model);
2. OpenAI-compatible vision API: any provider + API key.

---

Implementation details, full parameters and FAQ: see SKILL.md and deepseek-eyes.ps1.