{ "version": "1.0.0", "last_updated": "2025-12-14T00:00:00Z", "description": "Model-specific tips for vision LLMs used in prompt generation", "vision_models": { "Qwen2.5-VL-7B-Instruct": { "provider": "local", "parameter_count": "7B", "description": "Alibaba's Qwen2.5 Vision-Language model", "strengths": [ "Human detection and pose analysis", "Color accuracy", "Scene composition understanding", "Good instruction following", "Efficient at 7B size" ], "weaknesses": [ "Small accessories may be missed", "Fine patterns can be overlooked", "Fabric textures sometimes vague", "May miss jewelry details" ], "optimal_settings": { "temperature": 0.2, "quantization": "8-bit", "repetition_penalty": 1.2, "top_p": 0.9 }, "prompt_tips": [ "Explicitly ask for accessory details", "Request fabric material description", "Ask for pattern details separately if needed", "Use structured output requests" ], "known_issues": [ "At 4-bit quantization, detail detection reduces", "May generate verbose output without length guidance" ] }, "Qwen2-VL-72B-Instruct": { "provider": "local", "parameter_count": "72B", "description": "Large Qwen2 Vision-Language model", "strengths": [ "Excellent detail recognition", "Strong reasoning capabilities", "Better accessory detection than 7B", "Good at complex scenes" ], "weaknesses": [ "High VRAM requirement", "Slower inference" ], "optimal_settings": { "temperature": 0.3, "quantization": "4-bit", "repetition_penalty": 1.1 }, "prompt_tips": [ "Can handle complex multi-part requests", "Good for detailed analysis", "Responds well to structured prompts" ] }, "LLaVA-1.6-34B": { "provider": "local", "parameter_count": "34B", "description": "Large Language and Vision Assistant", "strengths": [ "Detailed descriptions", "Scene understanding", "Good conversation ability", "Strong at complex queries" ], "weaknesses": [ "Can be verbose", "Occasional hallucination", "High VRAM usage" ], "optimal_settings": { "temperature": 0.3, "repetition_penalty": 1.3, "max_tokens": 1024 }, "prompt_tips": [ "Use structured output requests", "Set clear length limits", "Be specific about what to include/exclude", "Request bullet points for structured output" ], "known_issues": [ "May add narrative commentary", "Can repeat information" ] }, "LLaVA-1.5-7B": { "provider": "local", "parameter_count": "7B", "description": "Efficient LLaVA variant", "strengths": [ "Fast inference", "Low VRAM usage", "Good general understanding" ], "weaknesses": [ "Less detailed than larger models", "May miss subtle elements" ], "optimal_settings": { "temperature": 0.3, "repetition_penalty": 1.2 }, "prompt_tips": [ "Keep requests focused", "Ask about specific elements", "Good for quick analysis" ] }, "Florence-2-large": { "provider": "local", "parameter_count": "0.7B", "description": "Microsoft's Florence-2 vision model", "strengths": [ "Very fast", "Low VRAM", "Good at detection tasks", "Structured output" ], "weaknesses": [ "Limited descriptive ability", "Best for specific tasks", "Less conversational" ], "optimal_settings": { "temperature": 0.2 }, "prompt_tips": [ "Use task-specific prompts", "Good for object detection", "Combine with other models for full descriptions" ] }, "Moondream2": { "provider": "local", "parameter_count": "1.6B", "description": "Compact vision-language model", "strengths": [ "Extremely efficient", "Fast inference", "Can run on CPU" ], "weaknesses": [ "Limited detail", "Basic descriptions", "May miss nuances" ], "optimal_settings": { "temperature": 0.3 }, "prompt_tips": [ "Keep questions simple", "Good for basic image understanding", "Use for quick checks" ] }, "InternVL2-8B": { "provider": "local", "parameter_count": "8B", "description": "OpenGVLab's InternVL2 model", "strengths": [ "Good balance of size/quality", "Strong OCR ability", "Multi-image support" ], "weaknesses": [ "May be verbose" ], "optimal_settings": { "temperature": 0.3, "repetition_penalty": 1.2 }, "prompt_tips": [ "Good for text in images", "Can compare multiple images" ] }, "Phi-3.5-Vision": { "provider": "local", "parameter_count": "4.2B", "description": "Microsoft Phi-3.5 with vision", "strengths": [ "Efficient size", "Good reasoning", "Fast inference" ], "weaknesses": [ "Less detailed than larger models" ], "optimal_settings": { "temperature": 0.3 }, "prompt_tips": [ "Good for reasoning tasks", "Efficient for multiple queries" ] }, "claude-sonnet-4-5-20250929": { "provider": "anthropic", "parameter_count": "unknown", "description": "Anthropic Claude Sonnet 4.5", "strengths": [ "Excellent accuracy", "Strong instruction following", "Good structured output", "Reliable and consistent", "Handles complex prompts" ], "weaknesses": [ "API cost", "May be conservative in descriptions" ], "optimal_settings": { "temperature": 0.3 }, "prompt_tips": [ "Can handle complex multi-part prompts", "Excellent with JSON output", "Good at following specific formats", "Reliable for production use" ] }, "claude-opus-4-5-20251101": { "provider": "anthropic", "parameter_count": "unknown", "description": "Anthropic Claude Opus 4.5 (most capable)", "strengths": [ "Highest accuracy", "Best reasoning", "Excellent detail detection", "Strong at nuanced descriptions" ], "weaknesses": [ "Higher API cost", "Slower than Sonnet" ], "optimal_settings": { "temperature": 0.3 }, "prompt_tips": [ "Best for evaluation and analysis", "Excellent for complex tasks", "Use for quality-critical applications" ] }, "gpt-4o": { "provider": "openai", "parameter_count": "unknown", "description": "OpenAI GPT-4o multimodal", "strengths": [ "Strong general ability", "Good at creative descriptions", "Handles ambiguity well" ], "weaknesses": [ "API cost", "May add interpretation" ], "optimal_settings": { "temperature": 0.3 }, "prompt_tips": [ "Clear instructions help", "Good at following examples", "Can be creative" ] }, "gemini-2.0-flash": { "provider": "google", "parameter_count": "unknown", "description": "Google Gemini 2.0 Flash", "strengths": [ "Fast inference", "Good multimodal ability", "Competitive pricing" ], "weaknesses": [ "May be less detailed" ], "optimal_settings": { "temperature": 0.3 }, "prompt_tips": [ "Good for quick analysis", "Handles multiple images" ] }, "llava:latest": { "provider": "ollama", "parameter_count": "7B", "description": "Ollama LLaVA model", "strengths": [ "Easy local deployment", "No API costs", "Good general ability" ], "weaknesses": [ "Depends on Ollama setup", "May vary by version" ], "optimal_settings": { "temperature": 0.3 }, "prompt_tips": [ "Good for local development", "Test with your specific version" ] } }, "provider_comparison": { "accuracy": { "ranking": [ "claude-opus-4-5-20251101", "claude-sonnet-4-5-20250929", "gpt-4o", "Qwen2-VL-72B-Instruct", "LLaVA-1.6-34B", "Qwen2.5-VL-7B-Instruct", "InternVL2-8B" ] }, "speed": { "ranking": [ "Florence-2-large", "Moondream2", "Phi-3.5-Vision", "gemini-2.0-flash", "Qwen2.5-VL-7B-Instruct", "LLaVA-1.5-7B" ] }, "cost_effectiveness": { "ranking": [ "Local models (free)", "gemini-2.0-flash", "gpt-4o-mini", "claude-sonnet-4-5-20250929", "gpt-4o", "claude-opus-4-5-20251101" ] } }, "quantization_effects": { "4-bit": { "vram_reduction": "~75%", "quality_impact": "Noticeable on fine details", "speed_impact": "Faster due to memory efficiency", "recommended_for": "Limited VRAM, quick iterations" }, "8-bit": { "vram_reduction": "~50%", "quality_impact": "Minimal quality loss", "speed_impact": "Slightly faster than FP16", "recommended_for": "Best balance of quality/efficiency" }, "fp16": { "vram_reduction": "0%", "quality_impact": "Full model quality", "speed_impact": "Baseline", "recommended_for": "Quality-critical applications" } } }