slahiri-ComfyUI-AI-Photography-Toolkit/CHANGELOG.md

Changelog

All notable changes to ComfyUI-AI-Photography-Toolkit will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[4.2.0] - 2024-12-11

Added

• SID_LLM_API - Unified cloud LLM node supporting 15+ providers:
  • Anthropic (Claude 4.x Sonnet, Opus, Haiku)
  • OpenAI (GPT-4o, GPT-4o-mini, o1, o3)
  • Google Gemini (2.0, 1.5 Pro, Flash)
  • xAI Grok (grok-2-vision, grok-vision-beta)
  • Mistral (pixtral-large, pixtral-12b)
  • DeepSeek, Cohere, Together AI, Fireworks, Groq, Perplexity
  • Ollama, LM Studio (local)
  • OpenRouter, Custom endpoints
• SID_LLM_Local - Unified local model node supporting:
  • Florence-2 Large (~4GB VRAM)
  • Moondream2 (~4GB VRAM)
  • SmolVLM (~4GB VRAM)
  • Phi-3.5 Vision (~8GB VRAM)
  • QwenVL 2B (~6GB VRAM)
• SID_ZImagePromptGenerator - Unified prompt generator with:
  • Auto pipeline selection (Single-Shot vs Agentic based on LLM capability)
  • Analysis modes: Quick, Standard, Detailed, Extreme
  • Preset styles: Auto-Detect, Portrait, Fashion, Artistic, NSFW
  • User guidance input for custom instructions
• Max tokens preset dropdown (Short/Medium/Long/Very Long/Maximum) replacing slider
• Enable reasoning toggle for extended thinking support
• Auto quantization for local models (4-bit/8-bit/None/Auto)

Changed

• Consolidated from 8+ nodes to 3 unified nodes
• Simplified workflow: LLM Provider → Prompt Generator → Output
• Provider-specific model dropdowns in SID_LLM_API
• Auto-detection of local vs cloud providers for API key requirements

Deprecated

The following nodes have been removed and replaced:

Deprecated Node	Replacement
SID_ZImagePromptGenerator (basic)	SID_ZImagePromptGenerator (unified)
SID_ZImagePromptGenerator_Advanced	SID_ZImagePromptGenerator (unified)
SID_ZImagePromptGenerator_Advanced_V2	SID_ZImagePromptGenerator (unified)
SID_Anthropic_LLM	SID_LLM_API (provider: Anthropic)
SID_OpenAI_Compatible_LLM	SID_LLM_API (provider: OpenAI/Custom)
SID_Grok_LLM	SID_LLM_API (provider: xAI Grok)
SID_GGUF_LLM	SID_LLM_Local
SID_QwenVL_LLM	SID_LLM_Local (model: QwenVL)

Migration

To migrate from v4.1.x:

1. Replace individual LLM nodes (SID_Anthropic_LLM, etc.) with SID_LLM_API or SID_LLM_Local
2. Replace any prompt generator with SID_ZImagePromptGenerator
3. Reconnect image and LLM model inputs

[4.1.0] - 2024-12-10

Added

• High-resolution GGUF models - all local models now support 1024x1024+ images natively:
  • Qwen3-VL 2B Q4 (4GB VRAM) - fast, dynamic resolution
  • Qwen2.5-VL 7B Q4/Q8 (6-10GB VRAM) - excellent quality, dynamic resolution
  • Llama 3.2 Vision 11B Q4 (8GB VRAM) - Meta's latest, 1120x1120 native
  • Pixtral 12B Q4 (10GB VRAM) - Mistral's vision model, 1024x1024 native
  • MiniCPM-V 2.6 (10GB VRAM) - multilingual, dynamic resolution
• Max Image Size option for GGUF provider - resize images before encoding for faster processing:
  • 512 (default): ~4x faster encoding
  • 768: ~2x faster encoding
  • 1024: ~1.5x faster encoding
  • Original: no resize (full quality, slowest)
• max_image_resolution field in GGUF model metadata for documentation
• Image caching infrastructure - tracks image hashes for future optimization

Changed

• Removed max_tokens slider from all LLM provider nodes (Anthropic, OpenAI, Grok, GGUF)
  • Now uses model's max_output_tokens from metadata directly
  • Eliminates the hardcoded 8192 cap that was limiting some models
• Updated GGUF model list - removed LLaVA 1.5/1.6 models (limited to 336x336/672x672 per tile)
• Default GGUF model changed from Moondream to Qwen2.5-VL 7B Q4
• Simplified GGUF node - removed quality_mode preset selector

Removed

• LLaVA 1.5 7B, 13B models (can't handle 1024x1024 images)
• LLaVA 1.6 Mistral 7B, 34B models (can't handle 1024x1024 images)
• Moondream2 model (replaced with higher resolution alternatives)

[4.0.1] - 2024-12-10

Added

• Read-only prompt display on V2 generator node - shows generated prompt after execution
• Max output tokens metadata for all LLM providers with model-specific limits:
  • Anthropic: Claude Sonnet/Opus 4.x (64K), Claude 3.5 (8K)
  • OpenAI: GPT-4o (16K), o1 series (32K-100K)
  • Grok: grok-2-vision (32K), grok-vision-beta (8K)
• Max tokens guardrail - automatically caps user's max_tokens to model's limit
• User prompt integration into analysis pipeline (both agentic and iterative modes):
  • Component analysis now receives user requirements as priority instructions
  • Agentic mode includes "CRITICAL: USER'S SPECIAL REQUEST" section
  • Final prompt shows user focus with [USER FOCUS: ...] prefix
• Logging for user requests in both iterative and agentic pipelines

Changed

• _analyze_component() now accepts optional user_prompt parameter
• _build_agentic_prompt() includes user requirements as high-priority analysis instructions
• _assemble_prompt() structures user requirements prominently at prompt start
• LLM provider nodes now store model_max_output_tokens in extra_params

Technical

• Added get_max_output_tokens() class method to all LLM providers
• V2 generator validates max_tokens against model limits at execution time
• is_output_node=True added to V2 schema for UI text display

[4.3.0] - 2024-12-09

Added

• High-resolution prompt generation for near-accurate image reproduction from source
• Detailed subject analysis with ethnicity detection and heritage identification
• Precise skin tone extraction including undertones (warm/cool/neutral) and surface quality
• Comprehensive facial feature analysis:
  • Eyes: color, shape, gaze direction, expression, eyelid position
  • Eyebrows: shape, thickness, grooming, arch, color
  • Nose: shape, bridge, tip, nostril visibility, proportions
  • Lips: shape, upper/lower lip ratio, natural color, position
• Detailed pose analysis for all body parts:
  • Head and neck positioning with tilt/rotation
  • Shoulder alignment and arm positions
  • Hand positions with palm orientation and finger details
  • Torso, hip, and leg positioning
  • Feet position and weight distribution
• Clothing coverage analysis:
  • Neckline type and depth
  • Cleavage and sideboob visibility levels
  • Back and midriff exposure
  • Leg and thigh visibility
  • Gaps, slits, and fabric behavior
• Cosmetics/makeup extraction:
  • Foundation and skin coverage
  • Eye makeup (shadow, liner, mascara)
  • Lip color and finish
  • Blush and contouring
• Structured JSON output with categorized data sections:
  • metadata (generator, version, timestamp, processing time)
  • model (provider, model name, template, settings)
  • settings (detail level, content detail, focus options)
  • image (dimensions, aspect ratio, color info)
  • classification, subject, cosmetics, pose, clothing
  • environment, lighting, colors, style
  • generated_prompt with word/token counts
  • zimage_recommendations
• LLM interaction tracking - all raw queries and responses included in output for debugging
• Token-optimized prompts that respect max token limits efficiently

Changed

• Enhanced Claude template with forensic-level image analysis prompts
• Enhanced Local GGUF template with multi-iteration Deep mode (6 focused analysis stages)
• Improved compound color descriptors for accuracy (e.g., "warm honey-golden blonde")

[4.2.0] - 2024-11-XX

Added

• High-quality GGUF models: LLaVA 1.6 34B, Qwen2.5-VL 7B, LLaVA 1.6 Mistral 7B
• Support for Qwen2-VL chat format

Fixed

• Moondream2 download URLs (correct HuggingFace repo)
• GGUF provider API key validation error

[4.1.0] - 2024-11-XX

Added

• SID_OpenAI_Compatible_LLM node - supports OpenAI, Together AI, LM Studio, Ollama
• SID_Grok_LLM node - xAI Grok vision models
• SID_GGUF_LLM node - local GGUF models with auto-download
• Auto-install llama-cpp-python with GPU detection (NVIDIA CUDA, Apple Metal, AMD ROCm, CPU fallback)
• Split into Basic (SID_ZImagePromptGenerator) and Advanced (SID_ZImagePromptGenerator_Advanced) nodes
• Welcome message shows dependency installation status on load

[4.0.0] - 2024-11-XX

Added

• Initial Z-Image optimized prompt generator
• 6-stage agentic pipeline for comprehensive image analysis
• Anthropic Claude API support
• Persistent disk caching for repeated analysis
• Multiple detail levels: Quick (1 LLM call), Standard (2 calls), Deep (3 calls)
• Focus options: subject, environment, lighting, colors, mood
• Content detail levels: minimal, standard, detailed, explicit
• Prompt modes: Image Only, Prompt Guides, Prompt First, Prompt Dominates

Technical

• SOLID principles architecture with template pattern
• BasePromptTemplate abstraction for provider-specific prompts
• Factory pattern for template selection
• ComfyUI V3 API compatibility