Files
1038lab-ComfyUI-MiniCPM/update.md
T
2025-08-27 10:50:43 -07:00

2.8 KiB

ComfyUI-MiniCPM Update Log

V1.1.0 (2025-08-27)

New Features

MiniCPM v4 VS v45

  • MiniCPM-V-4.5 Model Support: Added comprehensive support for the latest MiniCPM-V-4.5 models
    • New Transformers model: MiniCPM-V-4.5 (openbmb/MiniCPM-V-4_5) - Latest full precision version
    • New quantized model: MiniCPM-V-4.5-int4 (openbmb/MiniCPM-V-4_5-int4) - Memory-efficient 4-bit version
  • Performance Improvements: Added hf_xet dependency for faster downloads
  • Model Priority: V4.5 models now appear first in dropdown menus

Changes

  • V4.5 GGUF models temporarily disabled due to llama-cpp-python 3.16 compatibility
  • Enhanced error messages for unsupported models
  • Improved model ordering and descriptions

Usage Notes

  • Recommended: Use MiniCPM-V-4.5 Transformers models for latest features and enhanced capabilities
  • Alternative: Continue using MiniCPM-V-4.0 GGUF models (all quantizations available, see table above)
  • Coming Soon: V4.5 GGUF support will be restored once llama-cpp-python compatibility is resolved

V1.0.0 (2025-08-25)

Initial Release

  • MiniCPM-V-4 Transformers Support: Full integration with Hugging Face transformers models MiniCPM-V-4
    • MiniCPM-V-4 full precision model
    • MiniCPM-V-4-int4 quantized model
  • MiniCPM-V-4 GGUF Support: Comprehensive GGUF model support via llama-cpp-python MiniCPM-V-4-GGUF
    • Multiple quantization levels (Q4_0 to Q8_0)
    • Optimized memory usage and performance
  • Dual Node Architecture:
    • Basic nodes for simple usage
    • Advanced nodes with full parameter control
  • Image & Video Processing: Support for both image and video input processing MiniCPM-V-4_video
  • Memory Management: Three modes - Keep in Memory, Clear After Run, Global Cache
  • Caption Types: 14 different preset prompt types for various use cases
  • Legacy Compatibility: Backward compatible node for existing workflows

MiniCPM-V-4_batchImages

Core Features

  • Multi-Modal Processing: Image and video content understanding
  • Flexible Parameters: Customizable temperature, top-p, top-k, repetition penalty
  • Auto-Download: Automatic model downloading from Hugging Face
  • Cross-Platform: Windows, macOS, and Linux support
  • CUDA Acceleration: GPU acceleration support for faster inference
  • System Prompts: Customizable system prompts for different use cases