v1.0.10 – Qwen3.5 Support and Stability Improvements

Added local GGUF support for Qwen3.5 models

• Implemented proper Qwen3.5 handler routing with Qwen35ChatHandler
• Fixed incorrect fallback to Qwen3VLChatHandler for Qwen3.5 model names
• Improved mmproj handling for Qwen3.5 (requirement checks + auto-detection flow)

Added post-run cleanup() calls to VisionLLMNode, WanVideoPromptGenerator, and QwenImageEditPromptGenerator

• Refined cleanup lifecycle with cleanup(finalize=False/True) for regular unload vs final teardown

v1.0.9 - Local GGUF Discovery Expansion, mmproj Filtering, and Qwen/Wan Prompt Quality Improvements

Highlights

This release focuses on four major areas:

• Expanded local GGUF model discovery
• Safer and more intuitive mmproj selection
• Improved Qwen / Wan prompt rewriting quality
• Better robustness for text-only and image-to-video workflows

What Changed

Local GGUF model discovery

• Expanded local Qwen-family GGUF model search paths
• Added models/text_encoders and all subdirectories under both models/LLM and models/text_encoders to the search paths
• Centralized local model path and mmproj path resolution in local_gguf_utils.py
• Reduced duplicated path-handling logic across nodes

mmproj handling and model selection

• Added UI-side mmproj filtering so only mmproj files in the same directory as the selected model are shown
• Improved mmproj resolution behavior for local models
• Explicitly forces text-only mode when mmproj = (Not required) is selected
• Prevents unnecessary or incorrect Vision handler usage in text-only workflows
• Improves safety when switching between local model configurations

Prompt rewriting quality

• Added dedicated system prompt flows for:
  • qwen_image
  • qwen_image_edit
  • wan_t2v
  • wan_i2v
• Strengthened prompt instructions so outputs are more likely to contain only the final prompt body
• Reduced verbose analysis-style or heading-based outputs
• Added a second-pass Simplified Chinese normalization flow when Chinese output is requested but another language is returned
• Preserves quoted text during second-pass normalization to avoid breaking user-specified text
• Improved Qwen2.5-VL behavior in Qwen Image Edit Prompt Generator by fixing system prompt application issues

Node behavior and robustness

• Qwen-Image can now be used without an image input for text-only prompt generation
• Local text-only Qwen-Image runs no longer require mmproj
• Increased local inference max_tokens and n_ctx for longer prompt generation
• Added explicit validation for missing image input in Image-to-Video mode
• Improved output control for Wan prompt generation in Chinese-targeted workflows

Upgrade Notes

• After upgrading, you may need to reselect your GGUF model and mmproj file once
• This is because internal model path handling changed with the expanded search paths
• If you use local GGUF models, verify that the selected mmproj still matches the model directory
• Users of Qwen Image Edit Prompt Generator with Qwen2.5-VL should see improved output quality in this release

Notes

• Vision behavior still depends on the installed llama-cpp-python build and backend environment
• Some models may still show different output quality depending on whether they are used locally or via API
• Chinese-targeted prompt generation should now be more stable, but final output quality still depends on model behavior

v1.0.8 - Bug Fix

• Fixed issue where Qwen2.5-VL were always loaded in text-only mode even when a valid mmproj file was specified.
• Improved mmproj auto-detection logic."

v1.0.7 - Stability Update

• Fixed incorrect detection of Qwen3-VL when mmproj is set to (Not required).
  • Disabled automatic mmproj detection and prevented use of the VL handler in this case.
  • Updated GGUFModelManager.load_model and node-side mmproj interpretation to correctly respect (Not required).

v1.0.6 – Stability and Documentation Update

This release focuses on stability improvements and documentation cleanup ahead of the initial Comfy Registry publishing.

Changes

• Improved stability when switching between Qwen3-VL GGUF models
• Fixed mmproj reuse issues in local vision models
• Refined internal GGUF model lifecycle management
• Clarified project scope as a prompt generator for QwenImageEdit and Wan2.2
• Reorganized Credits and Dependencies for clearer attribution
• Updated llama-cpp-python installation notes to reference the JamePeng fork documentation

Notes

• No breaking changes to node interfaces
• This is the first version published to the Comfy Registry

v1.0.5 - Initial Release

v1.0.5 - Initial Release

Multimodal prompt generation nodes for ComfyUI with local Qwen-VL GGUF support.

What's Included

• Vision LLM Node - Local GGUF models (Qwen2.5-VL, Qwen3-VL) with multi-image input
• Qwen Image Edit - Image editing prompt optimization (local + cloud API)
• Wan Video Generator - T2V/I2V prompt enhancement for Wan2.2

Key Features

✅ Multi-image batch input support
✅ CPU/GPU device selection
✅ 5 style presets (raw, default, detailed, concise, creative)
✅ Auto-detect mmproj for Qwen3-VL
✅ Optimized for Chinese language prompts
✅ GPL-3.0 licensed with proper attribution

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/kantan-kanto/ComfyUI-MultiModal-Prompt-Nodes.git
cd ComfyUI-MultiModal-Prompt-Nodes
pip install -r requirements.txt

Full documentation: README.md