Release v1.0.10: add Qwen3.5 GGUF support, fix Qwen handler routing, and improve post-run cleanup across nodes

Author: kantan-kanto

.gitattributes

@@ -0,0 +1 @@
+* text=auto eol=lf

CHANGELOG.md

@@ -3,6 +3,19 @@
 All notable changes to ComfyUI-MultiModal-Prompt-Nodes will be documented in this file.
 
 
+## [1.0.10] - 2026-04-02
+
+- Added support for Qwen3.5 local GGUF models
+  - Added Qwen3.5 model detection and proper handler selection (`Qwen35ChatHandler`)
+  - Fixed incorrect fallback to `Qwen3VLChatHandler` for Qwen3.5 model names
+  - Updated mmproj handling for Qwen3.5 (requirement checks and auto-detection flow)
+
+- Improved post-run cleanup behavior for local model nodes
+  - `VisionLLMNode`, `WanVideoPromptGenerator`, and `QwenImageEditPromptGenerator` now call `cleanup()` at the end of execution
+  - Introduced `cleanup(finalize=False/True)` to separate regular unload from final teardown on process exit
+  - Added safe manager re-initialization after cleanup for stable repeated runs
+
+
 ## [1.0.9] - 2026-03-15
 
 - Expanded the search scope for local Qwen-family GGUF models

README.md

@@ -1,15 +1,15 @@
 # ComfyUI-MultiModal-Prompt-Nodes
 
-**Version:** 1.0.9
+**Version:** 1.0.10
 **License:** GPL-3.0
 
 Multimodal prompt generator nodes for ComfyUI, designed to generate prompts for **Qwen-Image-Edit** and **Wan2.2**.  
-Supports **local LLM / local GGUF models** (Qwen2.5-VL, Qwen3-VL) and **Qwen API** for image and video prompt generation and enhancement.
+Supports **local LLM / local GGUF models** (Qwen2.5-VL, Qwen3-VL and Qwen3.5) and **Qwen API** for image and video prompt generation and enhancement.
 
 ---
 ## Upgrade Notes for Existing Users
 
-The following notes are intended for existing users upgrading to `1.0.9`.
+The following notes are intended for existing users upgrading to `1.0.10`.
 
 ### Expanded search paths for local Qwen-family GGUF models
 In addition to `models/LLM`, this release now searches `models/text_encoders` and its subdirectories for GGUF files. Because this changes how model paths are handled internally, you may need to reselect your models the first time you run the node after updating.
@@ -48,7 +48,7 @@ Based on extensive testing, **Wan2.2** and **Qwen-Image-Edit** respond **signifi
   - `concise`: Minimal keywords, focused on core elements
   - `creative`: Artistic interpretation with unique perspectives
 - **Multi-image input**: Support batch image input via ComfyUI's batch nodes (e.g., Images Batch Multiple)
-- **Local GGUF support**: Run Qwen2.5-VL and Qwen3-VL models locally
+- **Local GGUF support**: Run Qwen2.5-VL, Qwen3-VL, and Qwen3.5 models locally
 - **Auto-detect mmproj**: Automatic detection or manual selection
 
 ### 2. Qwen Image Edit Prompt Generator
@@ -96,14 +96,15 @@ pip install dashscope pillow numpy
 
 **Important:** Model compatibility varies by llama-cpp-python version. Based on my testing environment:
 
-| Version | Qwen2.5-VL | Qwen3-VL | 
-|---------|------------|----------|
-| 0.3.16 (official) | ✅ | ❌ |
-| 0.3.21+ (JamePeng fork) | ✅ | ✅ |
+| Version | Qwen2.5-VL | Qwen3-VL | Qwen3.5 | 
+|---------|------------|----------|---------|
+| 0.3.16 (official) | ✅ | ❌ | ❌ |
+| 0.3.21+ (JamePeng fork) | ✅ | ✅ | ❌ |
+| 0.3.33+ (JamePeng fork) | ✅ | ✅ | ✅ | 
 
 ***Note:** Vision input support may vary depending on your environment and configuration.
 
-**Recommended Installation (JamePeng fork for Qwen3-VL support):**  
+**Recommended Installation (JamePeng fork for Qwen3-VL and Qwen3.5 support):**  
 Please follow the build and installation instructions provided in the JamePeng fork repository, as this fork requires a custom build and cannot be reliably installed via a simple `pip install`.
 
 **Source:** https://github.com/JamePeng/llama-cpp-python
@@ -168,7 +169,7 @@ Add your Alibaba Cloud Dashscope API key to this file.
 ### Qwen Image Edit Prompt Generator
 
 **Inputs:**
-- `image`: Primary input image (required)
+- `image`: Primary input image (optional)
 - `prompt`: Edit instruction or image description
 - `prompt_style`: 
   - `Qwen-Image-Edit`: For image editing tasks
@@ -245,14 +246,18 @@ Add your Alibaba Cloud Dashscope API key to this file.
 ## Model Compatibility
 
 ### Qwen2.5-VL (Separate mmproj)
-- ✅ Qwen2.5-VL(3B/7B): Full vision support
+- ✅ Qwen2.5-VL(3B/7B/32B): Full vision support
 - ✅ Requires matching mmproj file
 - ~~❌ Insufficient adherence to user prompts under the existing system prompt configuration with **Qwen-Image-Edit**~~
 
 ### Qwen3-VL (Separate mmproj)
 - ✅ Qwen3-VL(4B/8B): Full vision support with JamePeng fork
 - ✅ Requires matching mmproj file
 
+### Qwen3.5 (Separate mmproj)
+- ✅ Qwen3.5(9B/27B/35B-A3B): Full vision support with JamePeng fork
+- ✅ Requires matching mmproj file
+
 ### Model Sources
 - Qwen models: https://huggingface.co/Qwen
 - GGUF conversions: https://huggingface.co/models?search=qwen+gguf
@@ -447,9 +452,6 @@ Areas needing help:
 
 See [CHANGELOG.md](CHANGELOG.md) for detailed version history.
 
-### Current Version: 1.0.9
-- Expanded the search scope for local Qwen-family GGUF models
-- Improved mmproj selection behavior
-- Strengthened the local prompt rewrite flow for Qwen and Wan
-- Expanded Qwen Image Edit Prompt Generator
-- Improved the robustness of Wan Video Prompt Generator
+### Current Version: 1.0.10
+- Added support for Qwen3.5 local GGUF models
+- Improved post-run cleanup behavior for local model nodes

__init__.py

@@ -6,7 +6,7 @@
 # the Free Software Foundation, either version 3 of the License, or
 # (at your option) any later version.
 
-__version__ = "1.0.8"
+__version__ = "1.0.10"
 WEB_DIRECTORY = "./web"
 
 from .qwen_nodes import NODE_CLASS_MAPPINGS as qNODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS as qNODE_DISPLAY_NAME_MAPPINGS

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "multimodal-prompt-nodes"
 description = "Multimodal prompt generator nodes for ComfyUI, designed to generate prompts for QwenImageEdit and Wan2.2. Supports local LLM / local GGUF models (Qwen2.5-VL, Qwen3-VL) and Qwen API for image and video prompt generation and enhancement."
-version = "1.0.9"
+version = "1.0.10"
 license = {file = "LICENSE"} 
 # classifiers = [
 #     # For OS-independent nodes (works on all operating systems)

qwen_nodes.py

@@ -536,115 +536,122 @@ def INPUT_TYPES(s):
     DESCRIPTION = "Enhance your prompts using the Qwen LLM to align the behavior and capabilities of the Qwen-Image/Edit online version."
 
     def rewrit(self, prompt, prompt_style, target_language, llm_model, mmproj, max_retries, device, save_tokens, image=None, image2=None, image3=None):
-        # Collect all images
-        all_images = []
-        if image is not None:
-            all_images.extend(tensor2pil(image))
-        if image2 is not None:
-            all_images.extend(tensor2pil(image2))
-        if image3 is not None:
-            all_images.extend(tensor2pil(image3))
-        
-        # Local model processing
-        if llm_model.startswith("Local: "):
-            try:
-
-                model_filename = llm_model.replace("Local: ", "")
-                
-                # mmproj check
-                # vision_llm_node rewrite_prompt_with_gguf import
-                import sys
-                current_dir = os.path.dirname(os.path.abspath(__file__))
-                if current_dir not in sys.path:
-                    sys.path.insert(0, current_dir)
-                # Centralized import path handling
-                from import_utils import ensure_local_import
-                ensure_local_import(__file__)
-                from vision_llm_node import rewrite_prompt_with_gguf, resolve_local_gguf_path, resolve_mmproj_path_for_model
-                
-                # Model path retrieval
-                model_path = resolve_local_gguf_path(model_filename)
-                
-                # mmproj processing (same logic as Vision LLM Node)
-                if mmproj is None:
-                    raise RuntimeError("mmproj not specified. Please select an mmproj file in the optional inputs for Local models.")
+        try:
+            # Collect all images
+            all_images = []
+            if image is not None:
+                all_images.extend(tensor2pil(image))
+            if image2 is not None:
+                all_images.extend(tensor2pil(image2))
+            if image3 is not None:
+                all_images.extend(tensor2pil(image3))
+            
+            # Local model processing
+            if llm_model.startswith("Local: "):
+                try:
 
-                if prompt_style == "Qwen-Image" and len(all_images) == 0:
-                    mmproj_selection = "(Not required)"
-                else:
-                    mmproj_selection = mmproj
+                    model_filename = llm_model.replace("Local: ", "")
+                    
+                    # mmproj check
+                    # vision_llm_node rewrite_prompt_with_gguf import
+                    import sys
+                    current_dir = os.path.dirname(os.path.abspath(__file__))
+                    if current_dir not in sys.path:
+                        sys.path.insert(0, current_dir)
+                    # Centralized import path handling
+                    from import_utils import ensure_local_import
+                    ensure_local_import(__file__)
+                    from vision_llm_node import rewrite_prompt_with_gguf, resolve_local_gguf_path, resolve_mmproj_path_for_model
+                    
+                    # Model path retrieval
+                    model_path = resolve_local_gguf_path(model_filename)
+                    
+                    # mmproj processing (same logic as Vision LLM Node)
+                    if mmproj is None:
+                        raise RuntimeError("mmproj not specified. Please select an mmproj file in the optional inputs for Local models.")
 
-                mmproj_path = resolve_mmproj_path_for_model(model_path, mmproj_selection)
-                
-                print(f'[Qwen Prompt Rewriter] Using Local model')
-                print(f'[Qwen Prompt Rewriter] Model: {model_filename}')
-                print(f'[Qwen Prompt Rewriter] mmproj: {mmproj_selection}')
-                print(f'[Qwen Prompt Rewriter] Using {len(all_images)} image(s)')
-                
-                # Convert device selection to n_gpu_layers
-                n_gpu_layers = -1 if device == "GPU" else 0
-                
-                output_prompt = rewrite_prompt_with_gguf(
-                    prompt=prompt,
-                    model_path=model_path,
-                    mmproj_path=mmproj_path,
-                    style="qwen_image" if prompt_style == "Qwen-Image" else "qwen_image_edit",
-                    target_language=target_language,
-                    images=all_images,
-                    max_tokens=2048,
-                    temperature=0.7,
-                    n_ctx=4096,
-                    n_gpu_layers=n_gpu_layers,
-                )
+                    if prompt_style == "Qwen-Image" and len(all_images) == 0:
+                        mmproj_selection = "(Not required)"
+                    else:
+                        mmproj_selection = mmproj
 
-                if target_language == "zh" and not is_acceptable_zh_output(output_prompt):
-                    print('[Qwen Prompt Rewriter] Output language mismatch (expected simplified Chinese), converting output to simplified Chinese in a second pass')
-                    protected_text, placeholders = protect_quoted_text(output_prompt, "QTXT")
+                    mmproj_path = resolve_mmproj_path_for_model(model_path, mmproj_selection)
+                    
+                    print(f'[Qwen Prompt Rewriter] Using Local model')
+                    print(f'[Qwen Prompt Rewriter] Model: {model_filename}')
+                    print(f'[Qwen Prompt Rewriter] mmproj: {mmproj_selection}')
+                    print(f'[Qwen Prompt Rewriter] Using {len(all_images)} image(s)')
+                    
+                    # Convert device selection to n_gpu_layers
+                    n_gpu_layers = -1 if device == "GPU" else 0
+                    
                     output_prompt = rewrite_prompt_with_gguf(
-                        prompt=build_force_translate_to_zh_prompt(protected_text, prompt_style),
+                        prompt=prompt,
                         model_path=model_path,
-                        mmproj_path="(Not required)",
-                        style="zh_normalize",
-                        target_language="zh",
-                        images=None,
+                        mmproj_path=mmproj_path,
+                        style="qwen_image" if prompt_style == "Qwen-Image" else "qwen_image_edit",
+                        target_language=target_language,
+                        images=all_images,
                         max_tokens=2048,
-                        temperature=0.2,
+                        temperature=0.7,
                         n_ctx=4096,
                         n_gpu_layers=n_gpu_layers,
                     )
-                    output_prompt = restore_quoted_text(output_prompt, placeholders)
-                 
-            except Exception as e:
-                raise RuntimeError(f"Local model error: {str(e)}")
-        
-        # API processing (cloud models)
-        else:
-            # Load API key from api_key.txt
-            if not os.path.exists(key_path):
-                raise EnvironmentError(f"API key file not found: {key_path}\nPlease create this file with your Aliyun API key for cloud model usage.")
-            
-            with open(key_path, "r", encoding="utf-8") as f:
-                _api_key = f.read().strip()
+
+                    if target_language == "zh" and not is_acceptable_zh_output(output_prompt):
+                        print('[Qwen Prompt Rewriter] Output language mismatch (expected simplified Chinese), converting output to simplified Chinese in a second pass')
+                        protected_text, placeholders = protect_quoted_text(output_prompt, "QTXT")
+                        output_prompt = rewrite_prompt_with_gguf(
+                            prompt=build_force_translate_to_zh_prompt(protected_text, prompt_style),
+                            model_path=model_path,
+                            mmproj_path="(Not required)",
+                            style="zh_normalize",
+                            target_language="zh",
+                            images=None,
+                            max_tokens=2048,
+                            temperature=0.2,
+                            n_ctx=4096,
+                            n_gpu_layers=n_gpu_layers,
+                        )
+                        output_prompt = restore_quoted_text(output_prompt, placeholders)
 
-            if not _api_key:
-                raise EnvironmentError(f'API_KEY is not set in "{key_path}"\nPlease add your Aliyun API key to this file for cloud model usage.')
+                except Exception as e:
+                    raise RuntimeError(f"Local model error: {str(e)}")
 
-            if prompt_style == "Qwen-Image":
-                output_prompt = polish_prompt(_api_key, prompt, model=llm_model, max_retries=max_retries, target_language=target_language)
+            # API processing (cloud models)
             else:
-                # Qwen-Image-Edit requires at least one image
-                if len(all_images) == 0:
-                    raise ValueError("Qwen-Image-Edit style requires at least one image input!")
+                # Load API key from api_key.txt
+                if not os.path.exists(key_path):
+                    raise EnvironmentError(f"API key file not found: {key_path}\nPlease create this file with your Aliyun API key for cloud model usage.")
 
-                print(f'[Qwen Prompt Rewriter] Using {len(all_images)} image(s) for Image-Edit')
-                output_prompt = polish_prompt_edit(_api_key, prompt, all_images, model=llm_model, max_retries=max_retries, save_tokens=save_tokens, target_language=target_language)
-        
-        print(f'[Qwen Prompt Rewriter] Style: {prompt_style}')
-        print(f'[Qwen Prompt Rewriter] Target Language: {target_language}')
-        print(f'[Qwen Prompt Rewriter] Original: "{prompt}"')
-        print(f'[Qwen Prompt Rewriter] Enhanced: "{output_prompt}"')
-        
-        return (output_prompt,)
+                with open(key_path, "r", encoding="utf-8") as f:
+                    _api_key = f.read().strip()
+                        
+                if not _api_key:
+                    raise EnvironmentError(f'API_KEY is not set in "{key_path}"\nPlease add your Aliyun API key to this file for cloud model usage.')
+                
+                if prompt_style == "Qwen-Image":
+                    output_prompt = polish_prompt(_api_key, prompt, model=llm_model, max_retries=max_retries, target_language=target_language)
+                else:
+                    # Qwen-Image-Edit requires at least one image
+                    if len(all_images) == 0:
+                        raise ValueError("Qwen-Image-Edit style requires at least one image input!")
+                    
+                    print(f'[Qwen Prompt Rewriter] Using {len(all_images)} image(s) for Image-Edit')
+                    output_prompt = polish_prompt_edit(_api_key, prompt, all_images, model=llm_model, max_retries=max_retries, save_tokens=save_tokens, target_language=target_language)
+            
+            print(f'[Qwen Prompt Rewriter] Style: {prompt_style}')
+            print(f'[Qwen Prompt Rewriter] Target Language: {target_language}')
+            print(f'[Qwen Prompt Rewriter] Original: "{prompt}"')
+            print(f'[Qwen Prompt Rewriter] Enhanced: "{output_prompt}"')
+            
+            return (output_prompt,)
+        finally:
+            try:
+                from vision_llm_node import cleanup as vision_cleanup
+                vision_cleanup()
+            except Exception:
+                pass
 
 NODE_CLASS_MAPPINGS = {
     "QwenImageEditPromptGenerator": QwenImageEditPromptGenerator