--- title: Qwen_Image_2.1_Workflows canonical_url: "https://www.modelscope.ai/models/zuanfilm/Qwen_Image_2.1_Workflows" md_url: "https://www.modelscope.ai/models/zuanfilm/Qwen_Image_2.1_Workflows.md" repository: zuanfilm/Qwen_Image_2.1_Workflows last_updated: 2026-09-21 pipeline_tag: text-to-image-synthesis tasks: - text-to-image-synthesis downloads: 0 stars: 0 --- # Qwen_Image_2.1_Workflows > Qwen_Image_2.1_Workflows - An open-source model by zuanfilm on ModelScope. Qwen Image 2.1 Edit Style Transfer is a ComfyUI workflow for image-to-image style transfer using Qwen Image 2.1. It uses a target image and a style reference image, with LCS-based… - **Repository**: zuanfilm/Qwen_Image_2.1_Workflows - **Tasks**: text-to-image-synthesis - **Downloads**: 0 - **Stars**: 0 - **Last updated**: 2026-09-21 Source: https://www.modelscope.ai/models/zuanfilm/Qwen_Image_2.1_Workflows --- # Qwen Image 2.1 Edit --- Style Transfer Workflow for ComfyUI A ComfyUI workflow for **Qwen Image 2.1 Edit** designed for image-to-image style transfer. The workflow uses two images: - **Image #1** --- the edit target - **Image #2** --- the style reference The default instruction is: > Take the style of `` and apply it to `` This workflow is designed to transfer visual characteristics from a reference image while maintaining the main content and composition of the target image. ------------------------------------------------------------------------ ## Example The included comparison demonstrates the workflow using: - **Image #1:** photographic source - **Image #2:** stylized reference - **Qwen Image 2.1:** generated result The comparison output is automatically assembled by the workflow. ------------------------------------------------------------------------ ## Workflow Overview ``` text ┌─────────────────────┐ │ Image #1 │ │ Edit Target │ └──────────┬──────────┘ │ ▼ ┌─────────────────┐ │ Qwen Image 2.1 │ │ Conditioning │ └────────┬────────┘ ▲ │ ┌──────────┴──────────┐ │ Image #2 │ │ Style Reference │ └─────────────────────┘ │ ▼ ┌─────────────────────┐ │ Edit Instruction │ └──────────┬──────────┘ │ ▼ ┌─────────────────────┐ │ LCS Tone Adjustment │ └──────────┬──────────┘ │ ▼ ┌─────────────────────┐ │ Sampling │ └──────────┬──────────┘ │ ▼ ┌─────────────────────┐ │ VAE Decode │ └──────────┬──────────┘ │ ▼ ┌─────────────────────┐ │ GLSL Cleanup │ │ Nyquist Notch │ └──────────┬──────────┘ │ ▼ ┌─────────────────────┐ │ Output │ └─────────────────────┘ ``` ------------------------------------------------------------------------ # Features - Qwen Image 2.1 Edit - Two-image style transfer - Image #1 as the primary editing target - Image #2 as the style reference - Qwen3-VL 8B image/text encoder - INT8 ConvRot diffusion model - BF16 Qwen Image 2.1 VAE - Qwen Image 2.1 model caching - LCS-based latent tone adjustment - Contrast, brightness, saturation and color-temperature controls - Step-range control for latent tone adjustment - Resolution control - Optional latent switching - Automatic image comparison - Automatic comparison reel - GLSL-based artifact reduction - PNG output ------------------------------------------------------------------------ # Requirements ## ComfyUI The workflow was created with: - ComfyUI `0.36.0` - ComfyUI frontend `1.53.6` A recent version of ComfyUI is recommended. ------------------------------------------------------------------------ # Required Models The workflow uses the following Qwen Image 2.1 components. ## Diffusion Model ``` text qwen_image_2.1_int8_convrot.safetensors ``` ## Text / Vision Encoder ``` text qwen3vl_8b_int8_convrot.safetensors ``` ## VAE ``` text qwen_image_2.1_vae_bf16.safetensors ``` ### Official Model Repositories - [Hugging Face --- Comfy-Org/Qwen-Image-2.1](https://huggingface.co/Comfy-Org/Qwen-Image-2.1) - [ModelScope --- Comfy-Org/Qwen-Image-2.1](https://modelscope.cn/models/Comfy-Org/Qwen-Image-2.1) ------------------------------------------------------------------------ # Model Directory Structure Place the required models in the corresponding ComfyUI directories: ``` text ComfyUI/ └── models/ ├── diffusion_models/ │ └── qwen_image_2.1_int8_convrot.safetensors │ ├── text_encoders/ │ └── qwen3vl_8b_int8_convrot.safetensors │ └── vae/ └── qwen_image_2.1_vae_bf16.safetensors ``` ------------------------------------------------------------------------ # Custom Nodes The workflow uses functionality from: ### ComfyUI-FearnWorksNodes Used for: - `LoadAndResizeImage` - `LCSLoadData` - `LCSToneAdjust` ### ComfyUI LayerStyle Used for: - `LayerUtility: ImageReel` - `LayerUtility: ImageReelComposit` The workflow also uses standard ComfyUI nodes for model loading, Qwen Image 2.1 conditioning, sampling, VAE processing, image comparison, image saving, and GLSL shader processing. ------------------------------------------------------------------------ # How to Use ## 1. Load Image #1 `IMAGE #1` is the main image that you want to transform. This is the image whose subject and composition you generally want to preserve while transferring the visual characteristics of Image #2. The current workflow resizes Image #1 to: ``` text 1024 × 1024 ``` with dimensions divisible by 32. ------------------------------------------------------------------------ ## 2. Load Image #2 `IMAGE #2` is the style reference. It provides the visual characteristics that Qwen Image 2.1 should transfer to Image #1. The current workflow also resizes Image #2 to: ``` text 1024 × 1024 ``` ------------------------------------------------------------------------ # Prompt The default prompt is: ``` text Take the style of and apply it to ``` The Qwen Image 2.1 conditioning node supports multiple image references. Images can be addressed in the prompt as: ``` text ... ``` Image #1 is the primary editing image. Additional images can be used as references. ------------------------------------------------------------------------ # Prompt Examples ## Basic Style Transfer ``` text Take the style of and apply it to ``` ## Preserve the Subject ``` text Apply the visual style, color treatment, lighting and artistic characteristics of to while preserving the subject, pose, composition and environment of . ``` ## Cinematic Style ``` text Transfer the cinematic visual language, lighting, color grading, contrast and atmosphere of to . Preserve the subject and composition of . ``` ## Illustration Style ``` text Transform using the illustration style, line quality, color palette and rendering characteristics of . Preserve the subject, pose and composition of . ``` ------------------------------------------------------------------------ # Latent Color Subspace (LCS) This workflow includes an LCS-based tone adjustment stage using `LCSToneAdjust`. **Latent Color Subspace (LCS)** is a research approach introduced in *The Latent Color Subspace: Emergent Order in High-Dimensional Chaos* by Mateusz Pach, Jessica Bader, Quentin Bouniot, Serge Belongie, and Zeynep Akata. The research analyzes color representation in the VAE latent space of FLUX.1 and identifies a structured subspace associated with **Hue, Saturation, and Lightness**. The authors demonstrate that this representation can be used to predict and explicitly control color through training-free latent-space manipulation. The paper was accepted at **ICML 2026**. In this workflow, LCS is used as a **latent-space tone adjustment mechanism** within the Qwen Image 2.1 generation pipeline. The workflow exposes controls for: - Contrast - Brightness - Saturation - Color temperature - Start step - End step Current settings: ``` text Preset: Custom Contrast: 0.95 Brightness: 0.03 Saturation: 0.90 Color Temperature: 0.02 Start Step: 30 End Step: 33 ``` The LCS stage provides an additional way to influence the tonal and color character of the generation during the model's processing rather than applying a conventional pixel-space color grade after generation. > **Important:** The original LCS research was developed and evaluated > on FLUX.1. This workflow applies an LCS-based implementation to Qwen > Image 2.1. This should be understood as an application of the approach > in this workflow, not as a claim that the original LCS paper evaluated > Qwen Image 2.1. ### Research **The Latent Color Subspace: Emergent Order in High-Dimensional Chaos** Mateusz Pach, Jessica Bader, Quentin Bouniot, Serge Belongie, Zeynep Akata ICML 2026 - Paper: https://arxiv.org/abs/2603.12261 - Code: https://github.com/ExplainableML/LCS ------------------------------------------------------------------------ # Resolution Qwen Image 2.1 supports direct output up to approximately 2K resolution. The workflow starts with: ``` text 1024 × 1024 ``` The resolution system uses a total pixel budget while preserving the selected aspect ratio. For higher-resolution work, increase the resolution according to available GPU memory. ------------------------------------------------------------------------ # Sampling Settings The current workflow uses: ``` text Steps: 27 CFG: 1 Sampler: Euler Scheduler: Simple Denoise: 1.0 ``` The seed is randomized after generation by default. ## CFG The workflow uses: ``` text CFG = 1 ``` The negative prompt is currently empty. ------------------------------------------------------------------------ # Qwen Image 2.1 Cache The workflow uses: ``` text QwenImage21Cache ``` with: ``` text Device: auto Dtype: default ``` This provides the cached model path used by the sampler. ------------------------------------------------------------------------ # Artifact Reduction A GLSL shader is included after VAE decoding. The shader implements a **Nyquist Notch** filter designed to reduce small grid-like artifacts in the generated image. This is a separate post-processing stage and is not part of the Qwen Image 2.1 model itself. ------------------------------------------------------------------------ # Comparison Output The workflow automatically creates a comparison containing: 1. Image #1 --- original target 2. Image #2 --- style reference 3. Qwen Image 2.1 --- generated result The workflow includes: - `ImageCompare` - `LayerUtility: ImageReel` - `LayerUtility: ImageReelComposit` This makes it easy to compare the original target, style reference and generated result in a single image. ------------------------------------------------------------------------ # Recommended Workflow ### Step 1 Choose an Image #1 with a clearly defined subject and composition. ### Step 2 Choose an Image #2 that strongly represents the visual style you want. ### Step 3 Start with: ``` text Take the style of and apply it to ``` ### Step 4 If necessary, explicitly describe which characteristics should be transferred: ``` text color palette lighting rendering style texture contrast atmosphere art direction ``` ### Step 5 If the result changes important elements of Image #1, make the preservation requirements explicit: ``` text Apply the style of to . Preserve the identity, pose, composition, camera angle and environment of . ``` ------------------------------------------------------------------------ # Reference Images The Qwen Image 2.1 conditioning node supports up to 10 reference images. The workflow currently uses: ``` text Image #1 → Editing target Image #2 → Style reference ``` Additional reference images can be connected to the corresponding image inputs. Example: ``` text Use for the visual style and for the color palette. Apply both to . ``` ------------------------------------------------------------------------ # Workflow Structure ``` text IMAGE #1 │ ├─────────────────────┐ │ │ ▼ │ Qwen Image 2.1 │ Conditioning │ ▲ │ │ │ IMAGE #2 │ Style Reference │ │ │ └──────────┬──────────┘ │ ▼ Edit Instruction │ ▼ LCS Tone Adjust │ ▼ QwenImage21Cache │ ▼ KSampler │ ▼ VAE Decode │ ▼ GLSL Cleanup │ ▼ Output ``` ------------------------------------------------------------------------ # Notes - Image #1 is the primary edit target. - Image #2 is the style reference. - Explicitly identify reference images when using multiple images. - The workflow starts at 1024 × 1024. - Higher resolutions require additional GPU memory. - The default CFG is 1. - The workflow uses Euler sampling. - LCS tone adjustment is performed during the generation pipeline. - The GLSL stage is a separate post-processing artifact-reduction step. - The comparison output is generated automatically. ------------------------------------------------------------------------ # Credits This workflow uses: - Qwen Image 2.1 - Qwen3-VL 8B - ComfyUI - ComfyUI-FearnWorksNodes - ComfyUI LayerStyle - The Latent Color Subspace (LCS) research ### LCS Research Pach, M., Bader, J., Bouniot, Q., Belongie, S., & Akata, Z. (2026). *The Latent Color Subspace: Emergent Order in High-Dimensional Chaos.* International Conference on Machine Learning (ICML 2026). Research: https://arxiv.org/abs/2603.12261 Code: https://github.com/ExplainableML/LCS ------------------------------------------------------------------------ # License This workflow configuration is released under the **Apache License 2.0**. Third-party models, custom nodes and other dependencies remain subject to their respective licenses. The Qwen Image 2.1 model weights are not redistributed by this repository. ------------------------------------------------------------------------ # Author **zuanfilm** Workflow for ComfyUI.