Comfy-Org adapted Qwen3-VL models to work in ComfyUI

Team Comfy-Org released adapted versions of multimodal models Qwen3-VL designed for integration into the ComfyUI visual programming environment. The release includes optimized builds with 4B and 8B parameters, which allow using computer vision capabilities directly in image generation workflows.

For correct model operation, users need to place text encoder files in strictly defined system directories of ComfyUI. This folder structure ensures compatibility with current data processing nodes and allows the model to correctly interpret visual prompts within the operation execution graph.

Models Qwen3-VL are based on an architecture that supports processing high-resolution images and understanding scene context. The 4B parameter version is oriented towards operation on consumer equipment with limited video memory, while the 8B variant provides higher accuracy in recognizing small details and text elements in images.

Integration into ComfyUI simplifies the creation of complex AI pipelines, where the vision model can act as an input data analyzer or a tool for automatic description generation (captioning) before the generation stage. The source code and model weights are available in the project repository on the Hugging Face platform under the standard license for this model family.