{"id":15695,"date":"2026-11-17T11:03:58","date_gmt":"2026-11-17T02:03:58","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=15695"},"modified":"2026-11-17T11:13:27","modified_gmt":"2026-11-17T02:13:27","slug":"comfy-org-wevisone-vl-released-for-comfyui","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/comfy-org-wevisone-vl-released-for-comfyui\/","title":{"rendered":"WeVisOne-VL Vision-Language Model: 8GB+ VRAM"},"content":{"rendered":"<p><!-- lmw:facts --><\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Repository<\/td>\n<td><a href=\"https:\/\/huggingface.co\/Comfy-Org\/WeVisOne-VL\">Comfy-Org\/WeVisOne-VL<\/a><\/td>\n<\/tr>\n<tr>\n<td>Family guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-tencent-wevisone-vl-en\/\">WeVisOne-VL guide (2 articles)<\/a><\/td>\n<\/tr>\n<tr>\n<td>Publisher guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-tencent-en\/\">Tencent: models and licenses<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-11-09<\/td>\n<\/tr>\n<tr>\n<td>License<\/td>\n<td>apache-2.0<\/td>\n<\/tr>\n<tr>\n<td>Formats<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-safetensors-en\/\">safetensors<\/a><\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Unverified (not confirmed by a primary source)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Values determined by this site&#8217;s code when the information was collected. Dates are JST.<\/em><\/p>\n<p><!-- \/lmw:facts --><\/p>\n<h2>Overview<\/h2>\n<p>It has been reported that <code>Comfy-Org\/WeVisOne-VL<\/code>, a model file optimized and repackaged for the image generation and editing tool ComfyUI, has been released. Please note that at the time of writing, this information has not been officially confirmed.<\/p>\n<p>This repository is said to be based on <code>tencent\/WeVisOne-VL-2B<\/code>, a vision-language model developed by Tencent&#8217;s WeChat Vision team, adjusted for easier integration into the ComfyUI ecosystem. According to the paper linked from the model card, <a href=\"https:\/\/huggingface.co\/papers\/2610.31329\">WeVisOne-VL: Unified Visual Perception from Objects to Pixels<\/a>, this model is built on Qwen3-VL and was developed with the goal of processing everything from object-level visual recognition to pixel-level precise recognition within a single architecture. It is designed to handle region detection tasks such as open-vocabulary object detection, referring localization, object counting, pointing, and OCR, as well as detail extraction tasks like referring segmentation and prompt-directed alpha matting, all integrated into a single model. This is expected to become an option for ComfyUI users incorporating image recognition and segmentation processing into node-based workflows.<\/p>\n<h2>Specifications<\/h2>\n<p>The main structural specifications of the base model <a href=\"https:\/\/huggingface.co\/tencent\/WeVisOne-VL-2B\">tencent\/WeVisOne-VL-2B<\/a> confirmed from the documentation are as follows:<\/p>\n<ul>\n<li><strong>Base Model<\/strong>: tencent\/WeVisOne-VL-2B<\/li>\n<li><strong>Model Type<\/strong>: qwen3_vl<\/li>\n<li><strong>Architecture<\/strong>: Qwen3VLForConditionalGeneration<\/li>\n<li><strong>Decoder Configuration<\/strong>: 9.6M-parameter pixel decoder (<code>mask_decoder.pt<\/code> \/ 9,609,764 parameters)<\/li>\n<\/ul>\n<h2>Performance<\/h2>\n<p>Quantitative benchmark evaluation scores such as the Open LLM Leaderboard and comparative numerical tables with other models are not provided in the model card and related materials. Therefore, it is not possible to conclusively state objective comparisons based on numbers or superiority in benchmarks. The performance characteristics described below are qualitative evaluations based on the technical specifications published by the developers.<\/p>\n<p>According to the publisher&#8217;s explanation, the original model <code>tencent\/WeVisOne-VL-2B<\/code> seamlessly performs object recognition and pixel-level extraction by integrating structured next-token prediction with a lightweight pixel decoder composed of 9.6M parameters. In the visual feature extraction stage, it features a Visual Feature Adapter utilizing pre-merged ViT features from Qwen3-VL, and is reported to adopt a Box-Guided Query Decoder constructed from 6 Deformable Cross-Attention blocks initialized from Fourier-encoded box priors and bounding box hidden states.<\/p>\n<p>As high-resolution output functions for high-precision background removal and segmentation, an Instance-Conditioned Dense Prediction Head and <code>FullCanvasAlphaCascadeHead<\/code> are incorporated. This makes it possible to progressively refine predictions across S\/4, S\/2, and S\/1 resolution steps, outputting a continuous alpha matte at a stride-1 full resolution. Additionally, dedicated coordinate tokens are used to represent normalized bounding boxes and points, with an implementation where <code>None<\/code> is returned if the specified target does not exist in the image.<\/p>\n<p>On the other hand, practical limitations and precision considerations have also been reported. According to the developer&#8217;s notes, the following limitations exist:<\/p>\n<ul>\n<li>Fine alpha boundaries (such as delicate extraction of subject outlines and hair) are said to be susceptible to the original image resolution, severe occlusion, complex transparent structures, and the formulation of the input prompt.<\/li>\n<li>It is designed with an emphasis on inference speed and efficiency, directly generating structured outputs without performing explicit Chain-of-Thought (CoT) reasoning.<\/li>\n<li>At present, it is designed targeting single-frame still image recognition, and inputting multiple frames or grid images for videos is specified to be rejected by the pixel decoder.<\/li>\n<\/ul>\n<p>Because there are no independent third-party benchmark score tables, it is unclear how much precision difference exists compared to conventional vision-language models or segmentation models. When using it, performing hands-on verification using actual images is recommended.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>According to information from the model card and capability tags of the original model <code>tencent\/WeVisOne-VL-2B<\/code>, its greatest strength is its ability to integrally process everything from diverse object-level visual recognition to advanced pixel-unit region and boundary extraction. The specific strengths and assumed main tasks are as follows:<\/p>\n<ul>\n<li><strong>Open-Vocabulary Object Detection and Referring Localization<\/strong>: Detects and specifies the position of specific target objects present within an image as bounding boxes based on free text instructions or descriptive text.<\/li>\n<li><strong>Visual Pointing and Object Counting<\/strong>: Performs point grounding of targets present in the image and counts the number of objects matching specified classes and conditions.<\/li>\n<li><strong>OCR and Localization (OCR localization)<\/strong>: Reads text information within an image while simultaneously accurately identifying the region where that text exists.<\/li>\n<li><strong>Referring Segmentation<\/strong>: Generates precise region masks (segmentation masks) of target objects at the pixel level in response to natural language prompt instructions.<\/li>\n<li><strong>Promptable Alpha Matting<\/strong>: Outputs a high-precision alpha matte including the transparency of hair and borders for targets specified by text queries (e.g., people), supporting advanced background removal processing.<\/li>\n<\/ul>\n<p>According to the research and development team&#8217;s notes in the model card, the main intended uses and research targets for this model include verification and development in the areas of visual grounding, referring localization, pointing, counting, OCR localization, referring segmentation, and image matting.<\/p>\n<p>On the other hand, it is explicitly stated that this model is not available (not intended) for autonomous high-risk decision-making processes or purposes aimed at individual identification. In ComfyUI workflows, utilization is anticipated as pre-processing and post-processing nodes in image editing pipelines, such as automatic mask creation, background separation, and element extraction for specific subjects in generated or input images.<\/p>\n<p><!-- lmw:hardware --><\/p>\n<h2>Hardware Requirements<\/h2>\n<p><strong>Estimated requirements (calculated by Local Model Watch)<\/strong> \u2014 2.4B parameters (taken from the base model tencent\/WeVisOne-VL-2B)<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Your VRAM<\/th>\n<th>Quantization<\/th>\n<th>File size<\/th>\n<th>Est. memory needed<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>8GB (RTX 4060 \/ 3060 Ti, etc.)<\/td>\n<td>BF16<\/td>\n<td>4.5GB<\/td>\n<td>5.5GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model&#8217;s authors. Compare with other models in our <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/vram-guide-en\/\">VRAM quick reference<\/a>. What the quantization names mean: <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-quantization-en\/\">glossary<\/a>.<\/em><\/p>\n<p><!-- \/lmw:hardware --><\/p>\n<p><!-- lmw:runnability --><\/p>\n<h2>Can You Run It Locally?<\/h2>\n<p>The publisher distributes this model as safetensors.<\/p>\n<p><strong>License \u2014 <code>apache-2.0<\/code> (Commercial use allowed):<\/strong> Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.<\/p>\n<p><em>Compiled by this site&#8217;s code from the published formats and the license field. License summaries are not legal advice \u2014 check the publisher&#8217;s original terms before relying on them.<\/em><\/p>\n<p><!-- \/lmw:runnability --><\/p>\n<h2>How to Get It<\/h2>\n<p>It is reported that the model files for <code>Comfy-Org\/WeVisOne-VL<\/code> are distributed on Hugging Face as ComfyUI packages in safetensors format. The Apache 2.0 license is said to apply.<\/p>\n<p>Acquired model files are placed and used according to ComfyUI&#8217;s directory structure. According to the instructions in the repository, you are directed to place the corresponding files (<code>.safetensors<\/code>) in the following folders:<\/p>\n<p>Folder structure for placement:<\/p>\n<pre><code>\ud83d\udcc2 ComfyUI\/\n\u2514\u2500\u2500 \ud83d\udcc2 models\/\n    \u2514\u2500\u2500 \ud83d\udcc2 text_encoders\/\n        \u251c\u2500\u2500 wevisone_vl_4b_bf16.safetensors\n        \u251c\u2500\u2500 wevisone_vl_4b_int8_convrot.safetensors\n        \u251c\u2500\u2500 wevisone_vl_2b_bf16.safetensors\n        \u2514\u2500\u2500 wevisone_vl_2b_int8_convrot.safetensors\n<\/code><\/pre>\n<p>As described above, various model files corresponding to BF16 precision versions and INT8 quantized (convrot) versions of <code>wevisone_vl_2b<\/code> and <code>wevisone_vl_4b<\/code> are reportedly prepared. Please select the appropriate file according to your workflow and operating environment, place it under ComfyUI&#8217;s <code>text_encoders<\/code> folder, and use it.<\/p>\n<p>For official inference code called directly from Python and detailed specifications of the base model, please check the guidance in Tencent&#8217;s original model repository <code>tencent\/WeVisOne-VL-2B<\/code> and the official GitHub repository.<\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/comfy-org-minimax-h3-image-released\/\">MiniMax-H3-Image 16-Megapixel Image Generation Model: 4GB+ VRAM<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/24\/comfy-org-ming-image-released\/\">Ming-Image UI Design-Specialized Image Generation Model: ComfyUI Paths<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/comfy-org-releases-world-model-pan2-for-comfyui\/\">PAN2 World Model for Simulating Physical Changes: ComfyUI Paths<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/comfy-org-idol-comfyui-release\/\">Idol 3D Gaussian Splatting Generation Model: ComfyUI Paths<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:related --><\/p>\n<p><!-- lmw:next-steps --><\/p>\n<h2>What to Read Next<\/h2>\n<ul>\n<li><strong>Find models by VRAM<\/strong> (This model runs from the 8GB tier) \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/vram-guide-8gb-en\/\">Other models that run on a 8GB GPU<\/a><\/li>\n<li><strong>Explore the same model family<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-tencent-wevisone-vl-en\/\">WeVisOne-VL family overview (2 articles, 1 converted builds)<\/a><\/li>\n<li><strong>Formats this model is available in<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-safetensors-en\/\">Safetensors format guide and models<\/a><\/li>\n<li><strong>Learn about the publisher<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-tencent-en\/\">Tencent: models, licenses and articles<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/Comfy-Org\/WeVisOne-VL\">https:\/\/huggingface.co\/Comfy-Org\/WeVisOne-VL<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/tencent\/WeVisOne-VL-2B\">https:\/\/huggingface.co\/tencent\/WeVisOne-VL-2B<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/papers\/2610.31329\">https:\/\/huggingface.co\/papers\/2610.31329<\/a><\/li>\n<\/ul>\n<blockquote>\n<p><strong>This article contains unverified information.<\/strong> We will append an update note once it is confirmed by a primary source.<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"<p>Model files for Comfy-Org\/WeVisOne-VL have been released for ComfyUI, bringing object-to-pixel visual perception capabilities.<\/p>\n","protected":false},"author":1,"featured_media":15694,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[310],"tags":[1142,4824,677,209,4826,1565,4389,4037],"class_list":["post-15695","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-new-models","tag-comfy-org-en","tag-comfy-org-wevisone-vl-en","tag-comfyui-en","tag-safetensors-en","tag-tencent-wevisone-vl-2b-en","tag-unverified","tag-vision-language-model-en","tag-wevisone-vl-en"],"lang":"en","translations":{"en":15695,"ja":15693},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/15695","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=15695"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/15695\/revisions"}],"predecessor-version":[{"id":18080,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/15695\/revisions\/18080"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/15694"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=15695"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=15695"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=15695"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}