{"id":14646,"date":"2026-11-17T11:03:11","date_gmt":"2026-11-17T02:03:11","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=14646"},"modified":"2026-11-17T11:12:36","modified_gmt":"2026-11-17T02:12:36","slug":"comfy-org-idol-comfyui-release","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/comfy-org-idol-comfyui-release\/","title":{"rendered":"Idol 3D Gaussian Splatting Generation Model: ComfyUI Paths"},"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\/Idol\">Comfy-Org\/Idol<\/a><\/td>\n<\/tr>\n<tr>\n<td>Family guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-speridlabs-idol-en\/\">idol guide (2 articles)<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-11-02<\/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>Primary source (the publisher itself)<\/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>Comfy-Org has released &#8220;Comfy-Org\/Idol&#8221;, a ComfyUI-compatible repackaged version of &#8220;Idol&#8221;, an open-weight model that generates 3D Gaussian Splatting models from a single image.<\/p>\n<p>This model is a derivative distribution where the files from &#8220;speridlabs\/idol&#8221;, published by speridlabs, have been rearranged to match ComfyUI&#8217;s standard folder structure. It accepts a single image of an object (CG renderings, AI-generated images, or real-world photographs) as input, estimates up to 1,048,576 3D Gaussians including the backside and sides not visible in the original image, and outputs a standard 3DGS format <code>.ply<\/code> file.<\/p>\n<h2>Specifications<\/h2>\n<p>The main specifications described in the distribution repository and the documentation for the original model &#8220;speridlabs\/idol&#8221; are as follows. This model is positioned as a fine-tuned version of &#8220;TripoSplat&#8221; and is compatible with existing TripoSplat workflows.<\/p>\n<ul>\n<li><strong>Architecture and Components<\/strong>: It features a pipeline configuration combining multiple components:<\/li>\n<li>DiT (denoiser): 370.6M &#8211; Octree decoder: 50.8M &#8211; Gaussian decoder: 237.3M &#8211; BiRefNet (background removal): 220.7M &#8211; DINOv3 ViT-H+\/16 (image encoder): 840.6M &#8211; FLUX.2 VAE: 84.1M &#8211; Total parameters across the request: Approximately 1.80B (~1.80B)<\/li>\n<li><strong>Output Specifications<\/strong>:<\/li>\n<li>3D Gaussian Splatting data (<code>.ply<\/code> file) &#8211; Background-removed and cropped input image (<code>_input.png<\/code>) &#8211; Estimated camera parameters (<code>_camera.json<\/code>) &#8211; Number of output Gaussians: 32,768 to 1,048,576 (specifiable in multiples of 32, default is 262,144)<\/li>\n<li><strong>Recommended Sampling Parameters<\/strong>:<\/li>\n<li>Number of sampling steps (<code>--steps<\/code>): Default 50 steps (reducing steps speeds up generation but sacrifices fine details) &#8211; CFG strength (<code>--cfg_strength<\/code>): Default 3.0 (controls how strictly the object shape follows the input image) &#8211; Time shift (<code>--rescale_t<\/code>): Default 3.0 (higher values allocate more steps to rough overall shape formation, lower values allocate more to fine detail formation)<\/li>\n<li><strong>Input Mode Settings<\/strong>:<\/li>\n<li><code>synthetic<\/code> (default): For CG renderings or generated images of single objects. Objects are cropped to occupy 1\/1.2 of the frame &#8211; <code>real<\/code>: For real-world photographs. Objects are cropped to occupy 1\/1.7 of the frame<\/li>\n<li><strong>Optimization and Operational Settings<\/strong>:<\/li>\n<li>Fast sampling functionality is enabled by default, operating approximately 1.7x faster on GPU while maintaining the same output quality. The denoiser is compiled with <code>torch.compile<\/code>, paths with and without guidance are executed as a single batch, and cuDNN with TF32 matrix multiplication is used for attention (can be disabled with <code>--no_optimize<\/code>) &#8211; DINOv3 precision setting (<code>--dinov3_dtype<\/code>): Default is <code>bf16<\/code> (<code>fp32<\/code> can also be specified)<\/li>\n<li><strong>License Terms<\/strong>:<\/li>\n<li>Governed by the Apache-2.0 license<\/li>\n<\/ul>\n<h2>Performance<\/h2>\n<p>Below are the benchmark evaluations published in the model card of the original model &#8220;speridlabs\/idol&#8221;. Each model was evaluated using a single NVIDIA H100, comparing average scores across the input viewpoint (PSNR, SSIM, LPIPS, IoU) and seven novel viewpoints (DINOv2, CLIP, Sharpness distance).<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<p class=\"lmw-table-hint\" style=\"margin:0 0 4px;font-size:0.85em;opacity:0.7;\">\u2192 Scroll horizontally to see all columns<\/p>\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Model<\/th>\n<th>Setting<\/th>\n<th>PSNR \u2191<\/th>\n<th>SSIM \u2191<\/th>\n<th>LPIPS \u2193<\/th>\n<th>IoU \u2191<\/th>\n<th>DINOv2 \u2191<\/th>\n<th>CLIP \u2191<\/th>\n<th>Sharpness distance \u2193<\/th>\n<th>Request time \u2193<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/idol-3d-gaussian-splatting-model\/\">Idol<\/a><\/strong><\/td>\n<td>50 steps, 262,144 Gaussians<\/td>\n<td>12.25<\/td>\n<td><strong>0.726<\/strong><\/td>\n<td><strong>0.213<\/strong><\/td>\n<td><strong>0.862<\/strong><\/td>\n<td><strong>0.781<\/strong><\/td>\n<td><strong>0.885<\/strong><\/td>\n<td><strong>0.81<\/strong><\/td>\n<td>6.8 s<\/td>\n<\/tr>\n<tr>\n<td>TripoSplat<\/td>\n<td>50 steps, 262,144 Gaussians<\/td>\n<td>11.76<\/td>\n<td>0.712<\/td>\n<td>0.225<\/td>\n<td>0.860<\/td>\n<td>0.743<\/td>\n<td>0.868<\/td>\n<td>1.03<\/td>\n<td>9.9 s<\/td>\n<\/tr>\n<tr>\n<td>TRELLIS<\/td>\n<td>Paper setting, 50 + 50 steps<\/td>\n<td>11.89<\/td>\n<td>0.716<\/td>\n<td>0.235<\/td>\n<td>0.845<\/td>\n<td>0.754<\/td>\n<td>0.868<\/td>\n<td>0.97<\/td>\n<td>7.1 s<\/td>\n<\/tr>\n<tr>\n<td>LGM<\/td>\n<td>Default, 65,536 Gaussians<\/td>\n<td><strong>12.46<\/strong>*<\/td>\n<td>0.722<\/td>\n<td>0.237<\/td>\n<td>0.834<\/td>\n<td>0.575<\/td>\n<td>0.784<\/td>\n<td>1.22<\/td>\n<td><strong>1.1 s<\/strong><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>According to this comparison released by the publishers, Idol achieves the best scores across SSIM (0.726), LPIPS (0.213), IoU (0.862), DINOv2 (0.781), CLIP (0.885), and Sharpness distance (0.81), outperforming TripoSplat, TRELLIS, and LGM. Furthermore, compared to its base model TripoSplat, the request processing time has been reduced from 9.9 seconds to 6.8 seconds. On the other hand, its PSNR (12.25) and processing time (6.8 seconds) fall short of LGM (PSNR 12.46, Request time 1.1 seconds). Note that according to the publisher&#8217;s disclaimer, LGM achieves the highest numerical PSNR because PSNR is an metric averaging squared pixel errors, despite its generation results being visually blurry.<\/p>\n<p>Additionally, the original model documentation reports processing speeds on various GPUs (median of 8 images at 50 steps and 262,144 generated Gaussians), showing processing times provided by the publishers such as 3.4 seconds on B300, 6.7 seconds on H100, 14.1 seconds on RTX 5090, 19.1 seconds on RTX 4090, and 46.2 seconds on RTX 3080, ranging from consumer-grade to data-center hardware.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>According to the documentation for the original model &#8220;speridlabs\/idol&#8221; and the notes in this repository, this model specializes in taking a single still image as input and reconstructing 3D shapes\u2014including the backsides and sides of objects not visible in the image\u2014in 3D Gaussian Splatting (3DGS) format. Main use cases and specialized processing include:<\/p>\n<ul>\n<li><strong>3D Object Generation from Diverse Image Sources<\/strong>: Can generate 3D Gaussian Splatting data from 3D CG renderings, AI-generated images, and real-world photographs alike.<\/li>\n<li><strong>Automatic Framing and Background Removal Based on Input Image<\/strong>: Integrates BiRefNet internally to automatically remove backgrounds from input images and frame subjects at appropriate ratios before processing. Two modes are available depending on the type of input image:<\/li>\n<li><code>synthetic<\/code> mode: Settings for 3D CG renderings and AI-generated images. Cropping is performed so objects occupy 1\/1.2 of the entire frame. &#8211; <code>real<\/code> mode: Settings for real-world photographs. Cropping is performed so objects occupy 1\/1.7 of the entire frame.<\/li>\n<li><strong>High-Resolution and High-Density Gaussian Representation<\/strong>: The number of generated Gaussians can be specified from a minimum of 32,768 up to a maximum of 1,048,576 (~1 million). Adjustments can be made from lightweight previews to high-detail expressions according to use case and machine specs.<\/li>\n<li><strong>Integration with Existing TripoSplat Workflows<\/strong>: Because this model was developed as a fine-tuned version of TripoSplat, it can be dropped directly into existing TripoSplat workflow environments. It is well-suited for operations such as constructing nodes in ComfyUI and placing them downstream of an image generation pipeline to perform direct 3D conversion.<\/li>\n<\/ul>\n<p><!-- lmw:files --><\/p>\n<h2>Distributed Files<\/h2>\n<p><em>Weight files published in <a href=\"https:\/\/huggingface.co\/Comfy-Org\/Idol\/tree\/main\">Comfy-Org\/Idol<\/a>, listed by this site from the Hugging Face API. Sizes are the actual file sizes.<\/em><\/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>File<\/th>\n<th>Size<\/th>\n<th>ComfyUI folder<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><code>clip_vision\/dino_v3_vit_h_bf16.safetensors<\/code><\/td>\n<td>1.68GB<\/td>\n<td><code>models\/clip_vision\/<\/code><\/td>\n<\/tr>\n<tr>\n<td><code>diffusion_models\/idol_fp16.safetensors<\/code><\/td>\n<td>741MB<\/td>\n<td><code>models\/diffusion_models\/<\/code><\/td>\n<\/tr>\n<tr>\n<td><code>vae\/flux2-vae.safetensors<\/code><\/td>\n<td>336MB<\/td>\n<td><code>models\/vae\/<\/code><\/td>\n<\/tr>\n<tr>\n<td><code>vae\/idol_decoder_fp16.safetensors<\/code><\/td>\n<td>576MB<\/td>\n<td><code>models\/vae\/<\/code><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>The ComfyUI folder is shown only where the repository&#8217;s layout mirrors ComfyUI&#8217;s <code>models\/<\/code> directory.<\/em><\/p>\n<p><!-- \/lmw:files --><\/p>\n<h2>How to Get It<\/h2>\n<p>&#8220;Comfy-Org\/Idol&#8221; is available from the Hugging Face repository. Weight files are distributed in the single-file <code>safetensors<\/code> format.<\/p>\n<p>When using this model in a ComfyUI environment, place the respective repository files into ComfyUI&#8217;s designated directories as follows:<\/p>\n<pre><code>\ud83d\udcc2 ComfyUI\/\n\u251c\u2500\u2500 \ud83d\udcc2 models\/\n\u2502   \u251c\u2500\u2500 \ud83d\udcc2 diffusion_models\/\n\u2502   \u2502   \u2514\u2500\u2500 idol_fp16.safetensors\n\u2502   \u251c\u2500\u2500 \ud83d\udcc2 clip_vision\/\n\u2502   \u2502   \u2514\u2500\u2500 dino_v3_vit_h_bf16.safetensors\n\u2502   \u251c\u2500\u2500 \ud83d\udcc2 vae\/\n\u2502   \u2502   \u251c\u2500\u2500 flux2-vae.safetensors\n\u2502   \u2502   \u2514\u2500\u2500 idol_decoder_fp16.safetensors\n<\/code><\/pre>\n<p>Note that while the Comfy-Org\/Idol repository itself is not gated, the following two image encoders\/VAEs referenced by the original model pipeline are published as gated models on Hugging Face. When using them, you must agree to the terms of use (request access) on each respective model page in advance and log in to your Hugging Face account (<code>hf auth login<\/code> or setting <code>HF_TOKEN<\/code>).<\/p>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/facebook\/dinov3-vith16plus-pretrain-lvd1689m\">facebook\/dinov3-vith16plus-pretrain-lvd1689m<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/black-forest-labs\/FLUX.2-dev\">black-forest-labs\/FLUX.2-dev<\/a> (only the VAE is downloaded and used)<\/li>\n<\/ul>\n<p>Alternatively, if running via Python scripts or the command line using the original model&#8217;s SDK (<code>idol-sdk<\/code>) without ComfyUI, you can set up the environment with the following steps. Python 3.10 or higher, PyTorch 2.5 or higher, and a CUDA-capable NVIDIA GPU are required.<\/p>\n<pre><code class=\"language-bash\">git clone https:\/\/github.com\/speridlabs\/idol-sdk.git\ncd idol-sdk\npip install -e.\n<\/code><\/pre>\n<p>Required weights will be automatically downloaded to the Hugging Face cache directory upon first execution. If access permissions for the encoders have not been acquired, a notification regarding missing permissions will be issued before the download starts.<\/p>\n<p>Example commands for running from the command line are as follows. A 3DGS format <code>.ply<\/code> file, background-removed image (<code>_input.png<\/code>), and estimated camera information (<code>_camera.json<\/code>) are output as generation results:<\/p>\n<pre><code class=\"language-bash\"># When inputting a real-world photograph\npython -m idol.sample --image photo.jpg --mode real\n\n# When inputting a CG image or generated image folder\npython -m idol.sample --image renders\/ --mode synthetic --out outputs\n<\/code><\/pre>\n<p>When calling from Python code, you can run it using <code>IdolPipeline<\/code> as follows:<\/p>\n<pre><code class=\"language-python\">from idol import IdolPipeline\n\npipeline = IdolPipeline.from_pretrained(&quot;speridlabs\/idol&quot;)\noutput = pipeline(&quot;photo.jpg&quot;, mode=&quot;real&quot;, seed=42, num_gaussians=1_048_576)\noutput.gaussian.save_ply(&quot;photo.ply&quot;)\n<\/code><\/pre>\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<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/idol-3d-gaussian-splatting-model\/\">idol 3D Gaussian Splatting Generation Model: File List<\/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-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\/11\/17\/comfy-org-ming-3-0-tts-2\/\">Ming-3.0-TTS Multilingual TTS Model: 24GB+ VRAM, 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>Explore the same model family<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-speridlabs-idol-en\/\">idol family overview (2 articles)<\/a><\/li>\n<li><strong>Engines that run this model<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/engine-comfyui-en\/\">ComfyUI<\/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<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/Comfy-Org\/Idol\">Comfy-Org\/Idol (Hugging Face)<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/speridlabs\/idol\">speridlabs\/idol (Hugging Face)<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/speridlabs\/idol-sdk.git\">speridlabs\/idol-sdk (GitHub)<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/facebook\/dinov3-vith16plus-pretrain-lvd1689m\">facebook\/dinov3-vith16plus-pretrain-lvd1689m (Hugging Face)<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/black-forest-labs\/FLUX.2-dev\">black-forest-labs\/FLUX.2-dev (Hugging Face)<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Comfy-Org has released Comfy-Org\/Idol, a ComfyUI-compatible repackaging of the open-weight single-image 3D Gaussian Splatting model Idol.<\/p>\n","protected":false},"author":1,"featured_media":14645,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[203],"tags":[4493,4412,1142,4495,677,4416,1547],"class_list":["post-14646","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-image-video-and-audio","tag-3d-gaussian-splatting-en","tag-3dgs-en","tag-comfy-org-en","tag-comfy-org-idol-en","tag-comfyui-en","tag-idol-en","tag-verified"],"lang":"en","translations":{"en":14646,"ja":14644},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/14646","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=14646"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/14646\/revisions"}],"predecessor-version":[{"id":18018,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/14646\/revisions\/18018"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/14645"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=14646"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=14646"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=14646"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}