About ComfyUI 0.35.0 — Node-Based Generative AI for Images, Video and 3D
ComfyUI is a node-based engine for generative AI. You build a workflow by connecting nodes on a graph, and it generates images, video, 3D models, audio or text with open models running on your own GPU. It is GPL-3.0 licensed and developed by Comfy Org; the repository moved from comfyanonymous/ComfyUI to Comfy-Org/ComfyUI, and old links redirect.
This page links to the official v0.35.0 Windows portable builds and the source archive on GitHub. Bineret did not build ComfyUI and is not affiliated with Comfy Org.
Before you download 1.4 to 1.8 GB
The portable builds are Windows-only folders with Python and PyTorch already inside. The project itself says the portable build is not recommended for regular users, and points them to the Comfy desktop app for Windows and macOS, from comfy.org, instead. The portable build suits people who want a folder they can move, several installs side by side, or nothing to install.
Pick the file for your graphics card:
- Nvidia RTX 20 series and newer:
ComfyUI_windows_portable_nvidia.7z, 1.78 GB, with Python 3.13 and PyTorch for CUDA 13.0. - Nvidia GTX 10 series and older:
ComfyUI_windows_portable_nvidia_cu126.7z, 1.72 GB, with Python 3.12 and CUDA 12.6. The README warns against using it on 20 series and newer cards. - AMD:
ComfyUI_windows_portable_amd.7z, 1.47 GB. - Intel:
ComfyUI_windows_portable_intel.7z, 1.39 GB.
None of them include models. You download the models you want and put them in the right folders.
On Linux or macOS, use the source archive and the manual install, or the desktop app on macOS.
What it does
- A visual node graph for image, video, audio, 3D and text workflows without code, with reusable subgraphs, templates, App Mode for putting a simple interface over a complex workflow, and a local API.
- Efficient execution: a queue, re-running only the parts of a graph that changed, memory management that offloads models, and quantized models.
- Wide model support, including Stable Diffusion 1.5, SDXL and 3.5, Flux.1 and Flux.2, Qwen Image, Wan 2.1 and 2.2, LTX-Video 2, HunyuanVideo 1.5, ACE-Step 1.5, Stable Audio 3, Hunyuan3D 2.1, and Gemma and Qwen3 for text.
- Editing tools: inpainting, outpainting, masks, upscaling, frame interpolation, segmentation and depth estimation.
- Workflows that travel: save them as JSON, or recover the full workflow and seed from an image or video ComfyUI made.
- High bit depth and HDR: 16-bit PNG, 32-bit EXR and 10-bit AVIF among them.
- Custom nodes to extend it.
Offline by default, paid nodes optional
The core downloads nothing unless you ask it to. Partner nodes give access to closed models such as Nano Banana and Seedance through paid APIs. Start ComfyUI with --disable-api-nodes to switch them off and keep everything offline.
Custom nodes are code
Custom nodes run Python on your machine with your permissions, so install only ones you trust. The project also warns that commits between stable release tags can be unstable and break custom nodes. Stable versions come out roughly every two weeks.
What 0.35.0 changed
Released on 9 September 2026. It adds WAN3-Prime model support and Recraft V4 styles through partner nodes, retires Veo 2 and Veo 3.0 from the Google partner nodes, supports AVIF in Save Image Advanced, and fixes HEVC video remuxing. In a container it now respects the cgroup memory limit instead of the host’s RAM. A new File3DToMesh node reads GLB, GLTF, OBJ and STL files, and there is a new cfgpp_ud10_ab sampler.
When to use something else
If you would rather paint on a canvas with layers and brushes than wire up nodes, InvokeAI is built around that. Without a capable GPU, local generation is slow whichever tool you pick.
Key capabilities
Workflows as graphs
Connect nodes for models, prompts, samplers and image operations into a graph you can save, share and reuse. App Mode puts a simple interface over a complex workflow, and a local API runs it from other software.
Beyond still images
Video with Wan, LTX-Video and HunyuanVideo, audio with ACE-Step and Stable Audio, 3D with Hunyuan3D, and text with Gemma and Qwen3, alongside Stable Diffusion, Flux and Qwen Image.
Choose the build for your GPU
Nvidia RTX 20 series and newer, older Nvidia cards, AMD and Intel each have their own portable build, 1.39 to 1.78 GB. The project recommends its desktop app for regular users.
Offline unless you say otherwise
The core downloads nothing on its own. Partner nodes reach paid closed models, and --disable-api-nodes switches them off so every built-in feature stays local.
Custom nodes run code
Extensions are Python running with your permissions, so install only what you trust. Commits between stable tags can break them; stable versions arrive roughly every two weeks.
Questions & Answers
Which file do I need?
It depends on your graphics card: Nvidia RTX 20 series and newer take the nvidia build, older Nvidia cards take nvidia_cu126, and AMD and Intel cards have their own builds. All four are Windows only.
Portable build or desktop app?
The project recommends the desktop app for regular users. The portable build is for people who want a self-contained folder, several installs, or no installer.
Does the download include models?
No. You download models separately and place them in the models subfolders.
Is ComfyUI free?
Yes, it is GPL-3.0. Partner nodes that call closed models are paid, and Comfy Cloud is a paid hosted option, but neither is needed to use ComfyUI locally.
Does it work on a Mac or Linux?
Yes, through the manual install from the source archive, or the desktop app on macOS. The portable builds on this page are Windows only.
Are custom nodes safe?
They are Python code that runs with your permissions. Install only nodes from sources you trust, and expect some to break when you update ComfyUI.
Which GPU should I buy for it?
The project keeps a recommendations page in its GitHub wiki, linked from the README's QA section.
Is Bineret connected to ComfyUI?
No. This page describes ComfyUI and links to the files Comfy Org publishes on GitHub.
Tutorials
1. Pick the right file
Nvidia RTX 20 series or newer: ComfyUI_windows_portable_nvidia.7z. Nvidia GTX 10 series or older: ComfyUI_windows_portable_nvidia_cu126.7z, and never on a 20 series or newer card. AMD: ComfyUI_windows_portable_amd.7z. Intel: ComfyUI_windows_portable_intel.7z. If you are new to ComfyUI, the project recommends its desktop app from comfy.org over any of these.
2. Extract it
Extract the archive with 7-Zip, or with File Explorer on recent versions of Windows. Put the folder somewhere with plenty of free space; models are large.
3. Add models
For smaller models, put the checkpoint file (the large .ckpt or .safetensors) in ComfyUI\models\checkpoints. Larger models often come as several files that belong in different subfolders of ComfyUI\models\; follow each model's instructions.
4. Start it
Run the start script in the extracted folder. It opens ComfyUI in your browser. If the Nvidia build does not start, update your Nvidia drivers first.
Share models with another UI
Rename extra_model_paths.yaml.example in the ComfyUI folder to extra_model_paths.yaml and set the paths to your existing model folders.
Keep everything offline
Launch with --disable-api-nodes to turn off the paid partner nodes, so no built-in feature reaches the internet.
Linux and macOS: manual install
unzip ComfyUI-0.35.0.zip
cd ComfyUI-0.35.0
# install PyTorch for your GPU first, following the README
pip install -r requirements.txt
python main.pyPython 3.13 is the best supported. On Apple Silicon, install PyTorch as Apple's Metal guide describes before these steps. Alternatively, pip install comfy-cli and then comfy install sets it up for you.
Support
Where support comes from
ComfyUI is developed by Comfy Org and contributors. The documentation covers installation, nodes and workflows, bugs go to GitHub issues, and the project runs Discord and Matrix communities for questions.
What Bineret covers
This page. If a download link is broken or something written here is wrong, tell us and we will fix it. We do not provide models, GPUs or help with workflows.
Licences of models
ComfyUI's GPL-3.0 licence covers ComfyUI, not the models you run in it. Each model has its own licence, and some restrict commercial use; check before using outputs commercially.
