Getting started with ComfyUI for animation previs

Previs here means a still or a short clip you can throw away while you judge a shot. ComfyUI is the open graph in the directory: GPL-3.0, for Windows, Linux, and macOS, with GitHub release v0.42.1 published 19 October 2026 on comfyanonymous/ComfyUI. The steps are the manual install, the template browser, two camera nodes from the October changelog, and the Wan Animate 2 tutorial. Which changelog lines are partner nodes is a separate piece, ComfyUI’s October builds.

Install from the page, not from a screenshot

The install page is four moves: an environment, a clone, dependencies, then start. It says to install Miniconda, then run:

conda create -n comfyenv
conda activate comfyenv

The clone line is git clone https://github.com/Comfy-Org/ComfyUI.git. That remote is Comfy-Org. The directory row is the other GitHub path. This guide does not claim they are the same checkout. Follow the URL printed on the install page if you are following that page. On Windows it also says to install the Microsoft Visual C++ Redistributable when it is missing.

The GPU wheels are separate blocks on the same page: Nvidia’s cu130 index, AMD’s ROCm 7.2 index on Linux, a conda nightly for Mac ARM, and an XPU index for Intel Arc. The Intel note says not to install CUDA wheels. AMD on Windows is a different command and asks for Windows 11, a current driver, and 64-bit Python 3.13. Copy the block that matches the card.

From the cloned folder the shared commands are:

cd ComfyUI
pip install -r requirements.txt
python main.py

python main.py --help lists flags. Ctrl+C stops it. Update with git pull and pip install -r requirements.txt inside comfyenv. There is no official Docker image. Models outside the folder go in extra_model_paths.yaml, copied from the example. The desktop app uses a different path. Do not overwrite it.

Load a template, then check the package version

Open templates from the sidebar icon or from Workflow → Browse Workflow Templates. If a file is missing, the docs say the loader nodes turn red, a banner counts the errors, and the Errors tab lists them under Setup required / Missing Models, with Download and Download all. The desktop app saves into ComfyUI/models/…. A self-hosted browser starts a download you place yourself. Detection wants the filename directly under the matching top-level folder. A nested subfolder can still be selected on the node.

Templates are the package comfyui-workflow-templates, not the core repo alone. The template page shows example pins and says to check requirements.txt after an update. The October changelog numbers those bundles from v0.11.82 on October 13 to v0.11.93 on October 21. Our GitHub check stops at v0.42.1 on October 19. A missing new template usually means that package is behind. Official templates, the contribution rules say, do not use third-party nodes.

Framing, then a character

For a framing test, use the 13 October nodes in v0.40.0 before you load a video model. Camera Angle takes a horizontal angle, a vertical angle, and a zoom. It outputs a camera_info and a plain-English shot description for a prompt. Create Camera Info builds a camera_info in orbit, look-at, or quaternion mode for Render Splat and the other 3D camera nodes. The changelog does not say either node writes a Blender file.

Templates v0.11.82 also lists LTX-2.5 DFR: Image to Video, described as three-stage image-to-video with spatial and optional temporal upscaling, plus audio in the same pass. Load it from the browser and let the missing-model list fetch what the template embeds. The changelog does not file that template under New Open-Source Model Support, so this guide does not call the weights open.

For a character that should move like a reference clip, follow the Wan Animate 2 tutorial instead of drawing a graph from memory. It says the workflow is in Templates, searchable as “Wan Animate 2”, or as a downloaded JSON. The model takes the driving video directly: no pose extraction and no skeleton step. A text prompt can set the camera and a new background apart from that video. Update ComfyUI first. If nodes are missing, the page gives two reasons: the build is old, or nodes failed to import.

The documented workflow uses wan_animate_2_int8_convrot.safetensors. Other variants mean changing the UNETLoader selection. The tutorial points at the Comfy-Org/Wan-Animate-2 repository for that file plus a LoRA, a text encoder, CLIP vision, and a VAE, filed under diffusion_models, loras, text_encoders, clip_vision, and vae. The filenames are on that page.

Then the page’s order: models loaded, reference image uploaded, driving video uploaded, prompt set for scene, background, and camera, then Run or Ctrl(cmd)+Enter. Longer video stays in one Motion Transfer subgraph. StartLoop and EndLoop repeat until the driving clip is covered. One pass is 81 frames, about 3.4 seconds at 24 fps, which is the fragment_length integer. Legal lengths are 81, 85, 89, and so on, written as 4n+1. Each extra segment adds fragment_length - 1 frames because the overlapping frame is removed. The tail is cut to the driving video’s frame count. Those loop nodes need ComfyUI 0.36.0 or newer. Video Stitch is a side-by-side preview, not the extender.

Partner nodes are not part of this run. The October 5 changelog adds --disable-partner-nodes. HeyGen, Vidu, Higgsfield, and Kling are names on that list, and they are not linked here. The moving previs in this guide is the template and the tutorial above.