Freefield

ComfyUI-freefield

Freefield's 117 image, video, VFX and audio apps inside ComfyUI. A sidebar tab lists them by category; pick one, download its open-weight models with one button, and open it as a workflow with the lab's example already filled in. Then press Run.

<p> <img src="docs/panel.jpg" width="31%" alt="The Freefield tab: search, categories and app cards with what each one still needs"> <img src="docs/app.jpg" width="31%" alt="An app's page: summary, the lab example's time, the run on a 24 GB card, its models and their licenses"> </p>

There are no new nodes. Every Freefield app is a plain ComfyUI workflow, so what you open is ordinary ComfyUI that you can edit, save and share. The full Freefield studio (web UI, history, lab) lives in the main repo; this node is the shortest way to try its apps in a ComfyUI you already have.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/neuregex/ComfyUI-freefield

(On ComfyUI Desktop the folder is custom_nodes inside the ComfyUI folder you chose at install.) It is also in the Comfy Registry: search Freefield in ComfyUI Manager, or run comfy node install freefield.

Restart ComfyUI and open the Freefield tab in the left sidebar (the grid icon).

Use

1. Find an app by category or search. Each card says whether it is Ready here or how much it still needs to download, and how it does on a 24 GB card: Ran on 24 GB for the apps tested on one (see Tested), otherwise Freefield's estimate from the models' weight sizes.
2. On the app's page, Download fetches the model files the workflow loads from Hugging Face into ComfyUI's model folders, under their current names (diffusion_models, text_encoders, not the old unet and clip). Files arrive as .part and are renamed when complete, so an interrupted download resumes. Effects, Styles and Talking also list LoRAs the workflow does not load as it opens (the other effects and styles, a LoRA to add yourself); those have a button of their own.
3. Open with example loads the app in a new workflow tab with the values and input files of the example Freefield's lab made on a GPU. Open empty loads the app's defaults. Press Run.

Gated models (FLUX 3 Klein, LTX-2.5) need their license accepted on huggingface.co and a token: run hf auth login, or set HF_TOKEN, before starting ComfyUI.

What it needs

• ComfyUI 0.42 or newer. The apps use core nodes from that version; the tab lists any node type your ComfyUI lacks.
• Two custom node packs for some apps. ComfyUI-LTXVideo (the LTX-2.5 IC-LoRA apps) and ComfyUI-CrossViewWarp (Re-shoot); the app's page names them and links them. ComfyUI Manager installs both.
• Disk and VRAM. The app page lists every file with its size and license. Freefield's VRAM label adds up the weight sizes; it is an estimate, not a measurement. ComfyUI can keep part of the weights in system RAM, so an app labeled above 24 GB can still run on a 24 GB card with enough RAM: Dub and Cinema Studio Video did (see below).

Tested

On 5 November 2026, 72 of the 117 apps ran with their catalog example on a Windows PC with an RTX 3090 Ti (24 GB) and 104 GB of RAM, under ComfyUI 0.42.0 with dynamic VRAM and PyTorch 2.11 + CUDA 13.0, all without an error. Among them are Dub (135 s) and Cinema Studio Video (144 s), whose weights add up to more than 24 GB and which Freefield's estimate lists as needing 48 GB. tested/rtx3090ti-windows-2026-11-05.json has every app with its time, and the tab shows them. The other 45 apps have not run on a 24 GB card yet, among them the four MiniMax H3 apps (H3 Photo to Video, H3 Video, Extend Video, Re-shoot), whose text encoder alone is 27 GB.

How it works

• catalog/ is a snapshot of the main repo: every ready app's app.json and workflow.api.json, unchanged, plus catalog.json (categories, models, VRAM, the lab's example) and a thumbnail per app. tools/sync_catalog.py writes it from a checkout of the main repo. tested/ lists the apps that ran on a 24 GB card through this node.
• Download gets every file the app's workflow loads as the tab opens it, empty or with the example. That includes files Freefield marks optional, such as the small Gemma model of the video apps' prompt enhancer: ComfyUI checks every loader before it runs anything, even one on a branch that is switched off. A new file goes to the first folder ComfyUI uses for its kind (an extra_model_paths.yaml folder with is_default: true comes first).
• Opening an example downloads its input files once from the main repo, at the ref that catalog.json names, into ComfyUI's input folder, as freefield_<app>_<input>.<ext>, and fills the workflow exactly as Freefield's lab does.
• The node reaches the network only when you press a button: huggingface.co for models, raw.githubusercontent.com for example files.

Tests

FREEFIELD_SRC=../freefield python -m pytest tests

Checks the catalog and that every example opens as exactly the prompt Freefield's lab ran for it (117 apps). tests/e2e/ drives a real ComfyUI with Playwright: the tab, search, opening Face Swap with its example (18 nodes, both photos in their loaders), a model-free app run through the queue, and a download from a stand-in for Hugging Face (Smooth's model, then Styles: the files its workflow loads, and the other styles' LoRAs with the second button). Both pass on a CPU-only ComfyUI 0.42.0 with frontend 1.55.18 on Linux. On a GPU see Tested; it has not been tried on a Mac.

License

MIT. Model weights are not part of this repository and keep their own licenses, listed per model on each app's page. Freefield is an independent open-source project. It is not affiliated with, endorsed by or sponsored by Higgsfield.