Skip to main content
Ievgen Kovalevskyi gesturing next to a GPU and the FREEFIELD headline — an open studio on your GPU

Freefield: an Open Studio Like Higgsfield on Your Own GPU

An open-source rebuild of Higgsfield's creative studio apps on ComfyUI: ~120 apps for image, video, VFX and audio. What Freefield does, how to run it on your GPU, and an honest note on model licenses.

A lot of people pay for AI studios and hit the credit wall right when they need one more clip or frame. Another path: run an open studio on your own GPU and generate without a studio subscription's credit meter.

Freefield is an open-source project by neuregex: Higgsfield-style creative apps rebuilt on ComfyUI and open model weights. The menu covers image, video, VFX and audio — about 120 apps. The software is free. Generations run on your NVIDIA GPU, on Modal, or on a rented card (RunPod, Vast.ai) — without studio subscription credits.

Below is what Freefield can do, who it's for, how to run it on your own GPU from the repository README, and an honest licenses block. The GitHub link is at the end.

In short: clone the repo, set up the backend on your ComfyUI, download the model packs you need, open the studio in the browser. Start with 3–4 apps: Cinema Studio, effects, recast/relight, talking. Before any commercial use, check each model's license.

WHO THIS IS FOR

SituationWhat Freefield gives you
You pay for an AI studio and keep hitting creditsThe same kinds of tools on your own GPU, without studio credits
You already use ComfyUIA studio-style web UI on top of those workflows
You want cinema cameras, relight, character swap, lipsyncReady apps with open weights and lab examples
You don't have a strong cardModal, or a rented GPU with Freefield's public template

WHAT FREEFIELD CAN DO (4 EXAMPLES OUT OF ~120)

I won't list every app — the README has more than a hundred. These four are a solid place to start:

  1. Cinema Studio — cinematic stills: genre, era, camera body, lens, focal length, aperture, lighting and palette. Several aspect ratios, 1K–2K.
  2. Effects — ready video effects on LoRAs (bullet time and others from the app catalog). The form shows each model's trigger words.
  3. Recast and Relight — swap a person in a video / animate a character (recast); change lighting on a photo (relight: presets, or direction, softness, brightness, color).
  4. Talking — lipsync / speak: audio → video with lip motion on open weights.

Every app is a folder under apps/ with an app.json and a ComfyUI workflow. The web UI, model downloads and VRAM map are derived from those manifests. The How it works tab shows the model, size, license and commands.

WHAT YOU NEED

  • An NVIDIA GPU. For most apps in the README table, about 24 GB of VRAM (the “On a 24 GB card” column). Some apps are “tight” (e.g. relight, talking) — check the VRAM map in the repo.
  • Or: Modal, or a rented GPU via Freefield's public RunPod / Vast.ai template (no local install).
  • Node.js, Python with a PyTorch build for your card, Git.
  • A Hugging Face read token for gated models: accept the model license on the Hub first, then set HF_TOKEN or run hf auth login.

The repository software is free. Generations run on your hardware or a rented GPU — not on studio subscription credits.

HOW TO RUN IT ON YOUR OWN GPU

Steps from the README (*On your own GPU, without Modal*). Flags and edge cases live in the repo.

# 1. Once: ComfyUI, custom nodes and comfy-api-proxy
python tools/local_backend.py setup ~/freefield-comfy

# 2. Models for the packs you want (see packs/index.json)
python tools/local_backend.py download ~/freefield-comfy image-flux3 --dry-run
python tools/local_backend.py download ~/freefield-comfy image-flux3

# 3. Gateway token, npm, then the studio
python tools/setup_token.py --env-only
npm install
python tools/local_backend.py serve ~/freefield-comfy --studio

The studio opens locally (default http://127.0.0.1:3000). Without --studio, you can run the backend and the web app separately.

README alternatives: one-click RunPod/Vast.ai with the public template, or Modal (the GPU scales to zero between runs).

HONEST NOTES ON LICENSES AND GPU

  • Freefield code is MIT: you can read, change and use it under MIT terms.
  • Model weights are not in the repository. Each app.json lists its own licenses (Apache-2.0, MIT, OpenRAIL, FLUX/Krea/LTX community licenses, and others). Some are non-commercial / research-only. The UI shows the license next to the model — read it before you work.
  • I don't promise “for client ads.” If you need a commercial campaign, check every model's license and use footage and characters you own or have permission for. The authors state the apps are meant for SFW work with materials you have rights to.
  • GPU. 24 GB covers many apps end to end; heavier ones run by phases or need a larger card. The VRAM map and lab runs are in the repo's docs/.

WHERE TO GET IT

GitHub: github.com/neuregex/freefield

Built and maintained by neuregex. The README has a ~2-minute demo, the app table, VRAM notes, lab runs, and guides for Modal and rented GPUs. Credit to the authors for the open code and clear license labels.

I collect more useful skills and tools in the Telegram channel Ads is Easy.