What is Freefield?
Freefield is a web studio laid out like Higgsfield, the paid AI image and video platform, in which every app is a ComfyUI workflow running on open-weight models. There are 120 apps, 119 marked ready: 39 with workflows of their own, such as image creation on FLUX 3 Image Klein, video with sound on LTX-2.5 and MiniMax H3, relighting, upscaling, character swaps, lip-sync dubbing and music, plus 72 template apps that put a fixed prompt over one of those pipelines, as most of Higgsfield's catalogue does.
Each app is a folder with an app.json manifest and a workflow; the menu, forms, model downloads, VRAM map and tests are generated from those files, and a script drafts a new app from any ComfyUI workflow. No API or partner nodes are used: every model is a file you download, listed with its Hugging Face repo, size and licence.
It was built for Comfy Org's #OpenSourceHiggsfield challenge and made public on 27 October 2026. The code is MIT; the backend runs on Modal or, with a local script, on your own ComfyUI.
Is Freefield worth it?
Freefield shows how much of a credit-metered creative platform can be rebuilt from public parts. According to the author's launch post it covers 95 of the 129 tools and apps Higgsfield lists, and the README documents the evidence for each: 48 lab runs on Modal between 20 and 26 October, on L40S and H100 cards, in which 710 of 733 cases succeeded, with the time and peak VRAM of every example shown in the app.
The economics are the point. Higgsfield's Soul ID trains a model of a person; Freefield's character apps copy the person from one or three reference photos on each run instead, which took 13.0 seconds of L40S time, roughly $0.007. The README estimates that training a character LoRA would cost about $2 per person and leaves it for later. End-to-end runs from the browser are not instant: a morph video took 160.8 seconds on an H100, 128.7 of them on the GPU.
Hardware is the real barrier. Image apps fit a 24 GB card, from 0.4 GB for background removal to 17.8 GB for Create Image, but LTX-2.5 video sits at 21.5 to 22.8 GB, tight on a 4090, and the MiniMax H3 apps need 27.1 GB for the text encoder alone, so they run on Modal. Some Hugging Face models are gated and need a token and licence acceptance.
The catch for commercial work is licensing. MIT covers Freefield's code, not the weights, and each model carries its own terms; one image model the author considered, Fibo 2.0, was left out because its weights are non-commercial. Check the manifest of every app you plan to sell output from. The project is also one person's work, two weeks old, with 843 stars and a v1.0 tag from 3 November.
Where it earns its place
A two-person e-commerce video studio
Make product packshots and short ads without paying per credit
Deploy the backend to Modal and use the template apps such as Packshot, Volcano Ad and Product Shot; costs are Modal GPU seconds, shown per example in the app. Check each model's licence before client work.
A ComfyUI workflow author
Turn my own workflow into a web app with a form
Run tools/import_workflow.py on the API-format workflow to draft an app folder, edit the app.json manifest, and the menu, form, downloads and tests are generated from it.
A hobbyist with an RTX 4090
Run image edits and relighting at home
python tools/local_backend.py setup, then download only the image packs; Create Image peaks at 17.8 GB and relight at 21.5 GB, both inside a 24 GB card.
Before you commit
- Setup
- Node 22.6+, Python 3.10+, and either a Modal account or a local GPU; tools/doctor.py lists what is missing, then download model packs, deploy the backend and run npm run serve.
- Integrations
- ComfyUI 0.42.0 at a pinned commit, the Comfy SDK, ComfyUI-LTXVideo, Modal, Hugging Face for model downloads; a CLI (npm run app) as well as the web UI.
- Data and privacy
- Self-hosted: images and videos stay on your Modal volume or your own machine. A gateway token protects the backend.
- Cost at scale
- Modal billing by GPU second; the lab's examples show the time for each app. Large video models need H100 or L40S class cards.
- Limits
- 23 of 733 lab cases failed; one app is still a draft; MiniMax H3 needs more than 24 GB; no character training; MIT does not cover model weights.
What’s good
- 120 apps on open weights, with no API or partner nodes
- Every app is a manifest plus a workflow, easy to inspect or extend
- Lab runs with time and peak VRAM published for each example
- Runs on Modal or on your own ComfyUI with pinned versions
- Character apps need no training, about $0.007 a run on Modal
What’s not
- Video apps need 21.5 GB or more; H3 apps need Modal
- Model licences vary and some forbid commercial use
- No Soul ID-style training yet
- Single maintainer, two weeks old
Best for
ComfyUI users and small studios who want Higgsfield-style effects without credits and have a 24 GB GPU or a Modal account.
Freefield alternatives
Freefield: common questions
- Is Freefield worth it?
- Freefield is a thorough open-weights copy of Higgsfield's catalogue and an unusually well-documented ComfyUI project, though you need a big GPU or a Modal account and patience with model licences. We rate it 3.8 out of 5.
- How much does Freefield cost?
- Free under MIT. You pay for GPU time: on Modal, a character image took 13.0 s of L40S time in the lab, about $0.007 at Modal's base price of $1.95 an hour; locally, only your own hardware.
- What are the downsides of Freefield?
- Video apps need 21.5 GB or more; H3 apps need Modal; model licences vary and some forbid commercial use; no Soul ID-style training yet; single maintainer, two weeks old.
- Who is Freefield best for?
- ComfyUI users and small studios who want Higgsfield-style effects without credits and have a 24 GB GPU or a Modal account.
- How much does Freefield cost for a team or in production?
- Modal billing by GPU second; the lab's examples show the time for each app. Large video models need H100 or L40S class cards.
- Is Freefield safe for company data?
- Self-hosted: images and videos stay on your Modal volume or your own machine. A gateway token protects the backend.
- What are the best Freefield alternatives?
- The closest alternatives we have reviewed are ComfyUI, InvokeAI and ArtCraft.
Answers checked Nov 3, 2026.