120 Apps, One Folder Each, and the GPU Numbers Nobody Publishes

Freefield describes itself in one line: open-source Higgsfield on ComfyUI, where every app is one folder and every model is open weights.
That sentence is doing a lot of work, and the repository backs it up in the least glamorous way possible. There is a table of 120 apps. Each app is a directory with an app.json manifest. The web UI, the runner, the model downloads, the VRAM map and the tests are all generated from those manifests. Change the manifest, and the form in the browser changes with it.
What is actually in the box
The menus mirror a commercial product's layout, which is the point of the exercise. Image, video, VFX and audio studios, with the video menu alone holding 46 apps grouped into Create, Edit, Motion control and Templates.
Of the 120 apps, 119 are marked ready. Thirty-nine run a workflow of their own. Five are IC-LoRA apps covering Genjutsu-style character swap, reframing, SDR to HDR, mixed media and dubbing. One is a morph app. Seventy-two are template apps, where a fixed prompt gets run over one of the existing pipelines, generated by a script rather than hand-built. Two are model-free smoke apps for CI. One, moss-image, is still a draft.
That breakdown matters more than the headline number. A repository that says 120 apps and means 72 prompt presets is being honest about what a preset is.
The part that changes how you read every other project
Every example in the web UI is a real GPU run, and the repository publishes the run log. Forty-eight lab runs on Modal between October 20 and 26, on L40S and H100 hardware. Across 733 cases, 710 came back ok.
That is 96.9%, and the 23 that did not work are listed rather than deleted.
End-to-end timings are published the same way. On October 23, running from the browser through the web UI, the server, a gateway on Modal and ComfyUI on an H100: the morph app took 160.8 seconds from button press to finished video, with 128.7 seconds of that on the GPU. mixed-media took 149 seconds. dub took 72 seconds. A day later, face-swap spent 25.2 seconds on the GPU.
Peak VRAM is listed per app, with the GPU it ran on. cinema-video needs 44.3 GB on an L40S. h3-video needs 27.1 GB just for the text encoder. reshoot, which re-renders a clip from a new angle using MiniMax H3 plus MoGe and CrossView, peaks at 44.2 GB on a 48 GB card.
Anyone who has tried to size a GPU rental from a README knows why this is unusual. Most projects tell you the model. Very few tell you the peak memory of the pipeline around the model, which is the number that actually decides whether your card works.

The apps worth looking at first
cinema-video chains two models in one graph: a still from FLUX 3 Image Klein becomes the first frame of LTX-2.5 image-to-video, with up to three camera moves selected from a list of 33 and written in the order you pick them. The lab notes one finding worth stealing: a single camera move holds, while the order of two moves breaks down. That is the kind of detail you only get from running the thing.
h3-video and h3-photo run MiniMax H3 at int8 from Comfy's templates, text or photo to a clip with speech, sound and music, 20 steps, no turbo LoRA. extend-video takes the last frame of an existing clip and continues it with its own audio. reshoot changes the camera angle on a finished shot and fills the holes left by reprojection.
The image side is a shelf of the current open models: FLUX 3 Klein with and without the upsampler, Z-Image Turbo, Krea 2 Turbo, MOSS-Image. The repository shows six identical prompts run through each, with GPU seconds and peak VRAM printed under every frame, plus a note that a single image costs about $0.007 at Modal's base rate of $1.95 an hour.
Install
Two paths, both documented.
With Modal: Node 22.6+, Python 3.10+ and a Modal account. Run python tools/doctor.py first, which reports what is missing, whether Modal is logged in, whether secrets exist, and which gated Hugging Face licenses you have not accepted. Then python tools/setup_token.py to mint the gateway token, modal secret create huggingface HF_TOKEN=..., python tools/catalog.py packs to build the download packs, modal run infra/modal_app.py::download --pack image to pull weights, modal deploy infra/modal_app.py to deploy, and npm install && npm run serve to bring up the UI on port 3000.
Without Modal: python tools/local_backend.py setup ~/freefield-comfy installs ComfyUI at a pinned commit with pinned custom nodes, plus a Comfy API proxy and the token gateway. Then you download packs locally, run the backend, and start the same web app. The app table has a column for what fits on a 24 GB card.
You can also drive an app from the terminal without the browser, by calling the create-image app with a prompt such as a red bicycle against a white wall.
Where the seams show
The dependency on Modal is real for anything heavy. Several apps list "needs Modal" next to their VRAM number, which is a polite way of saying the pipeline will not fit on a consumer card. If your plan is a local 4090, read the 24 GB column carefully before you start downloading weights.
The gateway token and the per-app unload behavior exist for a reason: between two apps with different base models, the gateway calls ComfyUI's free endpoint so nothing from the previous app stays in memory. That is a duct-taped solution to a real problem, and it means cold starts are part of the experience.
The template apps are the weakest part of the catalog. Seventy-two entries that are a fixed prompt over an existing pipeline will feel like padding to anyone who opens the folder expecting code. The 39 apps with their own workflows are where the engineering lives.
There is also a licensing question that the repository raises but does not fully answer. A character LoRA keeps a GPU busy for tens of minutes, the bill lands on someone's Modal account, and a LoRA of a person's face has to be stored per user and with consent. The README flags this as an open issue. It is the right instinct, and it is unresolved.
Why this one is worth your afternoon
The AI studio category is crowded with products that hide their internals and with open-source clones that show you a demo and a broken install script. Freefield sits in a third place: it exposes the plumbing, publishes the failure rate, and prints the bill next to the output.
That combination makes it useful even if you never run it. The VRAM map alone is a reference document for anyone planning a self-hosted media pipeline, and the per-app time and memory numbers are the data most vendors keep private because it makes their marketing look worse.
The lab's own framing is a good test to apply elsewhere. Every example is a real run. If a project cannot say that, ask what the examples are.
Related articles
Google replaced the model behind Search, and the number that moved was a shell benchmark
Gemini 4 Hydrogen became the default on 12 November, taking over from Gemini 3.8 Flash across Search, the Gemini app and the developer API. It runs about 2.5 times faster, scores 57% on Terminal-Bench 4.1 where its predecessor scored 18.2%, and costs $1.50 in and $7.50 out per million tokens.
Three ways a coding agent wastes money, measured on 30 repositories
TokenEconomizer watches what an agent did in a repository you work in often, then stops it from rediscovering the same build command, loading tools it never uses and looping on one error. Across five agent and model combinations it cut mean cost by 18.4% to 30.9%, cheaper in all 30 repository cells, with Pass@1 up in four of five.
earshot gives coding agents 922 sounds, and every one is rendered from code
An agent can write a game, an interface or a film and never hear any of it. earshot answers requests like "three variations of a footstep on gravel" from a catalogue of 922 designs built from public-domain recordings, with no audio files in the repository at all. It runs on CPU, as a CLI, a Python package, an MCP server and a Claude Code plugin.
The agents stealing API keys were not sophisticated. The gateways were open.
Amazon's decoy servers spent October posing as inference APIs and logged agents built on ARTEX pulling provider keys out of LiteLLM, Dify, Langflow and Open WebUI. The way in was a default master key and a half-finished install. ARTEX itself ships with budgets, scope gates and evidence for every finding.