neuregex/freefield

You need Node 22.6+, Python 3.10+ and a Modal account, or a GPU of your own (below).

# 0. what is missing: Node, Modal login and secrets, Hugging Face license access per model, .env
python tools/doctor.py

# 1. secrets (once). Pick the Modal workspace first: in .env, FREEFIELD_MODAL_PROFILE=<profile from `modal profile list`>
python tools/setup_token.py                       # gateway token: Modal secret "freefield" and .env get the same value
modal secret create huggingface HF_TOKEN=hf_...   # accept the gated licenses on the Hub first (doctor.py checks them)

# 2. models into the "freefield-models" volume, per pack (see packs/index.json)
python tools/catalog.py packs
modal run infra/modal_app.py::download --pack image

# 3. backend (private: it needs the token)
modal deploy infra/modal_app.py                   # FREEFIELD_GPU=A10G to try the 24 GB plan

# 4. web app
npm install
# in .env (made by step 1): COMFY_BASE_URL = the URL printed by step 3
npm run serve                                     # http://127.0.0.1:3000
npm run app -- create-image --set prompt="a red bicycle against a white wall"   # or from the terminal

On Windows PowerShell, environment variables go before the command as $env:FREEFIELD_GPU = "A10G"; modal deploy ....

On your own GPU, without Modal

tools/local_backend.py runs the same backend as infra/modal_app.py, on your machine: ComfyUI at the pinned commit with the pinned custom nodes, comfy-api-proxy next to it and the token gateway in front. Run it in a Python virtualenv that already has the PyTorch build for your GPU (pytorch.org).

# 1. ComfyUI, the custom nodes and comfy-api-proxy (once)
python tools/local_backend.py setup ~/freefield-comfy

# 2. the models of the apps you want, by pack (packs/index.json says which apps each pack serves)
python tools/local_backend.py download ~/freefield-comfy image-flux3 --dry-run   # what is missing, and how big
python tools/local_backend.py download ~/freefield-comfy image-flux3             # gated repos: HF_TOKEN or `hf auth login`

# 3. a gateway token in .env, then the backend: ComfyUI :8188, comfy-api-proxy :8189, gateway :8000
python tools/setup_token.py --env-only
python tools/local_backend.py serve ~/freefield-comfy      # ComfyUI flags after --, e.g. -- --lowvram

# 4. in another terminal, with COMFY_BASE_URL=http://127.0.0.1:8000 in .env
npm install
npm run serve                                               # http://127.0.0.1:3000

Models go to <dir>/models; --models points at a model folder you already have instead. Which apps fit your card: the On a 24 GB card column of the app table and the VRAM map. Checked on 27 Oct on a CPU-only machine, setup to a finished run of the model-free smoke app through the gateway; the GPU apps send the same workflows as on Modal.

The web app

npm run serve opens a studio laid out like the one it rebuilds:

• Explore: featured apps as video cards, the studios, and one lab example per app.
• Pointing at Image, Video or Audio in the top bar opens a menu with every app of that studio, grouped as in the studio, and a column with the open models behind them and how many apps use each.
• Image and Audio: prompt-first apps. Results fill the page and a floating bar holds the prompt, reference images (click +, drop or paste them), one chip per setting, the number of generations and Generate, which shows how long the lab example took on the GPU.
• Video (and the upload-first tools): controls in a left column (Create video, Edit video, Motion control), results on the right.
• Every app has History (what you generated with it), Library (the lab example, Load this example) and How it works (models and licenses, VRAM by phase, the download and terminal commands, and Download the workflow: drop it on ComfyUI's canvas to open the app as nodes and chain it with others).
• A select marked multiple takes up to a few options at once and sends them in the order picked (Cinema Studio Video's camera moves are numbered; AI Influencer's facial features, distinctive traits and accessories). Apps that list each other under modes show a switch at the top of the form: Image / Video in Cinema Studio, New character / Your photo in AI Influencer.
• Use in… on any image or video result (in History and in the viewer) puts the file into another app and opens it: the usual next steps first (Image to Video, Motion Transfer, Relight, Upscale…), then every app that takes it.
• An image input marked paint in app.json (inpaint) opens a brush over the photo it names: what you paint becomes a white-on-black mask the size of the photo (web/js/paint.js).
• A thumbs-down on any lab example (Explore, Library, the viewer, the effects picker) marks it mediocre. The marks live in `lab/flags.

…