Introduction

Higgsfield promised to open source its platform, and one builder simply went ahead and did it. Freefield, launched on October 27 by developer neuregex as an entry in the #OpenSourceHiggsfield challenge, is an MIT-licensed web studio laid out like Higgsfield. It holds image, video, VFX, and audio apps, and every single one is a ComfyUI workflow running on open weights. There are no API or partner nodes anywhere. You can run it today on your own GPU or deploy the backend to Modal.

The headline number is 117 apps covering 95 of Higgsfield's 129 tools. Each app is just a folder: an app.json manifest plus a ComfyUI workflow in API format. The web forms, model downloads, VRAM map, and tests are all generated from those manifests, so there's no per-app code at all.

What's inside

The repo's status table currently lists 120 apps, with 119 marked ready and a MOSS-Image layout app still in draft. That count is 39 apps with their own workflow, five IC-LoRA apps, a morph app, 72 template apps, and two model-free smoke tests. The template apps are Higgsfield-style one-trick tools (Packshot, Billboard Ad, Plushies, Giallo Horror, Face Swap, ASMR Host and friends), each a fixed prompt baked over one of the core pipelines.

The model lineup

  • Create Image, Edit, Inpaint, Soul ID-style characters, Moodboard, Product Shot: FLUX 3 Image Klein (int8), with BFL's 9B prompt upsampler as a toggle in Create Image. Fits whole on a 24 GB card at about 11–17.8 GB.
  • Fast drafts: Z-Image Turbo (int8), about 12.2 GB.
  • Styles and style reference: Krea 2 Turbo plus Krea's nine official style LoRAs.
  • Relight: Qwen Image Edit 2509 (fp8) with Relight and Lightning LoRAs.
  • Create Video with audio, VFX, transitions, Genjutsu, Reframe, SDR-to-HDR, Dub, Speak: LTX-2.5 22B distilled plus Lightricks' official IC-LoRAs, tight on 24 GB at about 21.5–22.8 GB.
  • H3 video, Extend Video, Reshoot: MiniMax H3 (int8) from Comfy's templates, which needs Modal because the text encoder alone is 27.1 GB.
  • Character swap and motion transfer: SCAIL-2 with SAM 3.1 tracking, and Wan Animate 2.
  • Camera moves and effects: Wan 2.2 Fun Camera and Wan 2.1 i2v with 24 effect LoRAs, 20 of them by Remade-AI.
  • Audio: MiniMax Music 3 and Stable Audio 3 medium.
  • Utilities: SeedVR2 upscaling, BiRefNet background removal, MoGe-2 depth maps, and FILM frame interpolation.

Receipts, not vibes

What makes Freefield feel trustworthy is its lab log. The README reports 48 GPU runs on Modal between October 20 and 26 on L40S and H100 cards, with 710 of 733 cases passing. Every example in the web UI comes from one of those runs, with its time and peak VRAM attached. An end-to-end browser test on an H100 clocked the morph app at 160.8 seconds from button to finished video, mixed-media at 149 seconds, and dub at 72 seconds.

There's also a lovely little bake-off. Using the same seed on six prompts, the author found FLUX 3 Klein kept object counts and rendered label text where the other candidates slipped, which is why Klein now drives Create Image, Cinema Studio, and the meme app.

Soul ID without training

Higgsfield's Soul ID trains a model of your face. Freefield skips training: the character and character-3 apps copy a person from one or three reference photos on every run, at about 13 seconds of L40S time per image in the lab. A LoRA-training path built on ComfyUI 0.42's own TrainLoraNode is sketched in the README but not shipped yet.

Run it locally

You need Python 3.10+, Node 22.6+, and a PyTorch build for your GPU. Some models are gated on Hugging Face, so accept their licenses and keep an HF_TOKEN handy.

git clone https://github.com/neuregex/freefield && cd freefield

# 1. ComfyUI at the pinned commit, the custom nodes, and comfy-api-proxy
python tools/local_backend.py setup ~/freefield-comfy

# 2. download the models for the apps you want, by pack
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, then the backend (ComfyUI :8188, proxy :8189, gateway :8000)
python tools/setup_token.py --env-only
python tools/local_backend.py serve ~/freefield-comfy

# 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

Prefer the cloud? The README's Modal path (python tools/doctor.py, then modal deploy infra/modal_app.py) puts the same backend on rented GPUs, which is the route for the H3 apps that won't fit on a 24 GB card. And tools/import_workflow.py drafts a brand-new Freefield app from any ComfyUI workflow you already love.

Try a first prompt

Once the server is up, you can fire the Create Image app straight from the terminal:

npm run app -- create-image --set prompt="a red bicycle against a white wall"

Why it matters

Hosted creative platforms win on convenience: a big menu of one-click effects. Freefield shows that menu is mostly a handful of strong open pipelines wearing different prompts. Every model is a downloadable file listed with its Hugging Face repo, size, and license inside its app's manifest, so you can see exactly what you're running and where its license draws the line. Do check those per-model licenses before commercial work, because the MIT license covers Freefield's code, not the weights.

Original Source

Conclusion

Freefield is the most complete answer yet to the #OpenSourceHiggsfield dare: a full creative studio you can read, fork, and run on your own card, with lab receipts for every example. Clone it, start with the image pack, and build your own app from a workflow you already have.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=freefield-open-source-higgsfield-comfyui-117-apps

—Aurelia ♡