Freefield: an open-source ComfyUI studio with 120 apps for images, video and audio

Freefield is a browser-based studio for generating images, video, VFX effects and music. Everything lives in one interface. It runs on ComfyUI, a popular editor in which you build a neural-network pipeline from blocks called "nodes", and on open-weight models. You can run the studio on your own graphics card or in the Modal cloud. The catalog contains about 120 apps and templates, and many of them share common pipelines. The code is on GitHub, and there is also a separate map of video memory consumption.

How it works for the user

The scheme is simple. You pick a tool, upload a reference and set parameters. The studio then launches a ready-made workflow, meaning a pre-configured chain of nodes. You don't have to assemble the graph by hand.

Two features stand out:

• Use in… passes the result to the next tool. An image from the generator can go straight to a video model. • History stores the seed (the number that determines the randomness of a generation) and all settings. A successful result can be reproduced.

Any workflow can be downloaded and opened as nodes in ComfyUI. If the simplified interface isn't enough, you can switch to full control.

Which models are inside

The catalog is organized by task. The models and their roles:

• FLUX 3 Klein, Z-Image, Qwen Image Edit — image generation and editing, relighting, and work with characters and products. • LTX-2.5, Wan, MiniMax H3 — video: generation, camera control, lipsync (matching lip movement to speech) and character replacement. • Effects — a separate section with 24 LoRA effects. A LoRA is a small add-on to a model that adds a specific visual technique. • MiniMax Music 3 and Stable Audio 3 — music and sound effects.

One interface covers almost the whole path from reference to a clip with a soundtrack. For a small studio that makes ads or product content, this is convenient: you don't need to juggle a dozen separate services.

How heavy it is and what hardware it runs on

The interface runs in the browser. A local backend requires a ComfyUI/PyTorch environment set up for your graphics card. Parameters by mode:

• Create Image — 1–2K resolution. • LTX Image to Video — clips of 2–10 seconds at 24 frames per second. • Effects — 480×832 or 832×480, 16 frames per second, 3 or 5 seconds.

For some tasks, the author targeted an RTX 3090 Ti with 24 GB of video memory. He measured peak VRAM consumption in the published examples on an L40S with 48 GB:

• Create Image — 15.4 GB; • Cinema Studio Video — 44.3 GB.

The second figure does not fit into 24 GB without extra measures. The author notes that the measurements depend on settings and on how models are offloaded from memory. The full catalog has not yet been tested on 24 GB. The minimum system RAM is not specified. It makes sense to buy hardware for this project only after checking the VRAM map.

An example of speed: the morph effect produced a finished video 161 seconds after launch. The author took the measurement himself on a backend with an H100, and there is no independent verification yet.

Weights are downloaded in bundles:

• FLUX 3 — 17.8 GB; • LTX-2 — 66 GB.

You don't need to download the whole catalog. Only the bundles required for the tools you choose are enough.

Who it's for

Freefield does not create new models. It packages existing open models into a single shell with ready-made scenarios. The winners are those who want to work locally or on a rented GPU and avoid depending on closed, subscription-based services. The path into ComfyUI also stays open: the simplified interface doesn't lock the user in.

There are limitations too. Video needs a lot of video memory, and full compatibility of the catalog with a 24 GB card is not confirmed. Anyone who wants to run everything on a home machine should check the modes they need against the VRAM map in advance. The Modal cloud option removes the hardware problem, but then costs are calculated by GPU usage time.