ComfyUI-freefield: 117 Ready-Made Pipelines Instead of a Higgsfield Subscription, All Running Locally

If you have ever built a video pipeline in ComfyUI, you know the pain: finding the right workflow, guessing which models it needs, sorting them into folders, and hoping nothing breaks. Cloud services like Higgsfield solve this with a convenient interface, but they charge a subscription and keep you in their queue. The Freefield project tries to offer the same convenience for free, on your own machine.

What's inside

The ComfyUI-freefield extension adds a dedicated tab to the left panel of ComfyUI with 117 ready-made apps. They are organized into four studios: Image (70 apps), Video (46), plus VFX and Audio. These include video generation and animation, upscaling, audio work, visual effects and face swapping.

An important point: these are not black boxes. Each app unfolds into an ordinary, editable graph built on standard ComfyUI 0.42 and newer nodes. There are no proprietary binary components, and only open weights are used. You can take a graph apart, swap the model, save it to a personal library or send it to a colleague.

How it works in practice

Every pipeline has a card. On it, a single button downloads all the necessary models and LoRAs from Hugging Face and places them in the right folders: diffusion_models, text_encoders, loras. If a download is interrupted, it resumes from where it stopped thanks to .part files. For gated models, such as FLUX 3 Klein or LTX-2.5, your HF_TOKEN from the system is used.

Then there are two modes. "Open with example" opens the graph with test files and parameters from the Freefield lab so you can immediately check that everything works. "Open empty" gives you a clean topology for your own material. For developers and product managers, this is a quick way to understand what local generation can actually deliver without building everything from scratch.

Installation is simple: through ComfyUI Manager (search for Freefield) or with the command comfy node install freefield. The code is on GitHub under the MIT license, so it can be used in commercial projects too.

Limitations

The bold claim that it "replaces Higgsfield" should be read soberly. The authors themselves successfully ran 72 of the 117 workflows, just over half. The test machine is serious: Windows, an RTX 3090 Ti with 24 GB and 104 GB of RAM. Heavy processes such as Dub (135 s) and Cinema Studio Video (144 s) are pulled off thanks to dynamic VRAM management, which keeps extra weights in system memory. On a card with 8–12 GB and modest RAM, expectations should be lowered.

Another note: the claim that "Higgsfield is wildly nervous" in the original post is the channel author's emotion, not a fact. The source contains no comparisons of output quality with the commercial service. Most likely, the cloud product is still ahead in polish and stability, while Freefield's advantage lies in control, price and privacy of your material.

Practical takeaway

It is worth trying if you make content or prototypes with generative video and have a powerful card. Start with the "Open with example" mode on a few pipelines, see which ones actually run for you, and estimate the time per frame. For teams that pay for cloud subscriptions and run into queues, or don't want to send their material outside, this is a chance to move some tasks to their own hardware. It is also a ready-made library of graphs that is handy for learning: you can study how complex chains are built and assemble your own.