Free Tools

Fuobe: The Open-Source Creative Studio on ComfyUI — Images, Video, 3D and Audio on Your Own GPU

By Asad Ali ·

16 min read

Fuobe — the open creative studio on ComfyUI, with film, image, 3D cube and audio icons.

Fuobe is a free, open-source creative studio you run on your own computer. It puts image, video, 3D and audio generation into one workspace built on ComfyUI: you run open-weight models on your own GPU, or plug in your own fal.ai or Replicate key for frontier models. It is real, it is MIT-licensed, and it is also about a month old — so this post tells you exactly what it does, what it costs to run, and where the hype runs ahead of the software.

Fuobe's own description, from its GitHub README, is short: "The open creative studio on ComfyUI: images, video, 3D and audio on open weights, on your own GPU or the cloud." The project was created on 16 October 2026. When this post was written on 18 November 2026, the repository had 224 stars and 49 forks, with two contributors, and its last push was the same day. That is fast growth for a new project — the star count roughly doubled in a single day — but it is still a very young codebase, and you should judge it as one.

The short version: Fuobe is a genuine, working front end for ComfyUI with a large catalogue of ready-made workflows. It is not a magic one-click studio. The local route needs a real GPU, tens of gigabytes of model downloads, and some patience. The cloud route needs your own API keys and you pay per use. If that trade sounds fair, read on.

What Fuobe actually is

Fuobe is a web app (Next.js 16, TypeScript) that sits in front of ComfyUI. ComfyUI is the node-based engine that runs open image, video and audio models locally; Fuobe gives it a studio-style interface with menus, galleries, one-click apps and an AI agent, instead of a graph of nodes. You open Fuobe in a browser at http://localhost:3000, pick what you want to make, and Fuobe sends the job to an engine you choose.

The catalogue is the project's main selling point, and every number below comes from the README, checked on 18 November 2026:

  • 29 studios — grouped workspaces for images, video, 3D, audio, Cinema, Shorts, Marketing, Ads and more.
  • 132 apps — one-click tools for portraits, products, fashion, restoration, background changes, outfit swaps, relighting, upscaling and more.
  • 108 camera motions — dollies, cranes, arcs, orbits, drones and crash zooms; the README says you can mix any two in one clip.
  • 255 effects — weather, transitions, lens looks, film grades and transformations, for photos or clips.
  • 62 tools in Fuobe FX — a separate menu for FX work: recasting a plate, pulling a matte, cleaning it, extending, retiming, relighting and grading.
  • 3D in five modes — a model from one photo, several photos, a sentence, one object in a picture, or a soft Gaussian splat, plus Body Double, a 3D body for every person in a photo, with motion capture in six looks.
  • 268 API-format workflows, 191 of them validated against real node definitions — this is the number that matters most for the local route, because each workflow is a ComfyUI graph Fuobe can run.

Other pieces worth knowing about: Live edit, where the picture changes as you type; Kinetic Motion, motion-graphics templates drawn live in the browser in your brand; Marketing templates for product shots, UGC, TikTok, ads and posters; Layers, which splits a flat image into editable layers so you can change its text or repaint an area; Cinema, with a fly path through 3D scenes and a World mode you steer turn by turn; and voice and music tools, including Split soundtrack (separates a soundtrack into dialogue, music and effects) and Edit speech (changes the words in a recording).

How it works: two engines, one switch

Every Fuobe run goes to one of two engines, picked by a single switch in Settings:

  • Open source only. Every generation runs on open-weight models through your own ComfyUI, on your own GPU. Nothing is sent to a cloud provider. This is the free route — free in money, not in hardware or time.
  • Cloud. Frontier models run through your own fal.ai or Replicate key. You pay the provider's per-use prices; Fuobe itself charges nothing. The README notes that cloud runs are priced on the day they are planned, and pages show that day's price once they load.
  • Both. You can mix the two, and individual studios can fall back from one to the other.

With nothing connected, a run simply asks you to connect an engine. There is also a demo mode: turn on "Show placeholder results (for a tour)" in Settings and every run makes a labelled placeholder, so you can click through the whole studio without keys or a GPU.

The Assist agent

Assist is Fuobe's chat agent. You describe the whole job in plain words — say, a product shot turned into a short ad with a camera move and a caption — and Assist shows you the plan and its estimated cost before it runs anything. It only executes after you approve. Under the hood it can talk to a local model (Ollama, LM Studio, Unsloth Studio, or any OpenAI-compatible server) or, optionally, to Claude via an Anthropic key. That approval step matters: cloud jobs cost real money, and a plan you can see first is the difference between a tool and a bill.

Canvas

Canvas is the node-based workspace inside Fuobe. If the one-click apps are the easy path, Canvas is the workbench: you chain studios, effects and models into a pipeline — for example, generate a product image, apply a camera motion, extend the clip, then run a colour grade — and save the board to reuse it. Boards are saved with pages, and older Fuobe builds cannot read boards saved by newer ones, so the project advises backing up the data folder before upgrading.

Step-by-step setup

There are two setups: Fuobe itself, and the engine that does the work. Fuobe without an engine is a tour. Here is the full path, following the README's quick start.

Step 1 — Install the requirements

You need Windows, macOS or Linux, Node.js 22 or newer, and pnpm. On Windows, start.bat installs and starts everything for you. Check your versions first:

node --version   # needs v22 or newer
pnpm --version

Step 2 — Clone and start Fuobe

git clone https://github.com/workingprototype/fuobe.git
cd fuobe
pnpm install
pnpm dev          # opens at http://localhost:3000
# production: pnpm build && pnpm start

Step 3 — Connect an engine

Open Settings → Engine mode and connect what you have:

  • Local (free, needs hardware): a ComfyUI server. Fuobe ships scripts/comfy/install-comfyui.sh, which installs ComfyUI and its node packs at pinned commits. The library pins 52 node packs, each by commit. In Fuobe, the ComfyUI workflows page lists the model files each workflow needs and copies the exact install commands.
  • Cloud (needs keys, pay per use): a fal.ai or Replicate key, entered in Settings. No key, no cloud runs.
  • Nothing yet: turn on "Show placeholder results (for a tour)" to explore the interface with labelled placeholders.

Step 4 — Get the models (local route only)

This is the step the headlines skip. Open-weight models are files you download, and they are large. The README's own figures: the LTX-2.3 checkpoint is 29.53 GB plus a 9.45 GB text encoder; Wan VACE is 34.68 GB; a single round of new tools added about 182 GB of optional model files. You do not need all of them — the Workflows page lists what each tool needs — but a serious local setup means tens of gigabytes of downloads and the disk space to hold them. Plan for it before you start.

Step 5 — Optional: run your GPU from fuobe.com

An account on fuobe.com can run its jobs on the ComfyUI on your own computer, in three ways: start ComfyUI with a CORS header plus Fuobe's guard file (which keeps every other website out); install the optional Fuobe Connect add-on; or expose ComfyUI through a cloudflared quick tunnel, whose address changes each time the tunnel starts. To install Fuobe Connect, run this in ComfyUI's custom_nodes folder, then restart ComfyUI:

git clone https://github.com/workingprototype/ComfyUI-Fuobe-Connect

Step 6 — Pick a studio and run something small first

Start with one image app — Restore Photo, Product Shot or Upscale — on the open engine, and watch the first job end to end. Then try one motion, one effect, and one Assist plan with its cost estimate shown. If the first three jobs work, the setup is sound; if they fail, the failure message names the missing model or node pack, and the Workflows page gives you the install command for it.

Tip: the README's hardware line is "24 GB of VRAM covers almost every workflow" — not every workflow. Some need more: the LTX-2.3 route needs a 32 GB GPU, the fly path needs about 48 GB, and Kandinsky 6.0 Pro needs about 80 GB, all per the project's own manifest figures (which it labels as not measured). A 24 GB card is the honest starting point, not the ceiling.

Honest limitations and fine print

Fuobe is promising, but it is a one-month-old project with two contributors and 224 stars. Here is the honest version of the pitch, claim by claim.

What is claimedWhat is actually true
"A complete creative studio"A large catalogue (29 studios, 132 apps, 268 workflows) in one interface. 191 of the 268 workflows are validated against real node definitions — 77 are not yet validated, so expect some to fail.
"Free and open source"The code is MIT-licensed and free. The models are not all free to use however you like: weights, LoRAs and node packs carry their own licences, including noncommercial restrictions. Local runs cost electricity, hardware and disk; cloud runs cost per use.
"Runs on your own GPU"Yes — with a 24 GB GPU for most workflows, more for the heavy ones, plus tens of gigabytes of model downloads and 52 pinned ComfyUI node packs to install.
"Every preview rendered on one RTX 5090"True, per the README — but an RTX 5090 has 32 GB of VRAM and is a flagship card. Results on a smaller card will be slower, and some workflows will not fit at all.
"Licences are handled"Only partly. Licences are "information, not gates": everything is switched on by default, including noncommercial models, and a GPU-only LoRA can be sent to fal or Replicate with a note. You must switch things off yourself in Settings → Models & licences to stay commercial-only.
"A stable tool you can rely on"Not yet. The project is weeks old, moves fast (16 rounds of changes in its changelog), and upgrades can break saved work: Canvas boards saved by a new version cannot be read by older builds, and database migrations run automatically at start-up.

Fuobe vs the alternatives

OptionBest forThe catch
FuobeOne workspace for images, video, 3D and audio, local or cloud, with one-click appsVery young project; local route needs a strong GPU and large downloads
ComfyUI aloneFull control, no extra layer, huge communityNode graphs instead of a studio interface; you assemble every workflow yourself
fal.ai / Replicate directlyFrontier models with no local hardware at allPay per use, no unified studio, your media leaves your machine
Cloud video/image appsZero setup, polished results, subscriptionsMonthly cost, no local models, your work lives on someone else's servers

How a Fuobe job flows

How a Fuobe job flowsYour browserlocalhost:3000Fuobe serverstudios, apps, AssistEngine switchopen source / cloudComfyUI · your GPUopen weights, localfal.ai / Replicateyour key, pay per use→→→→Library — every result savedimages, clips, 3D models, audio↓ results

Security & privacy

Running your own studio is more private than a cloud app — but only if you set it up that way. The points that matter, all from the project's own docs:

  • Local runs stay local. On the open-source engine, your prompts, photos and videos are processed by ComfyUI on your own machine. Nothing is uploaded to a provider.
  • Cloud runs leave your machine. With a fal.ai or Replicate key, your media and prompts go to that provider under its own terms. That is the price of frontier models without a GPU.
  • Keys live in your install. Your cloud keys are stored in your own Fuobe setup, not on a vendor's account. Keep the install itself private: there is no admin role, and anyone who can sign in — or anyone who can reach a password-less install — can change settings, including licence switches and provider keys.
  • Do not expose it publicly without auth. The same warning applies to ComfyUI itself: it is a local tool, not a hardened public server. If you connect your GPU to fuobe.com, use the guard file or the Fuobe Connect add-on, which trusts fuobe.com by default and keeps other websites out.
  • Share links are tokenised. Fuobe stores only a hash of each share link's token, not the token itself.
  • Check licences before commercial use. As above, licences are information, not gates: noncommercial models are on by default. Before selling work made with Fuobe, open Settings → Models & licences and switch off anything that does not allow commercial use.
  • Face and voice tools need consent. Fuobe includes face swap, head swap, character swap, Body Double and voice tools. Swapping a real person's face or voice without their permission is a misuse of the tool, whatever the software allows.

Real use cases

Where Fuobe makes sense today, with its current maturity:

  • Product shots and ads for a small shop. Shoot one product photo, then use Product Shot, background changes and the Marketing templates to make listing images, a short ad clip with a camera motion, and a poster — all from the same source image.
  • Social Shorts from stills. Turn a photo into a clip with a dolly or orbit motion, add a film look, and export a vertical short. The Shorts studio is built for this.
  • Photo restoration and cleanup. Restore Photo, Upscale and relighting apps handle old family photos and low-resolution images without a subscription.
  • 3D product turntables. Image to 3D turns a product photo into a model or splat you can rotate in the browser — useful for a store page or a pitch deck.
  • Storyboards and animatics. The Storyboard studio and Cinema's fly path let a small team block out a scene before committing to a full render.
  • Voice and music fixes. Split soundtrack pulls a recording apart into dialogue, music and effects; Edit speech changes the words in a recording — handy for fixing a voiceover without re-recording it.

Where it does not make sense yet: as a production dependency for client work with deadlines (the project is weeks old and its changelog moves fast), on a laptop with a small GPU (many workflows will not fit), or for anyone who wants a one-click install with no model downloads.

Before you spend money: try the demo mode first, then run three small jobs on the open engine. If those work, add one cloud key and watch one Assist cost estimate before approving a real job. That is the cheapest way to find out whether Fuobe fits your machine.

Frequently Asked Questions

Is Fuobe really free?

The code is free under the MIT licence, and the open-source engine costs nothing beyond your own hardware, electricity and disk space. Cloud runs through fal.ai or Replicate are charged by those providers per use, and you supply your own keys. Model weights have their own licences, so "free" does not always mean free for commercial use.

Do I need a powerful GPU to use Fuobe?

For the local route, yes. The project says 24 GB of VRAM covers almost every workflow, and some workflows need 32 GB or more. Without a suitable GPU you can still use Fuobe through the cloud engine with a fal.ai or Replicate key, or explore it in demo mode with placeholder results.

Is Fuobe a replacement for ComfyUI?

No. Fuobe is a studio interface built on top of ComfyUI. The open-source engine is ComfyUI running on your machine, and Fuobe sends it workflows. If you already use ComfyUI, Fuobe adds menus, one-click apps, an agent and a library — it does not replace the engine.

Can I use Fuobe's output commercially?

It depends on the model that made it. Fuobe's code is MIT-licensed, but model weights, LoRAs and node packs have their own licences, including noncommercial restrictions — and Fuobe leaves them switched on by default. Check Settings → Models & licences and switch off anything that does not allow commercial use before selling work made with it.

How much disk space do the models need?

A lot, if you go local. Individual models run from under 1 GB to about 35 GB each, and the README's changelog mentions roughly 182 GB of optional model files added in a single round of updates. You only download what your chosen workflows need, and the Workflows page lists each tool's requirements — but budget real disk space, not a few gigabytes.

Is Fuobe safe to run?

As a local tool, yes, with normal care: keep the install private, do not expose it to the internet without authentication, and remember there is no admin role — anyone who can sign in can change settings. Its cloud path sends your media to fal.ai or Replicate under their terms, and its face and voice tools should only be used with the subject's consent.

Who is Fuobe for?

Developers, creators and small teams who already have a strong GPU or are happy paying per use in the cloud, and who want one workspace for images, video, 3D and audio instead of four separate subscriptions. If you want zero setup, a hosted app is still the easier choice.

Sources

All figures in this post were checked against the GitHub repository and its API on 18 November 2026: 224 stars, 49 forks, created 16 October 2026, MIT licence, two contributors. Star and fork counts move daily on a new project; treat them as a snapshot, not a fixed fact.

Want help setting up a self-hosted tool like Fuobe, or a website that shows off what you make with it? Asad Webs builds websites, web apps and AI-powered tools for businesses in the USA, UK, UAE, KSA and Spain. Get in touch and tell us what you are building.

Asad Webs has no fitting tool for AI image, video or 3D generation in its free tools collection — if you need one, that collection is where new tools land first.

Category: Free Tools

Tags: open-source · artificial-intelligence · comfyui

About Asad Ali

Asad Ali is a graphics designer with over 20 years of experience — designing since 2003 — and the founder of Asad Webs, based in Pakistan. He builds the free online tools on this site and writes the guides published here.

LinkedIn Facebook Website
KEEP READING

Next guide

Ghost Downloader 3: The Free Open-Source IDM Alternative →

Related articles

Join the conversation

No account needed. Comments appear only after approval. Your email address will not be published.

Be the first to share a useful question or thought.

You won’t be subscribed to any mailing list.
WELCOME BACK

Sign in

Access your orders, saved products and account.