ComfyUI

What ComfyUI is

ComfyUI is not an image generator in the sense that Midjourney is an image generator. It is the engine and the wiring diagram underneath one. You connect models, samplers, conditioning, upscalers and outputs as nodes on a canvas, and the graph you build is the product. It runs image, video, audio, 3D and text generation through the same interface, on Windows, Linux and macOS, across NVIDIA, AMD, Intel Arc and Apple Silicon, with Ascend NPUs, Cambricon MLUs, Iluvatar Corex and a CPU-only mode on the supported list too. The core is GPL-3.0 on GitHub with 131,400 stars, and the download page is unambiguous about the terms: "Free forever", with "No feature gates, no trial periods, no 'pro' tier for core functionality." If you own a GPU, ComfyUI costs nothing, and that is not a trial framing.

The commercial layer sits beside the engine rather than on top of it. Comfy Org raised $30 million in April 2026 at a $500 million valuation, led by Craft Ventures with Pace Capital, Chemistry and TruArrow participating, and puts its total funding at roughly $47 million. It sells things the free engine does not include: Comfy Cloud (managed GPUs, from $20 a month), Comfy Enterprise, a Developer Platform still labelled beta that bills GPUs by the second, Partner Nodes that call closed models such as Kling, Veo, OpenAI, Luma, ElevenLabs and Seedance from inside a graph on a prepaid credit balance, and a resold MiniMax commercial licence. The vendor claims 4 million users, 60,000-plus community-built nodes and 150,000-plus daily downloads. The structural point for anyone comparing this to a subscription product: because the engine is GPL-3.0, the company cannot sell the cloud by crippling the free product. The cloud sells you a GPU you do not own.

The cost of that power is the learning curve, and it is steep in a way marketing pages do not convey. A first workflow means understanding checkpoints, VAEs, CLIP encoders, samplers, schedulers and latent space as separate objects you wire together yourself. Most useful workflows people share depend on third-party custom nodes, which are unvetted Python packages from strangers, and which are the single most common cause of a ComfyUI install that used to work and now does not. There is an App Mode that collapses a graph into a simple form view, and it genuinely helps someone running a workflow they did not build. It does not help the person building it.

Pricing

The engine is free. Everything with a price attached is Comfy Cloud or something adjacent to it. All figures below were read from comfy.org/pricing and support.comfy.org in September 2026 and are in US dollars. The one number that ties the whole page together is the credit rate, which the support centre states as "$1 = 211 credits", and which reproduces exactly from every GPU row on the pricing page.

The three individual tiers are exactly 20% off on annual billing: $240, $420 and $1,200 at list become $192, $336 and $960. A 30% GPU price cut was applied on 23 January 2026, so any figure older than that is stale.

Team is a range, not a price

The card on the pricing page shows Team at $700 a month, and it is easy to read that as the price. It is the middle stop on a slider. The vendor's FAQ says it plainly: "Team plans come in five tiers from $200 to $2,500/month, set by a credit-commit slider." Team is also the only plan with seat language on the page, and it runs the other way: "Team pricing is based on your monthly credit commit, not per-seat", with membership "capped at 50 today".

One thing on this page we could not reconcile, so we are stating both halves. The FAQ says "A bigger monthly commit means a bigger discount: annual plans go up to 20% off, monthly plans up to 10%", starting at the $400 stop. But every stop we could price computes to exactly 211 credits per dollar, the same rate as the individual tiers, which is not what a monthly volume discount looks like. The public page does not resolve that. If the Team discount matters to you, get it in writing on your own quote.

What a credit actually buys on Comfy Cloud

Comfy Cloud runs on Blackwell RTX 6000 Pro cards with 96 GB of VRAM and, per the support centre and the vendor's own blog, consumes 0.266 credits per second of active GPU time, down from 0.39 before the January cut. That is 957.6 credits an hour, or $4.54 an hour at 211 credits per dollar. Building and editing a graph costs nothing; only the run is metered. Two consequences follow. First, the plan ladder for individuals is not a volume discount: $100 buys 21,100 credits at list and Pro grants exactly 21,100, so a higher tier buys features (LoRA import, a longer runtime cap, more concurrent API jobs), not cheaper compute. Second, Cloud is not the cheapest way to rent the same card from the same company. The Developer Platform table on the same page, captioned as applying to the Comfy API, lists an RTX PRO 6000 at $3.49 an hour, so Cloud carries roughly a 30% premium for the convenience of a managed workflow environment.

Credit expiry is split by source. Monthly plan credits reset each billing cycle and do not roll over. Purchased top-up credits are valid for one year. Both draw from one balance, and that balance is shared between Comfy Cloud and supported usage in a local ComfyUI install, which is how a "free" local graph calling a Partner Node ends up spending Cloud credits.

Developer Platform (beta), billed by the GPU second

These rates apply to the Comfy API on the Developer Platform, not to Comfy Cloud subscriptions. They are the only place the company publishes a dollar-per-hour figure for its hardware, and the credits column is the 211-per-dollar rate applied to each one.

MiniMax Commercial License

Comfy also resells a MiniMax commercial video licence on its pricing page. The Professional tier is from $5,000 a month on a monthly term, with about 46,250 video-seconds included, $0.108 per video-second inside the bundle and $0.036 per second of overage, for up to 10 licensed users on one domain. The Enterprise tier is contact-sales with no cap on users and a 12-month minimum. This is a separate product for studios that need MiniMax output cleared for commercial use, not something a Comfy Cloud subscriber pays.

What we could not see

Top-up credit pack sizes and prices sit behind a logged-in purchase modal, so we are not quoting them. Comfy Enterprise pricing is not published. The fourth Team stop's price is not printed on the card we could read. And whether an annual plan releases credits month by month or all at once is not stated outright, although the page's repeated wording of "monthly credits" that "reset each billing cycle" points to monthly.

What we tested, and what we did not

Reviews of a tool this large often imply a score covers hands-on use. This one does not, so here is the boundary in plain terms.

Verified: every price and plan rule on comfy.org/pricing, the credit rate and consumption rate on support.comfy.org and the vendor blog, the licence and star count on GitHub, the hardware and operating system list on docs.comfy.org, the free-forever wording on the download page, and the April 2026 funding round against TechCrunch. Every arithmetic claim on this page can be reproduced from those sources in a minute.

Not tested: installation on any machine, generation quality or speed, Comfy Cloud queue times, custom node stability across a core update, App Mode, Partner Node output, or the Developer Platform API. No workflows were built and no generations were run.

That means the pricing and the product structure are measured; the experience is reasoned. The score below reflects what we verified and takes the usability and reliability marks from documented design and the ecosystem's structure, not from a benchmark.

What the credits buy, in GPU hours

Because every plan is credits at the same rate, the honest way to compare tiers is hours of GPU time. Hours are monthly credits divided by 957.6 credits an hour, which is 0.266 credits a second. The advertised video counts are the vendor's, from a preset its FAQ names: the Wan 2.2 image-to-video template at default settings, 81 frames, 18 fps, 640x640, a 4-step sampler, about 11 credits per clip.

Hours are ToolChase arithmetic on vendor-published figures, rounded to one decimal. The advertised video counts are reachable only at the named preset. At 720p and a normal step count expect a small multiple of the per-video cost, and a small fraction of the count.

Where ComfyUI is the strongest pick

ComfyUI is at its strongest when the output has to be repeatable. A production shot, a product-photo batch or a house style distributed across a team is a graph file, and a graph runs the same way the hundredth time as the first. Hosted generators give you a prompt and a slider; this gives you every intermediate step, so generate, segment, inpaint a region, upscale and interpolate frames can be one pipeline rather than five tools and a folder of exports.

The second argument is where the models come from. New open-weight image and video models almost always land in ComfyUI first, often within days, because the community ships nodes faster than any hosted platform ships integrations. If you want to test a model the week it drops, or run one offline because the material is client-confidential, there is no hosted substitute. And the local install is genuinely the full product: a free tier here is not a watermarked sample of a paid one.

Best for

Technical creators, VFX and production artists, and small studios who need a repeatable multi-stage pipeline rather than a one-shot generation, and who either already own a capable GPU or are willing to rent one by the hour. It is also the right pick for anyone who must run open-weight models offline for confidentiality reasons, and for developers prototyping a generative feature before committing to a vendor API. It is the wrong pick for someone who wants to type a sentence and get a good image: Midjourney, KREA or a hosted Stable Diffusion front end will serve that person far better and in a fraction of the time. It is also wrong for teams with no one willing to own the install, because somebody has to maintain the custom node dependencies and that job does not go away.

Key features

• Node-graph canvas where models, samplers, conditioning, upscalers and outputs are wired together explicitly, with the graph itself as the reusable, shareable artefact
• One engine across image, video, audio, 3D and text generation rather than separate tools per medium
• 60,000-plus community-built custom nodes and thousands of shared workflows (vendor figures)
• Runs on NVIDIA, AMD, Intel Arc and Apple Silicon, plus Ascend NPUs, Cambricon MLUs, Iluvatar Corex and CPU-only, on Windows, Linux and macOS
• Comfy Desktop installer for Windows and macOS, a portable Windows package, and a CLI install from GitHub for developers
• App Mode, which collapses a finished graph into a simple form view for people who need to run a workflow but not build it
• Partner Nodes that call closed commercial models (Kling, Veo, OpenAI, Luma, ElevenLabs, Seedance and others) from inside a local graph on a prepaid credit balance
• Comfy Cloud: managed Blackwell RTX 6000 Pro GPUs with 96 GB of VRAM, metered at 0.266 credits a second and billed only for active GPU time
• Custom model and LoRA import on Cloud Creator, Pro and Team tiers
• Developer Platform API (beta) with published per-second GPU pricing across RTX PRO 6000, H100, H200 and B200
• Workflows run entirely offline on a local install once models are downloaded, with no account or network dependency for open models

Best-fit use cases

• Building a repeatable, deterministic pipeline for a production shot or a product-photo batch, where the same graph has to run identically a hundred times
• Running open-weight models such as Flux, Wan or Stable Diffusion locally on hardware you already own, with no per-generation cost
• Multi-stage work that hosted generators cannot express: generate, segment, inpaint a region, upscale, then interpolate frames, as one graph
• Testing a new open-weight model the week it drops, before any hosted platform has integrated it
• Prototyping a generative feature before committing to an API, then swapping a Partner Node for a local model once the pipeline is proven
• Studios standardising a house look across a team by distributing one workflow file rather than a page of prompt instructions

Pros

• The local engine is genuinely free and GPL-3.0, and the vendor's download page states "No feature gates, no trial periods, no 'pro' tier for core functionality"; the paid tiers sell hosting, not unlocks
• The most capable open generative media engine available: image, video, audio and 3D in one graph, with control over every intermediate step that hosted tools hide
• Comfy publishes its credit rate, its per-second consumption rate and its Developer Platform GPU hourly rates, which most managed-GPU vendors in this category will not do
• Billing is for active GPU time only; building and editing a workflow consumes nothing
• Broadest hardware support in the category, including AMD, Intel Arc and Apple Silicon, not NVIDIA-only
• Runs fully offline on local models, so sensitive or client-confidential material never leaves the machine
• New open-weight models usually land here first, often within days, because the community ships nodes faster than any hosted platform ships integrations
• A workflow is a file, so a pipeline is versionable, diffable and shareable in a way a prompt is not
• Financially stable for an open-source project: $30 million raised in April 2026 at a $500 million valuation, roughly $47 million in total by the vendor's count

Cons

• The learning curve is the steepest in the category by a wide margin; a first working video workflow is a day of reading, not an afternoon
• The vendor publishes no minimum VRAM figure anywhere we could find. The entire stated guidance is that a dedicated GPU is strongly recommended, which is useless to someone deciding whether their machine can run it
• Custom nodes are unvetted third-party Python from arbitrary authors, executed with full local privileges; they are both the main source of capability and the main source of broken installs
• Comfy Cloud's "~380 5s videos" style claims are derived from a deliberately cheap preset (640x640, 4 steps), not from settings anyone would ship work at
• Comfy Cloud costs about 30% more per GPU hour than the same company's Developer Platform rate for the same class of card
• The individual plan tiers give no volume discount at all; credits cost the same at $20 and at $100, so upgrading buys features, not cheaper compute
• Monthly plan credits do not roll over; only separately purchased top-ups persist
• The Team plan's advertised commit-based discount cannot be reconciled with the slider's own arithmetic from the public page
• Partner Nodes mean a "free forever" local install starts costing money the moment you wire in a closed model, drawing on the same balance as Cloud
• Support is a help centre, ticket forms and community channels; no support SLA is differentiated across the paid Cloud tiers below Enterprise
• GPL-3.0 is a real constraint if you intend to embed the engine in a commercial closed-source product

What to watch out for

Four things, in order of how much they will cost you.

Budget Comfy Cloud from the per-second rate, not the video count. The plans advertise "~380 5s videos" on Standard, and that figure is reproducible, but only from the preset the vendor's own FAQ names: the Wan 2.2 image-to-video template at default settings, 81 frames, 18 fps, 640x640, a 4-step sampler, roughly 11 credits a run. Almost nobody ships at 640x640 with four steps. The honest way to plan is 0.266 credits a second: 4,200 credits is 4.4 hours a month on Standard, 7.7 hours on Creator, 22 hours on Pro. Those are the numbers your spend actually follows, and they are arithmetic from figures Comfy publishes itself. If you are running 720p at a normal step count, expect a small fraction of the advertised video count.

Nobody will tell you if your GPU is big enough. This is the single most common reason a new ComfyUI user bounces, and the vendor does not address it. The download page says a dedicated GPU is strongly recommended and that more VRAM means bigger models. The system requirements documentation lists supported vendors, operating systems and a Python version but states no minimum VRAM, no minimum RAM and no disk figure. In practice VRAM is the binding constraint on which models you can run at all, and you will be working it out from community threads rather than from the vendor.

Custom nodes are the ecosystem and the liability at once. The 60,000-node figure is the reason ComfyUI can do things nothing else can, and it is also the reason installs break. These are third-party Python packages running with full local privileges, unreviewed, on your machine. A shared workflow that needs nine custom nodes is nine dependencies from nine strangers, any of which can break on a core update or on each other. Treat a ComfyUI install the way you would treat any machine running unaudited code, and pin your versions before a deadline rather than after one.

"Free forever" applies to the engine, not to everything in it. The local install genuinely costs nothing and is not crippled. But Partner Nodes calling Kling, Veo, OpenAI or the rest bill against a prepaid credit balance, and that is the same balance Comfy Cloud uses. A workflow you downloaded can contain one of these without it being obvious, so check before you run someone else's graph and assume it is free.

Is ComfyUI worth it?

Yes, for the buyer it is built for, and for that buyer there is no substitute at any price. The local engine is the full product, free under GPL-3.0, and it is the only place image, video, audio and 3D generation sit in one controllable graph on hardware you own. Comfy Cloud is a fair way to get the same engine without a GPU: the pricing is published to the credit and the consumption rate is published to the second, which is more than most managed-GPU vendors manage. It is not cheap, at $4.54 an hour against $3.49 for the same class of card on the company's own API, and the tier ladder buys features rather than volume.

It scores 4.3 rather than higher for reasons that have nothing to do with capability. A tool most buyers cannot successfully install should not sit alongside category leaders that anyone can pick up, and the vendor compounds the onboarding gap by publishing no minimum VRAM anywhere, leaving the most basic buying question to community forums. Scoring it lower would be wrong too. If you can run it, nothing else does what it does.

Worth saying plainly: no generations were run for this review. The usability and reliability marks are reasoned from documented design and the custom node ecosystem's structure, and the score can move in either direction once someone measures them.

How we arrived at 4.3

The score is weighted across eight categories. The Value line was revised down during fact-checking after we found that the GPU hourly table on the pricing page applies to the Developer Platform API, not to Comfy Cloud, which bills a higher rate for the same hardware.

The 5 point score is a product rating rather than a linear map of that total, and it reflects the most capable engine in the category held back by the steepest learning curve in it.

FAQ

Pricing, plan rules and feature claims verified September 2026 from comfy.org/pricing, comfy.org/download, support.comfy.org, docs.comfy.org, the Comfy blog and the ComfyUI GitHub repository. No hands-on testing was performed for this review. ToolChase is not an affiliate of Comfy Org and received no payment for this review. Scores, rankings and list positions are never sold. See our methodology for how we score tools.