Three ways to pay
Every service here runs the same ComfyUI. What differs is what you pay for: the seconds a workflow runs, the whole time a machine is on, or a bare GPU you set up yourself.
Pay per generation
You pay only while a workflow runs. Building, editing and waiting are free.
- Who
- Comfy Cloud, Floyo, ComfyICU
- Good
- No idle bills, nothing to install, files kept for you.
- Watch
- A monthly plan to start (Comfy Cloud from $20, Floyo from $15, ComfyICU from $10), a curated set of nodes, run time caps, and a higher price per hour.
Rent a ComfyUI machine
A machine of your own with ComfyUI ready, billed from when it is ready until it stops.
- Who
- RunComfy, ThinkDiffusion, MimicPC
- Good
- Any custom node, your own models, a saved setup.
- Watch
- Every idle minute counts until you stop it or a timer does.
Rent a raw GPU
A GPU server by the second. You install ComfyUI, or start from a template.
- Who
- RunPod, Vast.ai
- Good
- The lowest price per hour, and the whole machine.
- Watch
- You maintain it, storage is billed separately, and nothing documented stops it for you.
What a session costs
Most of a ComfyUI session is not generating: you load models, wire nodes, change a prompt, look at the result. The calculator shows what that costs on each service. It opens on a 2-hour session in which the GPU generates 25% of the time; move the sliders to match how you work.
Published hourly rates, read 9 October 2026 (RunPod 19 October 2026, Comfy Cloud 16 October 2026), before tax.
Cheapest for this session: RunPod Community at $0.68. Comfy Cloud costs $2.27; it beats a RunPod RTX 4090 only while the GPU generates less than 20% of the session.
| Service | This session | A month |
|---|---|---|
| RunPod Communitythird-party hostRTX 4090 · $0.34/h · raw GPU | $0.68 | $6.80 |
| ComfyICUplans onlyL4 · $3.24/h · per generation | $1.62 | $16.20 |
| RunPodRTX 4090 · $0.89/h · raw GPU | $1.78 | $17.80 |
| MimicPCL4 or A10 · $0.99/h · per session | $1.98 | $19.80 |
| Comfy Cloudplans onlyRTX PRO 6000 · $4.54/h · per generation | $2.27 | $22.70 |
| RunComfyA10G or A5000 · $1.75/h · per session | $3.50 | $35.00 |
| ThinkDiffusion24 GB machine · $1.75/h · per session | $3.50 | $35.00 |
| Floyoplans onlyH100 NVL · $9.00/h · per generation | $4.50 | $45.00 |
Each service is shown on one current card of at least that size. Older cards cost less and run slower: an RTX A5000 is $0.27 an hour on RunPod. Services marked "plans only" sell monthly plans (Comfy Cloud from $20, Floyo from $15, ComfyICU from $10), so their rows show what a session uses from a plan, not a bill. Raw GPUs do not include storage: 100 GB on a RunPod network volume adds $7.00 a month. RunComfy and ThinkDiffusion Pro take up to 20% off their hourly rates.
The pattern holds at every size. Paying per generation wins when your GPU mostly waits: against a RunPod RTX 4090, Comfy Cloud is cheaper while the GPU generates less than 20% of the session, and against an A10G or A5000 on RunComfy, less than 39%. Comfy Cloud's RTX PRO 6000 is also a faster card than an RTX 4090, so the same job takes less time there, which moves the line further in its favour. Once you queue long video jobs back to back, a machine billed by the hour is cheaper, and a raw GPU cheapest.
Comfy Cloud pricing
Comfy Cloud sells plans paid in credits; 957.94 credits buy one hour of generation on its RTX PRO 6000, counted by the second.
| Plan | Monthly | Billed yearly | Generation a month |
|---|---|---|---|
| Standard | $20 | $16 a month | about 4.4 hours |
| Creator | $35 | $28 a month | about 7.7 hours |
| Pro | $100 | $80 a month | about 22 hours |
| Team 147.7K | $665 | $630 a month | about 154.2 hours |
Plan credits do not roll over; top-ups last a year. Runs stop at 30 minutes, or an hour on Pro. Team, for up to 50 members who run workflows at the same time from one shared credit pool, comes in five sizes from $200 a month; the table shows the size the pricing page opens on. Your own models and LoRAs, imported from Civitai or Hugging Face, need Creator or above. Five free runs come first, with no card.
Which card your workflow needs
The memory each model's own ComfyUI template needs at its largest stage, the smallest card size that holds it with room to work (under 85% of memory, the site's rule for image and video), and the cheapest hourly rate here for a current card that size. Working memory for generation comes on top and grows with resolution, which is why a tight fit is the wrong pick for video.
| Model | Template peak | Card size | Raw GPU | Managed machine |
|---|---|---|---|---|
| AnimaBase v1 | 4.4 GB | 16 GB | $0.89/hRunPod RTX 4090 | $0.49/hMimicPC T4 |
| FLUX.2klein 4B | 8 GB | 16 GB | $0.89/hRunPod RTX 4090 | $0.49/hMimicPC T4 |
| Qwen-Image-2.1Text to image | 9.5 GB | 16 GB | $0.89/hRunPod RTX 4090 | $0.49/hMimicPC T4 |
| FLUX 3 Image KleinKlein (distilled) | 9.8 GB | 16 GB | $0.89/hRunPod RTX 4090 | $0.49/hMimicPC T4 |
| Z-ImageTurbo | 12.7 GB | 16 GB | $0.89/hRunPod RTX 4090 | $0.49/hMimicPC T4 |
| Krea 2Turbo | 13.4 GB | 16 GB | $0.89/hRunPod RTX 4090 | $0.49/hMimicPC T4 |
| Wan 2.2Text to video, A14B | 14.5 GB | 24 GB | $0.89/hRunPod RTX 4090 | $0.99/hMimicPC L4 or A10 |
| HunyuanVideo 1.5720p | 19.2 GB | 24 GB | $0.89/hRunPod RTX 4090 | $0.99/hMimicPC L4 or A10 |
| Qwen-Image-Edit-2511Image editing | 20.8 GB | 48 GB | $1.09/hRunPod L40S | $1.99/hMimicPC L40S |
| LTX-2.5Distilled | 24.3 GB | 48 GB | $1.09/hRunPod L40S | $1.99/hMimicPC L40S |
| MiniMax H3Text or frames to video with audio | 24.4 GB | 48 GB | $1.09/hRunPod L40S | $1.99/hMimicPC L40S |
Each model's page lists every file at every precision and which one loads on which card: start from the image and video models, or check a model against a card in the ComfyUI VRAM checker. For a workflow of your own, the ComfyUI workflow checker lists its files, the memory it needs and the custom nodes a hosted service would have to allow. Comfy Cloud and Floyo have enough memory for every one of these; check their node and model lists for the ones you use.
Price per hour by card
At each service's pay-as-you-go rate, or its plan rate where there is no other, before tax, on 9 October 2026 (RunPod 19 October 2026, Comfy Cloud 16 October 2026). Where a paid plan lowers the rate, the plan rate is shown beside it. ComfyICU's hourly figures are worked out from its per-second credits.
| Service | 24 GB | 48 GB | 80 to 96 GB |
|---|---|---|---|
| Comfy CloudPer second of generation, plans only | - | - | $4.54RTX PRO 6000 |
| FloyoPer second of generation, plans only | - | - | $9.00H100 NVL |
| ComfyICUPer second of generation, plans only | $3.24L4 | $5.76L40S | $11.52H100 |
| RunComfyPer session, idle included | $1.75 · $1.39 planA10G or A5000 | $2.50 · $1.99 planA6000 | $4.99 · $3.99 planA100 |
| ThinkDiffusionPer session, idle included | $1.75 · $1.49 plan24 GB machine | $2.50 · $1.99 plan48 GB machine | - |
| MimicPCPer session, idle included | $0.99L4 or A10 | $1.99L40S | - |
| RunPodRaw GPU, per second | $0.89RTX 4090 | $1.09L40S | $2.49RTX PRO 6000 |
| RunPod CommunityRaw GPU, per second, third-party host | $0.34RTX 4090 | $0.79L40S | $1.69RTX PRO 6000 |
- Comfy Cloud: Standard $20, Creator $35, Pro $100 a month (up to 20% off yearly), and Team from $200 a month in five sizes for up to 50 members sharing one credit pool, paid in GPU credits; plan credits do not roll over, top-ups last a year. One card only, the 96 GB RTX PRO 6000.
- Floyo: Explorer $15, Pathfinder $35, Trailblazer $89 a month ($144, $336, $852 a year), with 2, 5 and 14 GPU hours; $9 an hour beyond.
- ComfyICU: Subscription only: $10 to $240 a month in credits that do not roll over; L40S from Standard ($30), H100 from Pro ($60).
- RunComfy: Pay as you go, or Pro at $19.99 a month (listed at $29.99; $239.90 a year) for at least 20% off, a $10 monthly credit and 200 GB kept. ThinkDiffusion: Pay as you go, or TD-Pro at $29.99 a month ($19.99 billed yearly) for up to 20% off, a $10 monthly credit and 200 GB.
- MimicPC: Essential $13.95 or Advanced $26.95 a month, with $12 or $25 of credit and 50 GB more storage; interruptible machines are cheaper.
- Vast.ai: A marketplace where hosts set their own prices, billed per second. Its terms do not allow its prices in a comparison, so this page leaves them out; see its live prices.
For comparison, RunPod sells Wan 2.2 as a plain model API, outside ComfyUI, at $0.30 for a 5-second 720p clip, and Comfy Cloud's own example, a 5-second clip at 640×640, works out at about $0.05 of generation time. The two are not like for like: 720p has 2.25 times the pixels of 640×640, and Comfy's example is a four-step run.
Nodes, models and files
| Service | Custom nodes | Your own models | What is kept | Free to try |
|---|---|---|---|---|
| Comfy Cloud | Curated listThe most popular nodes, said to cover about 90% of local workflows; you cannot install others | From the Creator plan, imported from Civitai or Hugging Face; 900+ models preinstalled | Kept as long as needed to run the service; included storage not stated (Comfy API storage is $0.20 per GB a month) | 5 free runs, no cardRun cap: 30 minutes per workflow, 1 hour on Pro |
| Floyo | Curated list8,478 nodes from 352 packs preloaded; missing ones by request | Unlimited LoRA uploads, kept private | 20 to 200 GB by plan; trial files kept 72 hours | 15 minutes of GPU time, no cardRun cap: None |
| ComfyICU | No installs4,000+ preinstalled nodes; others by request on Discord | By public download link only; cached temporarily | Not stated | $0.50 of credits; free-tier images are public by default |
| RunComfy | Any nodeComfyUI Manager preinstalled; separate saved environments | Upload, or download from Civitai, Hugging Face or Google Drive | Pro: 200 GB kept; pay as you go: 10 GB, inactive files removed after 90 days | None beyond sign-up |
| ThinkDiffusion | Any nodeThrough ComfyUI Manager; no command line | Upload, or fetch by URL from Civitai or Hugging Face | TD-Pro: 200 GB while subscribed; free files deleted after 48 hours of inactivity | 15 minutes |
| MimicPC | Any nodeInstall any model or extension | Upload your own models | 50 GB on every account; data kept 30 days after cancelling | $0.50 of credit |
| RunPod | Any nodeAn official ComfyUI template with Manager, or your own container | Anything you upload or download into the pod | On a network volume ($0.07 per GB a month) or the pod's volume disk ($0.20 per GB a month while stopped); the container disk is erased when the pod stops | None |
| Vast.ai | Any nodeA ComfyUI template, or your own provisioning script | Anything you put on the instance | While the instance exists, at the host's storage price | None |
Privacy and training
What each service says about using your prompts, images and videos, and where it runs them. Three say plainly that they do not train on your content; the others are silent, or reserve broad rights.
Raw GPUs add a layer: on RunPod's Community Cloud and on Vast.ai the machine belongs to a third-party host, and both services say in their terms that they are not responsible for what hosts do. ComfyICU's free tier makes images public by default.
Running it on a raw GPU
RunPod and Vast.ai are two of many places to rent a bare GPU; the cloud GPU price guide compares 13 providers card by card.
- Start from a template. RunPod's official ComfyUI template comes with ComfyUI Manager; its first start can take up to 30 minutes. Vast.ai's template takes your own provisioning script.
- Keep models on a volume. On RunPod the container disk is erased when the pod stops; a network volume ($0.07 per GB a month) or the pod's volume disk keeps models and outputs. On Vast.ai, storage is billed for as long as the instance exists, even stopped.
- Stop it yourself. Neither documents an auto-stop for an idle machine, so a pod left on overnight bills every hour of the night.
- Secure or community. RunPod's Community Cloud is 28% to 62% cheaper than Secure Cloud on the same card, on machines run by third parties.
To run a workflow behind an API rather than in the browser, the ComfyUI API guide compares Comfy API, RunComfy, RunPod Serverless, Modal and more on what one job costs, custom nodes and cold starts.
Check your workflow first
Comfy Cloud runs only the custom nodes it pre-installs, and you cannot add your own. Drop a workflow into the ComfyUI workflow checker and it says whether Comfy Cloud has every node the workflow uses and whether it fits the 96 GB card, and lists every model file with its size, before you pay anyone. It runs in your browser; the workflow is not uploaded.
Questions
Is Comfy Cloud free?
It has five free runs on real GPUs, with no card needed. After that it is a plan: from $20 a month, which buys about 4.4 hours of generation on its 96 GB card.
How much does Comfy Cloud cost?
Plans are $20, $35, $100 a month, or $16, $28, $80 a month billed yearly, and Team plans start at $200 a month, all paid in credits worth $4.54 an hour of generation, counted by the second. Comfy says a 5-second Wan 2.2 clip at its defaults (640×640) takes about 11 credits, about $0.05. Plan credits do not roll over; top-ups last a year. Prices are before tax.
Is ComfyUI itself free?
Yes. ComfyUI is open source and free to run on your own computer. What you pay for in the cloud is the GPU, the storage and, on managed services, the setup.
Is Comfy Cloud private?
Its terms say Comfy will not use your inputs or outputs to train generative models, and that you keep the rights to them. It may use metadata, such as how prompts are classified, to improve its products, and keeps your content as long as needed to run and improve the service. Partner Nodes send your inputs to those providers, whose own terms apply. Comfy does not say where its servers are.
Does Comfy Cloud have a time limit?
Yes: a workflow stops after 30 minutes on the Standard and Creator plans, and after an hour on Pro. Long video jobs that need more run better on a rented machine.
Can I use my own custom nodes on Comfy Cloud?
No. Comfy Cloud runs a curated set of popular nodes that it says covers about 90% of local workflows. For any node, use a rented ComfyUI machine (RunComfy, ThinkDiffusion, MimicPC) or a raw GPU (RunPod, Vast.ai).
Can I upload my own LoRAs to Comfy Cloud?
From the Creator plan you can import models and LoRAs from Civitai or Hugging Face. Comfy says direct upload of larger files is on its roadmap.
What is the cheapest way to run ComfyUI in the cloud?
By the hour, a raw GPU: an RTX 4090 is $0.34 an hour on RunPod Community Cloud and $0.89 on RunPod Secure Cloud, and older cards cost less still. On Vast.ai each host sets its own price: check its live prices. By the session, it depends on how much the GPU works: against a RunPod RTX 4090, Comfy Cloud is cheaper while the GPU generates less than 20% of the session.
What is the difference between Comfy Cloud and RunPod?
Comfy Cloud is ComfyUI as a service: nothing to install, curated nodes, billed only while a workflow runs, at $4.54 an hour on an RTX PRO 6000. RunPod rents you the machine: any node and any model, billed every second it is on, $2.49 an hour for the same card, and you set it up and stop it yourself.
Does ComfyUI have an API?
ComfyUI itself serves an HTTP API. Hosted services with an API are Comfy Cloud, Floyo, ComfyICU and RunComfy; RunPod and Vast.ai run ComfyUI as a serverless worker. Our ComfyUI API guide compares the ways to run a workflow as an endpoint, from Comfy's own Comfy API to RunPod Serverless and Modal.
How much VRAM do I need for ComfyUI?
It depends on the model. In ComfyUI's own templates, Anima peaks at 4.4 GB and MiniMax H3 at 24.4 GB. A 24 GB card runs 8 of the 11 models here with room to work.
Sources, all read on 9 October 2026 (RunPod 19 October 2026, Comfy Cloud 16 October 2026): Comfy Cloud pricing, Floyo pricing, ComfyICU pricing, RunComfy pricing, ThinkDiffusion pricing, MimicPC pricing, RunPod pricing, plus each service's terms and privacy policy, linked in the tables. Memory figures come from this site's own ComfyUI template math: see each image and video model page for the files and their sources.