ComfyUI and Stable Diffusion experts get paid online by turning the jobs they already do for free in DMs (fixing a broken workflow, building a custom one, training a character LoRA) into fixed-scope offers on one link. On Brillay, the AI artist or small studio attaches their workflow JSON, error log or reference images, pays upfront, and you deliver the repaired .json or the trained .safetensors file inside the delivery time you set, up to five days. Quick "why is this node red?" questions sell separately as Paid Questions, answered in writing on your own time.
You probably know the pattern already. Someone loads a workflow you posted, half the graph turns red, and the screenshot lands in your inbox with "quick question." Twenty minutes later you're reading their traceback, guessing their VRAM and explaining which custom node pack was renamed last month. That's real expertise, and it's work people will pay for once there's a clear price and a clear place to pay.
What do AI artists actually ask a ComfyUI expert to fix?
The requests repeat, which is what makes them easy to sell. The common ones:
- Red nodes after loading a shared workflow. Missing custom nodes, nodes that were renamed or abandoned, or packs that won't install cleanly through ComfyUI Manager's "Install Missing Custom Nodes."
- "CUDA out of memory" on an 8 GB or 12 GB card, because the workflow was built on a 24 GB card with no tiled VAE decode, no model offloading and a batch size nobody lowered.
- "mat1 and mat2 shapes cannot be multiplied." Usually an SD 1.5 LoRA or ControlNet loaded into an SDXL graph, or the other way round.
- Grey, washed-out or burned images. A wrong or missing VAE, CFG far too high for the model, or a Flux workflow running SDXL-style sampler settings.
- A workflow that broke after an update. A new ComfyUI version, a torch or xformers mismatch, or a node pack that changed its inputs.
- A character that won't stay the same. The face, hair and outfit drift between images, which ruins a comic page, a game concept sheet or a picture-book mockup. This is where character LoRA requests come from.
- Studio pipelines. A small studio wants a repeatable product-shot, background-swap or upscale workflow that someone non-technical can run by changing one image and one prompt.
Every one of these comes with inputs you can name in advance: a .json file (or a PNG with the workflow embedded in its metadata), a console log, a GPU model, a list of reference images. That's why they work as fixed-scope offers instead of open-ended "help me" chats.
Which offers should a ComfyUI expert sell first?
Start with three Custom Work listings and one Paid Question. Custom Work lets you set the price, the scope and the delivery time; the follower sends a brief and up to 5 files and pays before you start, and you deliver a written note, files or links as one package.
1. Fix a broken workflow (Custom Work)
- What's included: one workflow and the error chain it's throwing. You find what broke, repair the graph, and send back a working .json plus a note explaining the cause, the node versions you used and anything they need to install.
- What the artist sends: the workflow .json (or the PNG with the embedded workflow), the full console traceback rather than a cropped screenshot, their GPU and VRAM, their ComfyUI version and install type (portable, desktop app or manual), and the base model and LoRAs the graph loads.
- Delivery: 1β2 days works for most fixes. Pick 24 hours only if you can actually keep it.
- Price: you set the price; a common starting point is $30β$60 for a single workflow.
2. Custom ComfyUI workflow (Custom Work)
- What's included: one workflow built for one job, for example "SDXL portrait with pose ControlNet, face detailer and a 2x upscale that runs on 12 GB." Nodes are grouped and labeled, with notes on the canvas, plus a list of every model and custom node pack and where to get it.
- What the artist sends: the goal in one paragraph, the base model they want to use, their VRAM, 1β5 reference or target images, and an example of the output they like.
- Delivery: 3β5 days.
- Price: a common starting point is $120β$300, depending on how many stages the graph has.
3. Character LoRA training (Custom Work)
- What's included: one character, trained for one base model. You check and caption the dataset, train, test, and deliver the .safetensors file, the trigger word, a recommended strength range and a sample grid at a few strengths and seeds.
- What the artist sends: 15β40 images of an original or properly licensed character, from varied angles, expressions and lighting; the target base model (SD 1.5, SDXL or a Flux variant whose licence fits the job); the trigger word they'd like; and confirmation that they own or have the rights to the images. Follower attachments are capped at 50 MB per file, so ask for the dataset zipped into one archive, or a shared folder link in the brief.
- Delivery: 3β5 days, depending on your GPU queue.
- Price: a common starting point is $60β$150 per character.
4. One ComfyUI question (Paid Question)
A Paid Question lets a follower pay upfront to ask one private question, with up to two follow-ups, and you answer in writing on your own time, with no live call or scheduling. It fits the short stuff: "Which VAE goes with this checkpoint?", "Why does my IPAdapter output ignore the reference?", "Is kohya_ss, AI-Toolkit or OneTrainer better for my 12 GB card?" Paid Questions are text only, so the follower pastes the error message into the question. If you need to open their file, that's a Custom Work fix. A common starting point is $10β$25, with a reply window you choose (24 hours is a good default).
How do you keep a workflow fix from turning into a week of remote debugging?
The hardest part of selling fixes is the gap between your machine and theirs. Scope protects you:
- Say what you fix. "The graph and its node settings" is a different job from "your Python environment." Put driver installs, broken venvs and remote desktop sessions in the listing's excluded section. If you also do environment repair, sell it as a separate listing at a different price.
- Say what "done" means. For example: the workflow loads with no red nodes and produces an image on your test setup using the same models, plus the exact install steps for anything you added.
- State your test setup in the delivery note. GPU, ComfyUI version, and the commit or version of each custom node pack. If it runs for you and not for them, both of you can see where the difference is.
- Add a short screen recording. Custom Work deliveries can include video up to 20 minutes, so a two-minute walkthrough of what you changed and why saves a lot of back-and-forth.
There are no built-in revision rounds. You write your own support, correction and refund policies, and the follower sees them before paying. After delivery, the follower can open a customer support conversation on the order (for example "File or link does not work" or "Correction or revision"), and you decide how to handle it under the policy you published. A clear line such as "one free adjustment if the fixed workflow fails on the setup listed in your brief" sets expectations without promising unlimited tweaks.
How long should LoRA training and custom workflows take to deliver?
Delivery and reply times are chosen from 1β24 hours or 1β5 days, and five days is the maximum. There are no milestones, so each order has to fit inside one window. For a character LoRA that usually means a day for checking and captioning the dataset, a training run, and a testing pass with a sample grid. If your GPU is busy with your own work, choose five days and deliver early when you can.
Bigger jobs need splitting into pieces that each fit one order. A studio that wants "a full pipeline plus LoRAs for our five main characters" is five LoRA orders and one or two workflow orders, not one giant order. Say this plainly in your listings: "One character per order. One base model per order." It keeps your queue sane and gives the buyer a finished result at each step.
Delivered files stay available for 90 days, so ask buyers to download the .safetensors and .json files and keep their own copies. Most character LoRAs fit comfortably under the 500 MB per-file delivery limit. If yours are bigger, send a download link instead.
What can you train, and what should you refuse?
Write your limits into every listing and decline orders that cross them. The rule for this kind of work on Brillay: SFW only: no adult content, no deepfakes or LoRAs of real people without consent; follow model licences.
In practice:
- Characters, yes; real people, only with consent. An original mascot, a client's comic protagonist or a game character the studio owns is fine. A LoRA or IPAdapter workflow built to reproduce a real person's face is not, unless that person has agreed. If a dataset looks like scraped photos of someone real, decline the order.
- No adult content, including "it's just for a private project."
- Dataset rights. Ask the buyer to confirm they own or have permission to use the training images, and put that confirmation in your listing's "what I need from you."
- Model licences. Base models ship under different terms. SD 1.5 and SDXL use OpenRAIL-style licences with use restrictions. Flux.1 [dev] ships under a non-commercial licence, while Flux.1 [schnell] is Apache 2.0. Before you sell training on a base model, or a workflow built around one, read its licence for the kind of use the buyer describes. If it isn't clear that a paid job is covered, train on a base model whose licence clearly allows it and tell the buyer why. Check custom node licences too if you're bundling someone else's code.
None of this is legal advice. It's the minimum you need to sell image-model work without putting your own name on something you'd regret.
How should you price workflow fixes, custom workflows and LoRA training?
You set every price. These are reasonable places to start, and you can raise them once orders arrive faster than you can deliver:
- Paid Question: $10β$25
- Fix a broken workflow: $30β$60
- Character LoRA, one character, one base model: $60β$150
- Custom workflow: $120β$300
Things that should move your price: how much GPU time the job takes (and what a rented cloud GPU would cost if your own card is too small for Flux training), how many stages the graph has, whether the dataset arrives captioned, and how fast you're promising delivery. A 24-hour fix should cost more than a three-day one.
No monthly fee, you only pay a small platform fee when you get paid. 8% platform fee when a follower pays for work, 5% platform fee on tips and goals. Card processing is separate.
That means a page with four listings and no orders yet costs you nothing to keep up. The full breakdown is on Pricing.
Why not keep taking these jobs in DMs or on a gig marketplace?
In DMs, every job starts with an unpaid back-and-forth: what's wrong, what's your VRAM, how much, can you send it over, then a payment-app request you have to chase. On a gig marketplace, the people you brought from your tutorials and shared workflows land on a search page that ranks other sellers next to you.
On your own Brillay link, you publish the price, the scope and the delivery time once. The artist reads the listing, writes the brief, attaches the JSON and the log, and pays upfront as a guest, with no account to create, right after tapping the link in your bio or under your video. What reaches you is a defined, paid job, not a lead to follow up. Buyers see the public name you choose for your page, not your legal name, which matters if you work under a handle. And there's no follower minimum: it works from the first person who asks. How it works walks through the follower's side step by step.
Where do Tips and Goals fit if you share workflows for free?
Plenty of ComfyUI experts build their reputation by giving workflows away on Civitai, OpenArt or GitHub. Those downloads can pay you back too:
- Tips are a one-time thank-you, with nothing for you to deliver and no membership for the follower to join. Put your link next to every free workflow: "If this saved you an afternoon, you can send a tip."
- A Goal is one named goal with a progress bar, such as "A 24 GB GPU so I can train Flux LoRAs locally" or "Time to maintain my free node pack." Followers make one-time contributions, every contribution is yours whether or not the goal is reached, there's no deadline unless you set one, and there's nothing to ship.
- Personalized Videos fit this trade least. A screen-recorded walkthrough usually belongs inside a Custom Work delivery instead.
How do you set up your page today?
- Create your page and pick the public name you work under.
- Add the three Custom Work listings above. For each one, fill in what's included, what the artist must send, what's excluded (environment repair, adult content, real-person likeness without consent, licence-restricted uses) and the delivery time.
- Add one Paid Question with a reply window you can keep on a busy week.
- Write your support, correction and refund policies in plain words.
- Put the link in your bio, on your Civitai and GitHub profiles, and in the description of every workflow you share. Next time someone sends "quick question" with a red-node screenshot, reply with the link to the listing it matches.
If you also fix configs and setups outside image generation, the approach in selling Docker or Kubernetes config fixes works the same way for environment work. When you're ready, start your Brillay page.
Frequently asked questions
Can a follower send me their workflow file with a Paid Question?
No. A Paid Question is text in, text out, so the follower pastes the error message and describes the setup. When you need to open their .json, PNG or dataset, point them to your workflow-fix listing, where they can attach up to 5 files with the brief.
Can I sell LoRA training or workflows built on Flux?
It depends on which Flux model and what the buyer will do with the output. Flux.1 [dev] ships under a non-commercial licence and Flux.1 [schnell] under Apache 2.0, so read the licence for the specific use before you accept the order. If a paid use isn't clearly covered, train on a base model whose licence allows it. Whatever the model: SFW only, no deepfakes, and no LoRAs of real people without their consent.
What if the fixed workflow still errors on the buyer's machine?
After delivery, the buyer can open a customer support conversation on the order and choose a reason such as a file that doesn't work or a correction request. You handle it under the correction policy you published before they paid. Stating your test setup in the delivery note makes these conversations short.
What does it cost me to sell ComfyUI work on Brillay?
There's no monthly fee and no paid plan. The only cost is the platform fee in the pricing section above, taken from orders that were actually paid, plus card processing.
Will buyers see my real name?
No. Followers see the public name you choose, on your page, at checkout and on their receipt. Your legal name is used only for identity verification and payouts.
Can I get paid if I'm not in the US?
Creators in 41 countries can get paid, and followers pay by card from anywhere, which suits an audience that's spread across time zones and Discord servers.
Explore more