Civitai LoRA training runs in the browser: upload 15 to 30 images, let the trainer caption them, pick a base model and pay in Buzz, from 500 Buzz (about 50 cents) for SDXL or Flux 2 Klein 4B. Two limits come first: Civitai's own notice says the site has not been accessible in Australia since March 16, 2026, and its rules prohibit content based on any real person's likeness. Start from the documented defaults of 2,000 steps and 10 saved epochs, then compare epochs before you keep one.
Civitai LoRA training runs in your browser: you upload 15 to 30 images, let the trainer caption them, pick a base model and pay in Buzz, from 500 Buzz (about 50 cents) for an SDXL or Flux 2 Klein 4B run. Two limits come before anything else. Civitai's own notice says the site has not been accessible to users in Australia since March 16, 2026, and its rules prohibit content based on the likeness of any real person, so this is a route for fictional personas only.
Checked in October 2026 against Civitai's on-site trainer guide, its Training Studio announcement, the training page and membership pricing, the Buzz guide, the changelog, and the developer docs for SDXL, Flux 1, Flux 2 Klein and other image models. The settings are Civitai's documented defaults and its own tuning notes. We did not run a test persona through the trainer for this page, so it shows no before-and-after images.
A LoRA is one way to hold a face; the consistent character AI guide compares it with reference-image methods that need no training. If Civitai is blocked where you are, or the face is your own, train a LoRA free online covers the alternatives.
Before you start: two limits, two trainers
Civitai currently has two trainers. The original wizard is still the default. Training Studio, a rebuilt version, went into beta on September 23, 2026 and is switched on per account. Both are paid in Buzz, and switching Training Studio on does not remove the old one.
- The default: Create, then Train a model
- Wizard: model type, data, review and submit
- You choose tag or caption labelling
- Email when the run is done
- Epochs listed with samples at the end
- Up to five runs submitted together
- Switch it on in preferences, under Features
- Pick character, style, concept or effect; it suggests a base model
- Labelling is automatic, in the format the model wants
- Progress, logs and epochs appear as they finish
- Generate with an epoch while the run continues
- Compare epochs grid, loss graph, zip download
What you will have, and what you need first
At the end you have up to 20 saved epochs of one LoRA, each with sample images, and 30 days to download or publish them. Here is what to have ready.
- Civitai loads in your country
- The persona is fictional: no real person's face anywhere in the dataset
- 15 to 30 sharp images; Civitai's training page says that is usually enough
- Angles, expressions, lighting and backgrounds vary from image to image
- Images are PNG, JPG, JPEG or WEBP; captions, if you bring them, are .txt files with matching names
- An invented trigger word is chosen
- You know which base model you will generate with
- You have the Buzz: 500 for SDXL or Klein 4B, 2,000 for Flux 1
For the image set itself, the character LoRA dataset checklist gives a shot list, and how many images to train a face LoRA explains why more is not always better.
How to train a LoRA on Civitai, step by step
These steps follow the classic trainer, which every account has. Training Studio asks for the same things in a different order.
- 1Open the trainer
Click the Create button, then Train a model. Result: the training wizard opens on the model type page.
- 2Choose the model type and name it
Pick character for a persona. The type pre-sets some advanced settings, which you can still change later.
- 3Add your images
Drop in loose images or a zip, or pull from your generations, your profile or a past dataset. Tick the acknowledgement to unlock the upload.
- 4Caption them
Auto Label writes tags with WD Tagger or sentences with JoyCaption. Put your trigger word in Prepend Tags so it leads every caption, then review the result.
- 5Pick the base model
Choose the family you will generate with. Result: the trainer selects the engine and shows the Buzz cost before you commit.
- 6Review the settings
Leave Training Parameters at their defaults for a first run. Add up to three sample prompts that include the trigger word.
- 7Submit and wait
Result: a status and ETA on the Training tab, then an email. A run cannot be stopped once it starts, so check the price first.
- 8Compare epochs and download
Each epoch arrives with its samples. Download the ones worth keeping within 30 days; at most 20 epochs are stored.
Two captioning choices matter for a face. Civitai's guide says tags are the preferred format for SD 1.5 and SDXL and natural-language captions are preferred for Flux; picking the other one only triggers an informational warning. And before you leave the data step, use the Download button to save the captioned set as a zip, so a second run does not start from raw images.
Run times in the developer docs are short: about 5 to 15 minutes for a default SDXL run, 3 to 10 for Klein 4B and 10 to 30 for Flux 1 or Klein 9B. They add that queue time on busy days can be longer than the run itself.
Civitai LoRA trainer settings: every field, for a face LoRA
The developer docs publish each training parameter with its default. The names below are the API keys; the on-site form labels them in plain words and does not expose all of them. The last column is Civitai's own guidance where it gives any, not ours.
| Field | Default | What it does | Civitai's note for a character LoRA |
|---|---|---|---|
| steps | 2,000 | Total training steps; the main driver of time and price | Raise to 1,500 to 2,500 if a Flux character LoRA comes out under-trained |
| epochs | 10 | How many checkpoints are saved, up to 20 | Each one is a version to compare; each adds 10 or 20 Buzz |
| batchSize | 1 | Images per step | Capped at 4 on SDXL and 2 on Klein 4B; fixed at 1 on Flux 1 |
| lr | 1e-4 | Learning rate | 5e-4 is called typical for character LoRAs on SDXL; keep Flux 1 and Klein at or below 5e-4 |
| networkDim | 32 on SDXL and Klein, 16 on Flux 1 | LoRA rank | Lower it when a LoRA memorises: 16 to 24 on SDXL, 8 to 12 on Flux 1 |
| networkAlpha | Matches networkDim | Scales the effective learning rate by alpha divided by rank | On SDXL, half of networkDim is one of the overfitting fixes |
| optimizerType | adafactor on SDXL, adamw8bit on Flux and Klein | The optimizer | None given; leave it for a first run |
| lrScheduler | cosine | How the learning rate falls during the run | None given; leave it for a first run |
| trainTextEncoder | On for SDXL at 5e-5, off for Flux and Klein | Whether the text encoder is trained too | On SDXL, switching it off halves memory cost but reduces prompt-following |
| triggerWord | None | The token that switches the LoRA on | Recommended for character and style LoRAs |
| noiseOffset, minSnrGamma | 0.1 and 5 on SDXL | Stabilising options | None given; leave them |
| shuffleTokens, keepTokens | Off and 0 | Shuffles caption tags, keeping the first few fixed | Only matters for tag-style captions |
| samples.prompts | Empty | Prompts rendered at each saved checkpoint at LoRA strength 1.0 | The form gives you three; write them with the trigger word |
Three things follow from that table. First, steps and epochs are separate dials: steps set how long the LoRA trains, and epochs only set how many checkpoints are saved along the way. Second, base models differ, so read the row for yours instead of reusing SDXL numbers on Flux. Third, the defaults are a reasonable first run. Training Studio now shows a recommended value and a reset beside each advanced setting, and describes steps in plain terms, such as each image seen about 100 times for a character.
To see how those numbers compare with other trainers, face LoRA training steps and learning rate puts Civitai's defaults beside kohya, AI Toolkit and the hosted trainers.
What Civitai LoRA training costs in Buzz
Default run: 2,000 steps and 10 saved epochs. Civitai's training page converts 1,000 Buzz to $1. Source: Civitai developer docs, checked October 2026
The price is steps times a per-step cost, plus a small charge per saved epoch, and it never drops below 80 percent of the default. So halving the steps on an SDXL run saves 100 Buzz, not 250. Training on a custom checkpoint adds Buzz, and so does a job with very high epochs and repeats.
You can pay without spending money. The Buzz guide says all types of Buzz work for training, including Blue Buzz: 100 on a new account, a daily claim from the image generator, rewards for reactions, and 0.25 Buzz per ad impression. It does not publish how quickly that adds up, so treat a free first run as days of earning, not minutes. Memberships are the paid route:
| Membership | Price | Buzz per month | Private models | Blue Buzz daily bonus |
|---|---|---|---|---|
| Bronze | $10 a month | 10,000 | 3 | 1.5x |
| Silver | $25 a month | 25,000 | 10 | 2.5x |
| Gold | $50 a month | 50,000 | 100 | 4x |
Prices from Civitai's membership page, checked October 2026. You do not need a membership to train.
What you may train: Civitai's rules on real people
The Real People rules are short and strict. Content that depicts or is based on the likeness of real people is prohibited, whether the person is a celebrity, an influencer, a private individual or a historical figure. A character trained on photos of an actor is not allowed if the output resembles the actor. Civitai says uploads are scanned by automated tagging and flagged for human review, and it calls the policy zero real-person.
The older trainer guide still describes an upload acknowledgement about having consent for any real person in a dataset. That guide is marked as only partially updated, and its change log dates the acknowledgement to September 2024. Go by the rules page: train an original AI persona here, and train your own face somewhere else. Our Flux LoRA walkthrough covers the local and hosted routes and what each allows.
After training: test, keep or publish
- Day 0Training completes
Epochs are listed with their sample images, and the model lands in your library.
- Days 1 to 30Free to use and download
Generate with it on-site with no subscription, and download the epochs you want to keep.
- Day 30Unclaimed files are deleted
Epoch files from a run you never accepted are removed, along with the stored dataset.
- After day 30Keep it private or publish it
A membership keeps a model private on Civitai. Publishing to the community is free and permanent.
Source: Civitai training page and trainer guide, checked October 2026
Judge epochs by images, not by the last number. Run the same prompts against each one: a neutral portrait, a profile, a full-body shot and an outfit that is not in the dataset. Keep the epoch that holds the face and still follows the prompt. Publishing one epoch is permanent and closes the results screen, so download the others first.
The file is the start of the work. The AI Influencers program covers what comes before and after it: creating the persona, LoRA training for consistent faces, ControlNet for body consistency, and a lesson on finding and vetting models on Civitai and Hugging Face. For settings outside Civitai, the LoRA training guide walks through a local run.
Troubleshooting a Civitai training run
| Symptom | What Civitai's docs say to do |
|---|---|
| The face is only loosely similar | Under-trained. Raise steps, or on SDXL raise the learning rate. For Flux, Civitai suggests 1,500 to 2,500 steps at 1e-4 to 3e-4. |
| Outputs copy the training images | Overfit. Use an earlier epoch. Next run, lower the steps and drop networkDim, or on SDXL set networkAlpha to half of networkDim. |
| Klein samples look overcooked or broken | The learning rate is too high. Civitai says to lower it to 1e-4 to 2e-4. |
| The run is rejected | The dataset failed automated moderation. Replace the flagged images; do not resubmit the same set. |
| The run fails | Civitai retries a failed run three times, then refunds the full Buzz amount. |
| Still processing after 48 hours | It should be refunded automatically. If it is not, contact a community manager through Discord or email. |
| The site shows a region notice | Civitai is not accessible from your country. Use another trainer. |
For the second row, face LoRA overfitting maps six symptoms to the setting or dataset fix each one needs.
Civitai LoRA training: FAQ
How do I train a LoRA on Civitai?
Click Create, then Train a model. Choose the model type, upload your images or a zip, let Auto Label write the captions, pick a base model and check the Buzz price, then submit. Civitai emails you when the run finishes and lists each saved epoch with sample images. Download the epochs you want within 30 days, or publish one as a model.
Is Civitai LoRA training free?
It can be. Civitai's Buzz guide says every type of Buzz pays for on-site training, including Blue Buzz, the free kind you earn on the site. A default SDXL or Flux 2 Klein 4B run costs 500 Buzz, and a new account starts with 100 Blue Buzz, so you earn the rest through the daily claim, reactions and ad impressions first. Checked October 2026.
How much Buzz does it cost to train a LoRA on Civitai?
Default runs cost 500 Buzz on SD 1.5, SDXL and Flux 2 Klein 4B, 1,000 Buzz on Klein 9B and Z-Image, and 2,000 Buzz on Flux 1, Qwen and Chroma. Civitai's training page puts 1,000 Buzz at $1, so that is about $0.50 to $2. A run never costs less than 80 percent of the default price. Checked October 2026.
What are the best Civitai LoRA trainer settings for a face?
Start from Civitai's defaults: 2,000 steps, 10 saved epochs, a learning rate of 1e-4 on a cosine schedule and a rank of 32 on SDXL or 16 on Flux 1. Its developer docs call 5e-4 a typical learning rate for character LoRAs on SDXL and suggest 1,500 to 2,500 steps when a Flux character LoRA is under-trained. Change one value per run.
Can I train a LoRA of myself or another real person on Civitai?
Plan on no. Civitai's Real People rules say content that depicts or is based on the likeness of real people, living or deceased, public or private, is strictly prohibited, and they call it a zero real-person policy. Uploaded images are scanned and flagged for moderator review. Train a fictional persona on Civitai, and use a local trainer for your own face.
Is Civitai available in Australia?
No. Civitai's own notice says that as of March 16, 2026 the site is no longer accessible to users in Australia. It cites the age-verification duties in Australia's Age-Restricted Material Codes and says it could not feasibly comply with a team of its size. If you are in Australia, use a notebook, a hosted trainer or a local trainer instead.
What is Civitai Training Studio?
Training Studio is Civitai's rebuilt trainer, released in beta on September 23, 2026. You switch it on in your preferences under Features. It shows progress, logs and epochs as they finish, lets you generate with an epoch while the run is still going, and compares every sample prompt across every checkpoint in one grid. The older trainer is still available.
Civitai trains the file. You still have to build the persona.
AI Influencers, included in All Access, covers creating the persona, LoRA training for consistent faces, ControlNet for body consistency and navigating Civitai and Hugging Face, then video and content, with the other three programs, live coaching and the private community in one subscription.
See what other persona builders are training on
Join the free Telegram channel for the trainers, base models and persona workflows that are working this month.