You can train a LoRA free online through Civitai's trainer paid in free Blue Buzz, a Kaggle notebook with 30 GPU hours a week, a Google Colab notebook, or free compute credits on Modal and Lightning AI. Every one of them is a quota with a documented limit, and FluxGym, the option with no quota, runs on your own GPU instead of online. Pick by what you are training: Civitai for a fictional character with no setup, a notebook for a real face or a blocked region.
You can train a LoRA free online in four documented ways: Civitai's on-site trainer paid with free Blue Buzz, a Kaggle notebook with 30 GPU hours a week, a Google Colab notebook on the free GPU tier, or free compute credits on Modal and Lightning AI. None of them is unlimited. Every free option is a quota, and FluxGym, the one with no quota at all, is not online: it runs on your own graphics card.
- Pick Civitai if the character is fictional, you want no setup, and the site loads in your country.
- Pick Kaggle if you want the largest documented free allowance and can follow a notebook.
- Pick Colab if you already work in it and can live with a GPU that is not always available.
- Pick Modal or Lightning AI if you need a bigger GPU than the free notebooks give and are comfortable with a terminal.
- Pick FluxGym if you own a 12 GB or larger NVIDIA card.
Limits and prices checked in October 2026 against Civitai's Buzz guide, trainer guide and developer docs, the Google Colab FAQ, Kaggle's GPU documentation, Modal pricing, Lightning AI pricing, Hugging Face pricing and the FluxGym README. This page compares what each service documents. We did not run one dataset through all of them, so it makes no claims about queue times or output quality.
Free training is one piece of a larger decision. If you have not settled on a LoRA yet, our consistent character AI guide compares it with reference-image methods that need no training at all.
Train LoRA free online: the options side by side
The useful question is not "which is free" but "what exactly is free, and when does it stop". Three different things get called free: site currency you earn, GPU hours on a notebook, and dollar credits on a cloud.
| Option | What is free | Documented limit | Models | The catch |
|---|---|---|---|---|
| Civitai on-site trainer | Training paid with free Blue Buzz | 500 Buzz per default SDXL run; new accounts get 100 | SD 1.5, SDXL, Pony, Illustrious, Flux 1, Flux 2 Klein and more | No real people; not accessible in Australia |
| Kaggle notebook | GPU time | 30 hours a week, 9 hours per session | Anything that fits its GPU; you bring the trainer | You set up the notebook yourself |
| Google Colab notebook | GPU time | Unpublished; 12-hour maximum runtime | SDXL on the free T4, per kohya-colab | A GPU is not always available; no web-UI trainers on free |
| Modal | Compute credit | $30 a month on the Starter plan | Up to A100 and H100 class GPUs | Command-line setup; billed per second after the credit |
| Lightning AI | GPU hours to start | Up to 75 T4 hours or 31 L4 hours, listed as a starting allowance | A cloud workspace; you install the trainer | Free Studios restart every 4 hours |
| FluxGym | The software | None, but it needs your own 12 to 20 GB GPU | FLUX.1 only | Local, not online |
Civitai: train a LoRA free with Blue Buzz
Civitai is the only hosted trainer here where free site currency pays for the run. Its Buzz guide says all types of Buzz can be used to train LoRAs on the site, and that includes Blue Buzz, the free kind you earn by using it.
- How you earn it: new accounts receive 100 Blue Buzz. After that the guide lists a daily claim from the image generator, reacting to images, reviews and comments, and 0.25 Buzz per ad impression. Civitai does not publish how fast that adds up.
- What a run costs: the developer docs give the default price per base model, shown below, as steps times a per-step cost plus a small charge per saved epoch. A run never costs less than 80 percent of the default, so the cheapest SDXL run is 400 Buzz.
- What you get: up to 20 epochs to download, sample images per epoch, an automatic refund if the run fails, and 30 days to collect your files before they are deleted.
Default run: 2,000 steps and 10 saved epochs. A new account starts with 100 Blue Buzz. Source: Civitai developer docs, checked October 2026
The arithmetic matters. One hundred starting Buzz does not buy a run, so "free" on Civitai means a period of earning first, and a Flux 1 LoRA costs four SDXL runs. If you plan to train on Flux for free here, Klein 4B at 500 Buzz is the realistic target.
Sources: Civitai's Buzz guide, on-site trainer guide, developer recipes for SDXL, Flux 1 and Flux 2 Klein, and its rules on real people.
Kaggle and Google Colab: free GPU notebooks
A notebook gives you a free GPU and nothing else. You load a trainer into it, point it at your images and download the result. It is more work than a hosted trainer and has fewer rules about what you train.
Kaggle publishes its allowance. Its documentation says notebook users get free access to NVIDIA Tesla P100 GPUs, up to 30 hours a week, with individual sessions running up to 9 hours. That is the only weekly GPU quota in this comparison that is written down.
Google Colab is free of charge, but Google's FAQ says access to resources is not assured, the limits are not published and they change over time. Free notebooks run for at most 12 hours and time out when idle. Two rules matter for LoRA training. Free runtimes can be terminated for bypassing the notebook to work mainly through a web UI, which covers trainers that open their own interface. And creating deepfakes is disallowed on every Colab runtime.
What fits on a free 16 GB card is a documentation question too. The kohya-colab notebooks say their SDXL LoRA trainer works on the free T4. The sd-scripts FLUX.1 guide documents 16 GB training at batch size 1 with block swapping, and Black Forest Labs lists 12 GB of VRAM for FLUX.2 [klein] 4B and 22 GB for 9B. So SDXL is the comfortable choice on a free notebook, FLUX.1 and Klein 4B are documented as possible, and Klein 9B is out of reach.
- Upload a zip, pick a base model, press train
- Paid in Buzz you earn on the site
- Default settings chosen for you
- Fictional characters and styles only
- Files kept for 30 days
- You install and configure the trainer
- Paid in GPU hours from a quota
- Every setting is yours to change
- Your own face is allowed, with consent
- You download the file before the session ends
Free compute credits: Modal and Lightning AI
Credit-based clouds give you dollars or hours on real data-centre GPUs, including cards far larger than a free notebook offers. They suit Flux-class models, and they need a terminal.
Modal has a Starter plan with no monthly fee that includes $30 of compute a month. The AI Toolkit README has a section on running training jobs on Modal, with an A100 40 GB and a two-hour timeout as its defaults. At Modal's listed per-second rates, the credit works out like this:
| GPU | Listed rate per second | Per hour | GPU time in $30 |
|---|---|---|---|
| T4 | $0.000164 | About $0.59 | About 50 hours |
| L4 | $0.000222 | About $0.80 | About 37 hours |
| A10 | $0.000306 | About $1.10 | About 27 hours |
| A100, 40 GB | $0.000583 | About $2.10 | About 14 hours |
The last two columns are our arithmetic from Modal's pricing page (checked October 2026). Modal bills CPU and memory on top of the GPU, so real hours come out a little lower.
Lightning AI advertises up to 80 free GPU hours to start, and its pricing table lists the starting allowance per card: 75 hours on a T4 with 16 GB or 31 hours on an L4 with 24 GB. Its FAQ says you get 5 credits on registration and 25 more when you add a card, and that unused credits expire after 12 months. Free Studios need a restart every 4 hours. Treat it as a one-time allowance, not a monthly one.
FluxGym: free, but not online
FluxGym shows up in every "free LoRA training" thread, so it belongs here with the right label. It is free software on GitHub: a simple web interface over kohya sd-scripts that trains FLUX.1 LoRAs on 12, 16 or 20 GB cards. Its README covers a one-click Pinokio install, a manual install and Docker. All three run on a machine you control.
That makes it the best free option if you own the hardware and no option at all if you do not. Renting a GPU to run it is a paid route; our cloud GPU hosting comparison covers what that costs. And because FluxGym is a web interface, running it on free Colab falls under the web-UI restriction above.
What looks free but is not
- Hugging Face Spaces. The free hardware is a CPU with 2 vCPUs and 16 GB of RAM. GPU Spaces are billed by the hour, from $0.40 for a small T4 and $0.80 for an L4 (pricing page, checked October 2026). A trainer Space you duplicate runs on your bill.
- Replicate. Its FLUX fine-tuning guide puts a run at about $1.46. Cheap, not free.
- fal. Its FLUX.1 fast trainer lists $2 per run at the default 1,000 steps, scaling with steps.
- Other generation sites. Several bundle LoRA training with daily credits. We could not confirm their current training allowances on an official page, so they are left out of the table.
Those paid prices are worth keeping in mind as a ceiling. If earning Buzz or setting up a notebook will cost you an evening, a two-dollar run can be the cheaper choice.
Free LoRA training: how to spend a small quota well
A free quota punishes sloppy runs, because every failed attempt uses up the allowance that would have paid for the good one. Do the thinking before you press train.
- 1Finish the dataset first
Cull, caption and zip it before you open the trainer. Dataset problems are a common reason to retrain.
- 2Pick the cheapest model that does the job
On Civitai, SDXL and Klein 4B cost a quarter of Flux 1. On a 16 GB notebook, SDXL is the comfortable fit.
- 3Decide the step count in advance
Derive it from your image count, not from a default you have not read.
- 4Save every checkpoint you can
Several saved epochs let you pick the best one without paying for a second run.
- 5Download the files straight away
Notebook sessions end and Civitai deletes unclaimed files after 30 days.
- 6Test before you train again
Run fixed prompts against each checkpoint. Retrain only when you can name what failed.
Two of those steps have their own guides: how many images to train a face LoRA and face LoRA training steps and learning rate. The character LoRA dataset checklist covers the culling, and the free LoRA Training Steps Calculator turns images, repeats and epochs into a step total and a GPU-time estimate.
- The dataset is culled, captioned and zipped
- The base model matches the one you will generate with
- The subject is allowed on that service (no real people on Civitai)
- The run fits the GPU: SDXL or Klein 4B on 16 GB, larger models on rented cards
- Step count and checkpoint interval are chosen, not left at an unread default
- Sample prompts are set so you can judge checkpoints without extra generations
- You know where the files go and how long the service keeps them
Which free option to pick
Match the option to the subject and the hardware you have, not to the size of the free allowance.
| Your situation | Use | Why |
|---|---|---|
| Fictional persona, SDXL or Klein 4B, no setup wanted | Civitai with Blue Buzz | Cheapest default price and no code |
| A real person with consent, such as your own face | Kaggle or Colab notebook | Civitai prohibits real likenesses |
| Civitai does not load in your country | Kaggle notebook | The only documented weekly GPU quota |
| Flux-class model, larger GPU needed | Modal credit with AI Toolkit | $30 a month covers A100 time |
| You own a 12 GB or larger NVIDIA card | FluxGym or AI Toolkit locally | No quota at all |
| One LoRA, today, and a couple of dollars is fine | Replicate or fal | Not free, but no setup and no queue to earn through |
Whichever route you take, the trainer is the smallest part of the job. Getting a persona that stays consistent across hundreds of posts is a dataset, testing and workflow problem, and that is what the AI Influencers program teaches end to end. For the settings themselves, the LoRA training guide walks through trainers, captions and checkpoint testing.
Free LoRA training: FAQ
Can I train a LoRA free online?
Yes, within a quota. Civitai's on-site trainer accepts free Blue Buzz, Kaggle documents 30 hours of free GPU a week, Google Colab has a free GPU tier with unpublished limits, Modal includes $30 of compute a month and Lightning AI gives free GPU hours to start. Each one runs out, so plan the run before spending it. Limits checked October 2026.
Is Civitai LoRA training free?
It can be. Civitai says all types of Buzz, including the free Blue Buzz you earn on the site, can pay for LoRA training. A default SDXL or SD 1.5 run costs 500 Buzz, Flux 2 Klein 4B costs 500, Klein 9B 1,000 and Flux 1 2,000. New accounts start with 100 Blue Buzz, so you have to earn the rest first. Checked October 2026.
Can I train a Flux LoRA for free?
It is the hardest one to get free. On Civitai a default Flux 1 run costs 2,000 Buzz, four times an SDXL run. In a free notebook you need a trainer set up for a 16 GB card: sd-scripts documents FLUX.1 training at that size with block swapping, and Black Forest Labs lists 12 GB for FLUX.2 klein 4B. Modal's $30 monthly credit covers roughly 14 hours of an A100.
Is FluxGym free?
The software is free to download, but it is not an online service. FluxGym is a local web interface over kohya sd-scripts that trains FLUX.1 LoRAs on your own 12, 16 or 20 GB graphics card. If you do not own one, you rent a GPU to run it, which is paid. Running its web interface on free Google Colab falls under Colab's restriction on web UIs.
Can I train a LoRA on Google Colab for free?
Yes, with caveats. Google says Colab is free of charge, that free notebooks run for at most 12 hours, and that usage limits are not published and can change. Community notebooks such as kohya-colab state they work on the free T4 for SDXL LoRAs. Free runtimes may be terminated for using a web UI in place of the notebook, and creating deepfakes is disallowed.
Which free LoRA trainer is best for a face?
For a fictional character, Civitai's trainer is the least work if it is available in your country: upload, pick a base model, pay in Blue Buzz. For a real person, including yourself, Civitai is out, because its rules prohibit likenesses of real people. Use a Kaggle or Colab notebook or free cloud credits there, and only with that person's written consent.
A free run gets you one LoRA. The persona takes a system.
AI Influencers, included in All Access, covers the whole build: character design, dataset, LoRA training, checkpoint testing and a content workflow, with the other three programs, live coaching and the private community in one subscription.
Compare notes before you spend your quota
Ask in the free Discord which trainer people are using this month, and work out your step total with the free calculator before you start a run.