Where should you train a LoRA?
Train a LoRA where you can pick the base model and download the .safetensors file at the end. We train on a rented RTX PRO 6000 in AI Toolkit on RunPod: $2.09 an hour on secure cloud (checked 4 Oct 2026), 2 to 3 hours per character. Hosted trainers are simpler, your own PC works if it has the VRAM, and free Colab rules out web UIs. A LoRA is what keeps a consistent AI character.
Affiliate note: RunPod and Fanvue links on this page are referral links. If you sign up through them, AI Empire may earn credits or a commission. You pay the same. How this works.
What are your options for training a LoRA?
Six, from a GPU you own to an app that never hands you a file. The two questions that sort them: can you train on the base model your generator uses, and can you download the result? Everything below comes from each provider’s own pages, checked 4 Oct 2026. Prices change often.
| Where | How you pay (4 Oct 2026) | Krea 2? | Download the file? |
|---|---|---|---|
| Your own PC | The card you own, plus power | Yes, in AI Toolkit, if the VRAM allows | Yes |
| RunPod pod + AI Toolkit (ours) | RTX PRO 6000 $2.09/hr, secure cloud | Yes | Yes, from JupyterLab or the job page |
| Ostris Cloud | RTX PRO 6000 $2.51/hr, billed per second, container disk and bandwidth included | Yes (it runs AI Toolkit) | Through AI Toolkit’s UI |
| fal (API) | Krea 2 trainer $0.003/step; Z-Image trainer $2.26 per 1,000 steps | Yes | Yes, a URL to the .safetensors |
| Civitai trainer | Buzz, from 500 for SD 1.5 and SDXL; newer models cost more | Not listed in its guide (updated 1 Apr 2026) | Yes, up to 20 saved epochs |
| Higgsfield Soul ID | 25 credits, 20+ photos | No, its own model | Not mentioned on the pages we read |
Sources: RunPod pricing, Ostris Cloud, fal Krea 2 trainer, fal Z-Image trainer, Civitai’s trainer guide (updated 1 Apr 2026), Higgsfield on Soul ID. Google Colab is left out of the table on purpose: see below.
How much does it cost to train a LoRA on RunPod?
About $4.20 to $6.30 of GPU time for one character at our settings: an RTX PRO 6000 at $2.09 an hour on secure cloud, for a 3,000-step run of 2 to 3 hours. Add the pod’s disk while it exists, then terminate it. Budget for a second run: the first one teaches you what to fix.
What a run looks like with our template:
- Deploy the trainer template on an RTX PRO 6000 (96 GB), with
AI_TOOLKIT_AUTHandJUPYTER_PASSWORDset. - Upload her 50-image dataset in AI Toolkit’s New Dataset screen.
- Set up the job (Krea 2, rank 64, 3,000 steps, a save every 250) and press Start. You can close the tab.
- Download the final file plus the 2,500 and 2,000-step saves.
- Stop and terminate the pod.
Every field is in video 3. The disk is the cost people forget: a stopped pod’s volume disk bills $0.20 per GB per month (RunPod docs), so a 100 GB volume left stopped is about $20 a month for nothing. To add up your own GPU and hours, use the RunPod cost calculator.
Should you use a hosted LoRA trainer instead?
If you want no setup and your base model is on its list, a hosted trainer is a fair choice. fal’s Krea 2 trainer returns the LoRA as a downloadable .safetensors file. The catch is control: its output is one final LoRA file plus a config, with no in-between saves to compare.
Things to check on any hosted trainer, with what we found on 4 Oct 2026:
- Defaults. fal’s Krea 2 trainer defaults to 100 steps at 768 px (API docs), far from our 3,000 steps at 1024. We haven’t tested fal’s defaults on a character, so set steps and resolution on purpose (it offers 768 or 1024). Its learning rate differs too, so step counts don’t compare one to one. At $0.003 a step, 3,000 steps would be $9.
- Captions. fal’s trainer can auto-caption. Our captions never describe her face or hair; LoRA captioning explains why. Upload your own caption files.
- Base models. Civitai’s guide (updated 1 Apr 2026) lists SD 1.5, SDXL, Flux, Chroma, Qwen, Z-Image and video models, but not Krea 2. If your base model is on its list and you’d rather not rent a pod, Civitai keeps up to 20 epochs for you to download and compare.
- Faces. Civitai requires that you’re the person in a dataset of a real person, or have their explicit consent. We never train on real people.
- Inside-the-app trainers. Higgsfield’s Soul ID trains on 20+ photos (your own, or someone else’s if you hold the usage rights) and stays inside Higgsfield. More in Higgsfield alternative.
Can you train a LoRA on your own PC?
Yes, if your card has the VRAM for your base model. AI Toolkit, OneTrainer and Kohya all run locally. For Krea 2 at our settings, 80 to 96 GB trains at full quality. On 32 to 48 GB cards, including the RTX 5090, it only fits with fp8 quantization and Low VRAM, and the LoRA comes out worse.
So a home PC works, with trade-offs:
- No hourly bill, but you pay for the card and the power, and the run ties up your machine for hours. The arithmetic is in rent a GPU or buy one.
- Smaller base models need less. musubi-tuner’s README recommends 12 GB or more for image training in general; check the numbers for your model.
- Which trainer depends on your base model: AI Toolkit vs OneTrainer lists what each supports. VRAM by task is in how much VRAM ComfyUI needs.
Can you train a LoRA for free on Google Colab?
Not with a web-UI trainer on the free tier: Colab’s FAQ disallows “bypassing the notebook UI to interact primarily via a web UI” for free users, and AI Toolkit’s trainer is a web UI. On top of that, no Colab tier guarantees which GPU you get, so you can’t count on a 96 GB card.
Colab’s other rules, side by side with RunPod, are in RunPod vs Google Colab.
Why do we train on a rented pod?
Because the course’s base model is Krea 2, we want the file, and a 96 GB card for an afternoon costs a few dollars. Our trainer template runs the official AI Toolkit image, needs no Hugging Face account, unzips a dataset by itself and keeps everything on the volume. It’s also where our other templates live.
We rent on RunPod, and earn a referral if you sign up through that link. We use it because the face template, the Dataset Maker, the LoRA trainer and the Telegram bot’s serverless endpoint all run there, so one account and one balance cover the whole build. Ostris Cloud is a fair alternative for training alone: it’s run by AI Toolkit’s author, and he says the money funds the project. Its RTX PRO 6000 cost more per hour than RunPod’s secure cloud on 4 Oct 2026 ($2.51 vs $2.09), but Ostris includes container disk and bandwidth in that rate, while RunPod bills disk separately.

How many images go in, and how to choose them: how many images for LoRA training.
What can go wrong?
- You can’t download the result. Then her face lives in someone else’s app. Check for a download before you pay for training.
- Wrong base model. A Krea 2 LoRA needs a Krea 2 workflow; a Z-Image LoRA needs Z-Image. Train on the model your generator uses.
- You keep a hosted trainer’s defaults without checking them. fal’s are 100 steps at 768 px. Set steps and resolution on purpose.
- Civitai deletes your epochs. Its guide says epoch files of models left unpublished are deleted after 30 days. Download the ones you want first.
- Auto-captions describe her face. Upload your own scene-only captions.
- You terminate before downloading. Terminating deletes everything not on a network volume. Download first.
- You stop the pod and forget it. A stopped pod still bills its volume disk. Terminate when the files are safe.
- Out of memory on a 32 to 48 GB card. Turn on fp8 quantization and Low VRAM, or move to an 80 to 96 GB card.
- You train on someone’s real photos. Don’t. Start from a generated face and keep her original.
Questions people ask
Can I train a LoRA for free?
How much does it cost to train a LoRA online?
Can I train a LoRA on my own PC?
Which hosted trainer lets me download the LoRA?
Can I train on photos of a real person?
Read next
- AI Toolkit vs OneTrainer: which LoRA trainer should you use?
AI Toolkit, OneTrainer, Kohya's GUI and musubi-tuner compared from their own READMEs (4 Oct 2026): models, interface, cloud use, and why we use AI Toolkit.
- How many images do you need to train a character LoRA?
A character LoRA needs 20 to 50 images. We train on 50. What they should show, how to caption them, how to pick a trigger word, and what goes wrong.
- How much VRAM does ComfyUI need? It depends on the job
ComfyUI needs little; the model sets the VRAM. What faces, datasets, LoRA training and video need in a real pipeline, from 16 GB to 96 GB, with sources.
- RunPod for AI images: templates, GPUs and what it costs
Rent a GPU by the hour instead of buying one. Which RunPod GPU to pick, what storage costs, how to run ComfyUI, and the stop vs terminate trap.