A character LoRA is a small add-on file, trained on about 10 to 20 pictures of one character, that teaches an image model to draw that character whenever you use a made-up trigger word. Train only on pictures you have the right to use.
Reference pictures are the easy way to keep a face the same; a LoRA is the next step when references are not enough. You can train one in a browser on a hosted service, or run the software yourself on a graphics card.
- Try references first. A LoRA is extra work: a dataset, a trigger word, a training run and testing.
- Variety in, consistency out. Same character in different places, expressions and backgrounds.
- Numbers differ by trainer. Replicate says 10 to 20 pictures; fal’s FLUX.2 trainer says 20 to 1,000. Follow your trainer.12
- Test with and without it, using the same prompt and seed.
- Your own characters only, or people who said yes.
Think of a camera and a lens filter. The camera, the base image model, stays as it is. You train a small filter that adds one thing: your character. Say the secret word and the filter switches on. The filter is a small file, so it is easy to keep, share and swap.
Do you need a LoRA, or are reference pictures enough?
The plain point: for one comic or a few dozen pictures, reference pictures are usually enough. A LoRA is for when you need the character everywhere, in hundreds of pictures or in a tool that takes no references. That is this site’s reading; the table sets out the trade.
| Way | What you do | Good for | Cost in effort |
|---|---|---|---|
| Reference picture | Do: attach a picture of the character to each prompt. | For: a comic, a picture book, a short series. | Effort: low. See the comic and picture-book guide. |
| Written description | Do: paste the same character words every time. | For: any tool, as a base layer. | Effort: lowest. See identity lock templates. |
| Character LoRA | Do: train a small add-on on 10 to 20 or more pictures. | For: hundreds of pictures, or a model that takes no references. | Effort: high: dataset, training, testing. |
For what a LoRA is in the first place, see what is a LoRA?. Hugging Face, which publishes the tools many trainers use, says LoRA “is a popular and lightweight training technique that significantly reduces the number of trainable parameters” and makes the finished weights “a few hundred MBs, which are easier to store and share”.3
How many pictures does a character LoRA need?
The plain point: the trainers give different numbers, so follow the one you use. All of them agree that more variety beats more copies.
- Replicate’s FLUX.1-dev trainer: “Aim for a dataset of 12-18 images of your subject.”4
- Replicate’s fast trainer: “Aim for a dataset of 10-20 images of your subject.”1
- fal’s Turbo Flux trainer: “Try to use at least 10 images, although more is better.”5
- fal’s FLUX.2 trainer: “you should gather 20 to 1,000 images that reflect the same style, subject, or identity you want to teach the model.”2
Replicate’s blog on training with your own face adds that you can fine-tune “with as few as two training images, but for best results you’ll want to use at least 10 images or more.”6
How do you choose the pictures?
The plain point: the same character, in different settings, with different expressions and backgrounds, and nothing that changes the character itself.
Replicate’s advice for character LoRAs is to “use images of the subject in different settings, facial expressions, and backgrounds” and to “avoid different haircuts or ages, and showing hands in a lot of face framing positions as we found this led to more hand hallucinations”.4 Its blog says the same of variety: “choose training images with different settings, clothing, lighting, and angles.”7
If your character is drawn, not photographed, you need a set of pictures of it first. In this site’s reading, make them with the reference methods above and keep only the ones that match: front and side views, close-ups, full-length shots, a few expressions and backgrounds. One bad picture in a small set teaches a bad habit.
One setting to know about: fal’s Turbo trainer has a face_crop option, “Whether to try to detect the face and crop the images to the face,” switched on by default.8 That suits a face LoRA; for a character whose costume or body shape matters, check it before you train.
What are a trigger word and captions?
The plain point: the trigger word is the made-up name you will type to switch the LoRA on. Captions are short descriptions that go with each training picture.
On the trigger word, Replicate’s two pages differ. Its blog says it “should be something unique like MY_UNIQ_TRGGR”, “should not be an existing word in any language” and “should not be TOK, because it will clash with other fine-tunes”.7 The trainer’s readme says a “Trigger word can be a generic proper name (ex. sarah, john)” and, for character LoRAs, suggests pairing the trigger word with a gender word such as man or woman to improve results.4 This site’s view: follow the page for your trainer and, if unsure, pick a made-up word that means nothing else.
For captions, fal’s FLUX.2 guide says to put a text file next to each picture with the same name (oil_01.png and oil_01.txt), and that “Each caption should describe the image content and include a unique trigger phrase.” It also says “caption files usually lead to far better LoRA learning retention, especially when your dataset contains a variety of subjects.”2 Replicate’s fast trainer will write captions for you with an autocaptioning model, or take your own text files.1 A caption in the fal style, for an invented character, might read:
How do you train it?
The plain point: there are two routes. A hosted service trains it for you on its computers. Or you run training software on your own graphics card.
- Put your pictures in one zip file. Both Replicate and fal take a zip of images, with optional caption files of the same name.
- Pick a trigger word and, if the trainer asks, whether this is a “subject” or a “style”. For a character, choose subject.
- Leave the settings alone at first. fal’s Turbo trainer defaults to 1000 steps and a “subject” style, and says of FLUX.2 that “for most cases, the defaults work well.”8 Replicate’s blog says to leave steps “at 1000. Any less and your training process will not properly learn the concept”, while its readme gives a range of 1000 to 3000.7
- Start the run and wait. Replicate’s blog says about 20 pictures and 1000 steps is “typical” for a 20-minute training; fal says a run may take “anywhere from a few minutes to hours for larger datasets”.7
- Test it (next section), then keep the file.
To train on your own computer, you need a graphics card with enough memory. The open-source AI Toolkit lists “Nvidia GPU with enough ram to do what you need” as a requirement.9 Hugging Face’s example script is slower: it says a full training run “takes ~5 hours on a 2080 Ti GPU with 11GB of VRAM” for its own example.10 Its guide points to two settings: --rank, “the inner dimension of the low-rank matrices to train; a higher rank means more trainable parameters”, and --learning_rate.10 Replicate’s readme says “LoRA Rank between 16-32 produce good results”, with 64 for likeness and 8 for styles in its own runs.4
Good pictures matter more than clever settings: the same character, varied scenes and expressions, and a made-up trigger word. Train with the defaults first and change one thing at a time.
How do you test a character LoRA?
The plain point: run the same prompt with and without it, then try the character in places your training pictures never showed.
fal’s advice: “To test whether your LoRA learned what you intended to teach, generate using the same prompt and seed with and without the LoRA.” (spelling corrected from the page) The prompt must include the trigger phrase.2 You can also turn the LoRA up or down. Hugging Face says “A value of 0 is the same as only using the base model weights, and a value of 1 is equivalent to using the fully finetuned LoRA.”11 Replicate says that for stronger style effects you can reduce it, “reducing the lora strength to 0.8 - 0.95 can make a difference”.4
A short test list (this site’s own):
- The character in a new outfit, a new place and a new pose.
- A close-up and a full-length picture.
- A different art style, to see whether the character survives it.
- Prompts without the trigger word, to confirm nothing leaks in.
If the character looks right only in poses like your training pictures, the dataset was too narrow. Replicate puts the link bluntly: “The worse the dataset, the less likely that the model will be flexible enough to apply different art styles to the subject.”4 Add varied pictures and retrain.
Whose pictures can you train on, and who can use the result?
The plain point: train on characters you invented, your own face, or people who have clearly agreed. The licence of the base model also matters.
A LoRA trained on someone’s photos can put that person in pictures they never posed for, which is one way deepfakes are made. See what a deepfake is and voice and likeness law. This is this site’s advice, not a legal opinion.
Check the licence too. Replicate says: “If you generate images on Replicate with FLUX.1 models and their fine-tunes, then you can use the images commercially. If you download the weights off Replicate and generate images on your own computer, you can’t use the images commercially.”1 Other base models have their own terms, so read the one you train on. More on ownership is on who owns what you generate.
Questions people ask
Can I train a LoRA without a graphics card?
Yes, on a hosted service, which trains it on its own computers. Replicate’s blog says its FLUX trainings run on “Nvidia H100 GPU hardware”.7 The pages read here do not state plan prices, so check each service’s own page.
Is a LoRA the same as a reference image?
No. A reference picture steers one result and nothing is trained. A LoRA is a trained file that changes every result. See what a LoRA is and how to use reference images.
Can I use a LoRA in ChatGPT or Midjourney?
The pages read for this guide do not describe it. LoRAs are mostly used with open models, run at home or on studio sites that support them: see open weights. For ChatGPT, Gemini and Midjourney, use their reference-picture features, covered in the comic and picture-book guide.
Why does my LoRA ignore my prompt or look stiff?
Often the dataset was too narrow or the LoRA too strong. Lower its strength and add varied pictures, as above. Replicate also found that “Increasing the step count of the inference improves LoRA coherence in the case of a weaker dataset.”4
Where to go next
- “What is a LoRA?” Read what is a LoRA?
- “Do I even need one?” See keeping the same character in a comic or picture book and character consistency in AI.
- “What are the other ways to steer a model?” See what is ControlNet? and ComfyUI explained.
- “Is it safe and legal?” See voice and likeness law and who owns what you generate.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. When does the page say a LoRA is worth the extra work?
- For every single picture
- When you need the character in hundreds of pictures, or in a tool that takes no references
- Only when you have no pictures at all
Show the answer
Answer: When you need the character in hundreds of pictures, or in a tool that takes no references. For a comic or a few dozen pictures, reference pictures are usually enough; a LoRA is for when you need the character everywhere.
2. What kind of training pictures does the page recommend for a character?
- Many copies of the same pose
- The same character in varied settings, expressions and backgrounds
- A different-looking character in each picture
Show the answer
Answer: The same character in varied settings, expressions and backgrounds. Replicate’s advice quoted on the page is varied settings, expressions and backgrounds, and nothing that changes the character, such as different haircuts or ages.
3. How does the page say to test a trained LoRA?
- Run the same prompt and seed with and without it
- Trust it without testing
- Delete it and retrain from scratch
Show the answer
Answer: Run the same prompt and seed with and without it. fal’s advice on the page is to generate with the same prompt and seed with and without the LoRA, then try the character in new places.
- 1Replicate, replicate/fast-flux-trainer (readme), read at source 27 Oct 2026“Aim for a dataset of 10-20 images of your subject.” ↑a“Aim for a dataset of 10-20 images of your subject. Images with resolutions around 1024x1024 are ideal”. ↑b“We’ll run an autocaptioning model on the input images that will generate captions which contain your trigger word. You can also provide your own captions.” ↑c“If you generate images on Replicate with FLUX.1 models and their fine-tunes, then you can use the images commercially. If you download the weights off Replicate and generate images on your own computer, you can’t use the images commercially. Off Replicate, the Flux-Dev LoRAs have the same license as the original base mode for FLUX.1-dev.” ↑d
- 2fal, Training FLUX.2 LoRAs, 25 Nov 2025, read at source 27 Oct 2026“For FLUX.2, you should gather 20 to 1,000 images that reflect the same style, subject, or identity you want to teach the model.” ↑a“For FLUX.2, you should gather 20 to 1,000 images that reflect the same style, subject, or identity you want to teach the model.” ↑b“Each caption should describe the image content and include a unique trigger phrase.” “caption files usually lead to far better LoRA learning retention, especially when your dataset contains a variety of subjects.” ↑c“To test wether your LoRA learned what you intented to teach, generate using the same prompt and seed with and without the LoRA.” and “just make sure it includes your trigger phrase.” (The sentence on this page corrects the spelling slips in the original.) ↑d
- 3Hugging Face, Diffusers: LoRA (training guide), read at source 27 Oct 2026“LoRA (Low-Rank Adaptation of Large Language Models) is a popular and lightweight training technique that significantly reduces the number of trainable parameters. It works by inserting a smaller number of new weights into the model and only these are trained. This makes training with LoRA much faster, memory-efficient, and produces smaller model weights (a few hundred MBs), which are easier to store and share.” ↑
- 4Replicate, ostris/flux-dev-lora-trainer (readme), read at source 27 Oct 2026“Aim for a dataset of 12-18 images of your subject”; “Use high resolution images, ideally around 1024x1024 or larger, but smaller than 1440x1440”. ↑a“For character LoRAs use images of the subject in different settings, facial expressions, and backgrounds. For character LoRAs avoid different haircuts or ages, and showing hands in a lot of face framing positions as we found this led to more hand hallucinations”. ↑b“Trigger word can be a generic proper name (ex. sarah, john).” “For charater LoRAs, pair the trigger work with a gender (man, woman, etc) to improve results.” ↑c“LoRA Rank between 16-32 produce good results. We’ve gone as high as 64 for likeness, and as low as 8 for styles”. ↑d“For more style LORA influence (ex: watercolor or cartoon styles) reducing the lora strength to 0.8 - 0.95 can make a difference”. ↑e“The worse the dataset, the less likely that the model will be flexible enough to apply different art styles to the subject.” ↑f“Increasing the step count of the inference improves LoRA coherence in the case of a weaker dataset.” ↑g
- 5fal, Turbo Flux Trainer (model page), read at source 27 Oct 2026“URL to zip archive with images of a consistent style. Try to use at least 10 images, although more is better.” ↑
- 6Replicate, Fine-tune FLUX.1 to create images of yourself, 30 Aug 2024 (updated May 2025), read at source 27 Oct 2026“You can fine-tune Flux with as few as two training images, but for best results you’ll want to use at least 10 images or more.” ↑
- 7Replicate, Fine-tune FLUX.1 to create images of yourself, read at source 27 Oct 2026“Variety is key. For best results, choose training images with different settings, clothing, lighting, and angles.” ↑a“It should be something unique like MY_UNIQ_TRGGR.” “It should not be an existing word in any language, like dog or cyberpunk.” “It should not be TOK, because it will clash with other fine-tunes if you ever want to combine them.” ↑b“For steps, leave it at 1000. Any less and your training process will not properly learn the concept in your training images.” Replicate, ostris/flux-dev-lora-trainer (readme), read at source 27 Oct 2026: “For steps select a value from 1000-3000”. ↑c“For a 20-minute training (which is typical when using about 20 training images and 1000 steps)”. fal, Training FLUX.2 LoRAs, read at source 27 Oct 2026: “a LoRA may take anywhere from a few minutes to hours for larger datasets”. ↑d“Trainings for the Flux model run on Nvidia H100 GPU hardware”. ↑e
- 8fal, Turbo Flux Trainer, read at source 27 Oct 2026“face_crop (boolean, optional): Whether to try to detect the face and crop the images to the face. Default value: true”. ↑a“Number of steps to train the LoRA on. Default value: 1000”; “Training style to use. Default value: subject”. fal, Training FLUX.2 LoRAs, read at source 27 Oct 2026: “You can change the number of steps and learning rate but for most cases, the defaults work well.” ↑b
- 9GitHub, ostris/ai-toolkit, read at source 27 Oct 2026“Requirements: python >=3.10 (3.12 recommended) Nvidia GPU with enough ram to do what you need python venv git”. ↑
- 10Hugging Face, Diffusers: LoRA, read at source 27 Oct 2026“A full training run takes ~5 hours on a 2080 Ti GPU with 11GB of VRAM.” ↑a“--rank: the inner dimension of the low-rank matrices to train; a higher rank means more trainable parameters”; “--learning_rate: the default learning rate is 1e-4, but with LoRA, you can use a higher learning rate”. ↑b
- 11Hugging Face, Load adapters, undated, read at source 1 Oct 2026 (re-used from this site’s LoRA page): “A value of 0 is the same as only using the base model weights, and a value of 1 is equivalent to using the fully finetuned LoRA.”↑
Read at source 27 Oct 2026: Replicate, flux-dev-lora-trainer readme, fast-flux-trainer readme and Fine-tune FLUX.1 to create images of yourself (30 Aug 2024); fal, Turbo Flux Trainer and Training FLUX.2 LoRAs (25 Nov 2025); Hugging Face, Diffusers: LoRA; GitHub, ostris/ai-toolkit.
Re-used from this site’s LoRA page (read 1 Oct 2026): Hugging Face, Load adapters.
The comparison table, the picture-selection advice for drawn characters, the test list, the advice on trigger words when the pages disagree, and the advice on whose pictures to use are this site’s own, not claims from these sources. No trainer on this page was run for this guide, and none is ranked. Replicate’s readmes are undated, and trainers, sizes and defaults change.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK