How to Train Your Own Character LoRA on Raydance.ai
A character LoRA teaches a model one specific face and body, so the same person shows up in every image you generate - new outfit, new location, new lighting, same character. Training one used to mean renting a GPU, installing a trainer and guessing at a config file. The Raydance.ai LoRA Trainer turns it into four screens: Subject, Dataset, Train, Result.
This guide walks through all four with a real 28-photo character dataset. The tool does the heavy lifting, but the result is decided by your dataset - so that is where we spend the most time.
What You'll Learn
What You Need
- A Raydance.ai account with credits (our example run costs 724 credits on the Balanced preset)
- 10-25 photos of one character, at least 768px on the short side
- The rights to those photos - your own face, your own AI character, or someone who explicitly agreed
Part 1: Build the Dataset
Nothing you set later in the trainer can rescue a weak dataset. Get this part right and the default settings will do the rest.
What a Good Character Dataset Looks Like
A LoRA learns whatever your photos have in common. If the only thing they share is the character, it learns the character. If they also share a bedroom, a denim jacket and a phone in the left hand, it learns those too - and paints them into every image.
So the rule is simple: keep the character constant, change everything else.
Six photos from our example dataset: the same character, six different framings, outfits and lighting setups
| Aim for | Why it matters | | --- | --- | | 10-25 photos | Enough variety to learn from. Past roughly 30 you mostly pay for more training steps. | | A third close-ups | The face is what makes the character recognisable - give the model plenty of it. | | Half-body and full-body shots | Otherwise the LoRA only knows the face and invents the rest. | | Different outfits and backgrounds | Stops clothes and rooms from being baked into the character. | | Different lighting | Daylight, golden hour, night, indoor - the LoRA stays flexible. | | Different expressions and angles | Front, three-quarter, profile, looking away, smiling, neutral. | | Sharp, 768px+ on the short side | Soft or tiny photos produce a soft LoRA. |
The Most Common Mistake: Near-Duplicates
Five shots from the same burst are not five training images. They add training time and cost, teach nothing new, and pull the LoRA towards that one pose and room. Keep the best frame and delete the rest.
Two frames from the same moment - same pose, same room, same outfit. Keep one.
Leave these out:
- Group shots or photos where another face is visible
- Sunglasses, masks, hands or hair covering most of the face
- Heavy beauty filters, stickers, text overlays and watermarks
- Blurry, dark or heavily compressed images
- Anything showing nudity, a minor or a public figure - the trainer blocks these (more in Step 5)
No Photos Yet? Generate the Dataset
Our example character is AI-generated. If you are building a virtual character, generate one strong reference image first, then create variations of it in different scenes, outfits and framings until you have 20-30 candidates. Pick the best 15-25 - consistency of the face matters more than the number of images.
Collect the Photos in Your Library
You can upload photos straight into the trainer, but keeping them as a dataset in your Library is the better habit: the set stays reusable, you can train it again on a different base model, and you can review everything in one grid before spending credits.
- Open "Library" in the left sidebar of the AI Suite
- Create a dataset and give it a clear name
- Click "Add media" and add your photos
- Scan the grid for near-duplicates and weak shots and remove them now
Our "Character" dataset in the Library - 28 photos before the cleanup
Spot the Problem
Look at the first two rows: three almost identical mirror selfies on the same rug and two near-identical shots on the same bed. That is exactly the kind of repetition the dataset check will flag in Step 5.
Part 2: Train the LoRA
Open "LoRA Trainer" in the left sidebar under Tools. The wizard at the top shows where you are: 1 Subject, 2 Dataset, 3 Train, 4 Result. Your LoRAs - drafts, running trainings and finished ones - are listed under My LoRAs on the left.
Step 1: Name Your LoRA and Choose What It Learns
- Click "New LoRA" above the My LoRAs list
- Enter a name - this is only the label you see in your list
- Select "Person / character" under What should it learn?
The Subject step: name, subject type, base model and trigger word on one screen
The subject type is not cosmetic. It decides which training presets you get in Step 6 - a character is trained differently from a style or a product.
Step 2: Pick the Base Model
A LoRA only works with the model it was trained for, so choose the model you want to generate with afterwards.
The four base models, each with its minimum photo count and price level
| Base model | Best for | Min. photos | Cost | | --- | --- | --- | --- | | Krea 2 | Running on Raydance Red pods and stacking with the VIP LoRAs | 6 | Standard | | Z-Image Turbo | Photorealistic characters, fast and cheap - the recommended pick | 10 | Low | | FLUX.2 [dev] | Maximum detail and prompt following | 10 | Premium | | Qwen Image | Text, illustration and graphic styles | 10 | Low |
Which one should you pick?
- You already generate on Raydance Red and want your character together with the VIP LoRAs: Krea 2. That is what we use in this guide.
- You want a realistic character at the lowest price: Z-Image Turbo.
- You need the finest detail and are happy to pay for it: FLUX.2 [dev].
Pro Tip
Unsure? You can change the base model for as long as the LoRA is still a draft. It is only locked once training starts.
Step 3: Set the Trigger Word
The trigger word is the token the LoRA is tied to. The trainer suggests one from your LoRA name plus a short random tail.
The trigger word field - the wand button suggests a new one
Keep the random tail. A plain name like anna already means something to the base model - it drags in every Anna the model has ever seen. A made-up token like anna_k7q means nothing, so the model attaches it to your character alone.
- Lowercase letters, numbers and underscores only, 3-32 characters, starting with a letter
- Click the wand button to roll a new suggestion
- Click "Continue" when you are happy
Fixed Once Training Starts
You can rename the LoRA any time, but the trigger word is locked the moment training begins - it is baked into the weights.
Step 4: Add Your Photos
The Dataset step starts empty. The panel on the right tells you what is still missing - here, six usable images for Krea 2.
An empty training set: drop photos in, or pull a whole dataset from your Library
You have two ways to fill it:
- Drop photos into the upload area (JPG, PNG or WebP, up to 100 images), or
- Click "Add from library", select your dataset and click "Add images"
Importing the "Character" dataset from the Library
Your Library Dataset Stays Untouched
The photos are copied into the training set. Removing an image in the trainer does not delete it from your Library, and the Library dataset is not deleted after training.
Every image is checked and captioned as it comes in. A 28-photo import takes about a minute - wait for the "images added" confirmation before you continue.
Step 5: Read the Dataset Check
This is the step that separates a good LoRA from a wasted training run. Every photo gets a badge, and the panel on the right sums up the whole set.
Our 28 photos after the import: "Not ready yet", with three problems listed
Our import came back "Not ready yet". The readiness panel lists the reasons, most important first:
The readiness panel: red entries block training, grey entries are advice
- Red entries must be fixed before the "Continue to training" button unlocks
- Grey entries are advice - you can train anyway, but the result is better if you act on them
What the badges on each photo mean:
| Badge | Meaning | What to do | | --- | --- | --- | | OK | Passed the check and counts as usable | Nothing | | Blocked | Can't be used - the reason is shown on the card | Remove it. A block can't be overridden. | | Not checked | The check didn't finish | Click "Check again", or remove the image | | Duplicate | Nearly identical to an earlier photo | Keep the better one, remove the other | | Blurry | Visibly soft | Replace it with a sharp photo | | Low res | Under 768px on the short side | Replace it if you can | | Too small | Under 512px on the short side - not counted as usable | Remove or replace it |
A blocked image next to accepted ones. Hover any card to reveal the remove button in its top right corner.
Why Images Get Blocked
The trainer refuses photos it reads as showing a public figure, a minor or nudity. The check is deliberately strict and can be over-cautious - it flagged three images of our AI-generated character. Don't fight it: remove the blocked images and move on. With 20+ good photos left you lose nothing.
Our cleanup: we removed the three blocked images, the two that could not be checked, one near-duplicate and one weak beach shot. That left 21 photos - and the panel switched to "Great dataset".
21 usable images, score 100/100, and "Continue to training" is unlocked
Aim for Green, Not for 100 Images
Ten sharp, distinct photos are already enough for a full score. Twenty clean, varied images beat sixty with duplicates and blur - and they are cheaper to train.
Check the captions
Under every photo sits an automatically written caption, starting with your trigger word. Captions tell the trainer what in the picture is not the character.
Each caption starts with the trigger word, then describes pose, outfit and setting
The principle: everything you describe stays changeable, everything you leave out becomes part of the trigger word.
- Describe the outfit, pose, location, lighting and camera angle
- Don't describe what should always belong to the character - face shape, eye colour, freckles, hair colour
Look at the last card in the blocked-images cut-out above: its caption mentions green eyes and freckles. Those are traits of the character itself, so we click into the caption and delete them. Otherwise the LoRA learns that it only needs to draw those features when the prompt asks for them.
Note
You don't need to rewrite every caption. Skim them, fix what is plainly wrong, and remove descriptions of the character's permanent features. Two minutes here is well spent.
Click "Continue to training" when the panel is green.
Step 6: Choose a Training Preset
Each preset is priced for your exact dataset, so the numbers you see are the numbers you pay.
Fast, Balanced and Max quality - with credits, steps, resolution and expected duration
Here is what our 21-photo dataset costs on Krea 2:
| Preset | Steps | Resolution | Credits | Time | Use it for | | --- | --- | --- | --- | --- | --- | | Fast | 1,000 | 768px | 496 | ~11 min | A quick first version to test the idea | | Balanced | 1,500 | 768px | 724 | ~16 min | The best mix of likeness, flexibility and price | | Max quality | 2,550 | 1024px | 1,202 | ~40 min | Finest skin and fabric detail |
Start with Balanced. It is the recommended preset for a reason: enough steps for a strong likeness, not so many that the LoRA starts copying your photos.
Why Your Numbers Will Differ
Steps scale gently with the size of your dataset - a bigger set needs more passes to be seen at all. More photos means slightly more steps and a slightly higher price, but doubling the photos does not double the cost.
Don't Jump Straight to Max Quality
If Balanced gives you a weak likeness, the cause is almost always the dataset, not the step count. Fix the photos first. Max quality is worth it once Balanced already looks right and you want more detail at 1024px.
Step 7: Advanced Settings (Optional)
Open "Advanced settings" only if you know why you want to change something. It replaces the preset with your own values and shows the price live.
Advanced settings: steps, learning rate and resolution, priced as you type
- Steps (300-3,000) - how long the model trains on your photos. More steps capture more likeness, but too many make the LoRA rigid: it reproduces your photos instead of following the prompt.
- Learning rate - how strongly each step changes the model. Higher learns faster but can overshoot and cause artifacts; lower is steadier and needs more steps.
- Resolution (Krea 2 only) - 1024px keeps finer detail but takes longer and costs more; 768px is enough for most characters.
How to read a finished LoRA and what to change next time:
| What you see | What it means | Next run | | --- | --- | --- | | Likeness is weak even at strength 1.0 | Undertrained | More steps, or more close-ups in the dataset | | Every image looks like a training photo | Overtrained | Fewer steps, more variety in the dataset | | Waxy skin, harsh contrast, artifacts | Learning rate too high | One notch lower | | Same outfit or room keeps appearing | Dataset bias | Remove the repeats, fix the captions |
Step 8: Confirm Your Rights and Start
Below the presets you confirm that you may use the photos, and you see the final price next to your balance.
The consent checkbox, the price for the selected preset and the "Start training" button
- Read and tick "I have the rights to these images" - you confirm that you are the person shown or have their explicit consent, and that the set contains no minors, public figures or nudity
- Compare the price with your balance
- Click "Start training"
When You Are Charged
The credits are charged when training starts. If the training fails, the full amount is refunded automatically - you don't pay for a broken run.
Only Train People Who Agreed
A character LoRA can put a real face into any scene. Train your own face, your own invented character, or someone who has clearly said yes - nothing else.
Step 9: Wait for the Training
The wizard jumps to Result and shows the live status. First the job waits for a free GPU:
Queued: "Waiting for a GPU…" with the elapsed time and the usual duration
Then the progress bar starts moving:
Training in progress at 42% - about seven minutes into a 16-minute run
- Waiting for a GPU… - the job is queued
- Training your LoRA… - the actual training
- Saving your LoRA… - the weights are being stored
You don't have to keep the page open. The training runs on Raydance's side, and you get a notification as soon as the LoRA is ready. It also stays in My LoRAs with a "Training" badge, so you can come back any time.
Pro Tip
Use the wait to write three or four test prompts - a close-up, a full-body shot, and one scene that is nothing like your training photos. That last one tells you how flexible the LoRA really is.
Step 10: Use Your LoRA
When training finishes, the Result step shows your LoRA with everything you need to use it.
The finished LoRA: trigger word, default strength, and the buttons to use or download it
- Click "Use in Krea 2" - the Image Generator opens with the right model and your LoRA already selected
- Write your prompt as usual - the trigger word is added automatically when the LoRA is active
- Generate and compare the result with your training photos
Dial in the strength
Default strength - the value the LoRA starts with in every generation
- Start between 0.8 and 1.0
- Lower it if images look over-cooked - harsh, stiff, or too close to a training photo
- Raise it if the likeness is weak
- The value you set here becomes the default for this LoRA; you can still change it per generation
More you can do from here:
- Stack it - combine up to three of your own LoRAs with the Krea 2 VIP LoRAs, for example your character plus a style
- Download .safetensors - take the weights with you and use them in ComfyUI or any other tool that supports the base model
- Train again - not happy? Click "New LoRA", import the same Library dataset, and change one thing
Troubleshooting
| Problem | Likely cause | Fix | | --- | --- | --- | | "Continue to training" stays greyed out | Blocked or unchecked images, or too few usable ones | Read the red entries in the readiness panel and clear them | | Face looks like a relative, not the character | Too few close-ups | Add 5-8 sharp close-ups from different angles and retrain | | The same outfit appears in every image | Outfit repeated across the dataset | Remove repeats and make sure captions name the clothing | | The same background keeps coming back | Too many photos from one location | Vary locations; name the setting in the captions | | Character ignores the prompt | Overtrained, or strength too high | Lower the strength to 0.7-0.8 first; retrain with fewer steps if that isn't enough | | Skin looks plastic or burnt | Strength too high or learning rate too high | Lower the strength; next run, one notch lower learning rate | | "You need … credits to start this training" | Balance below the preset price | Top up, or switch to the Fast preset |
Recommended Workflow
The loop that gets you a good character LoRA for the fewest credits:
- Collect 20-30 candidate photos in a Library dataset and delete duplicates right there
- Import the dataset into the trainer and clear everything red in the dataset check
- Skim the captions and remove descriptions of the character's permanent features
- Train once on Balanced - don't touch the advanced settings yet
- Test with three prompts at strength 0.8, 1.0 and 1.2
- Change one thing if you retrain: either the dataset or one setting, never both
Pro Tip
Keep a short note per LoRA: number of photos, preset, steps and what you thought of the result. After two or three trainings you will know exactly what your character needs - and stop paying for experiments.
Wrapping Up
You now have the full path from a folder of photos to a trained character LoRA: a varied dataset, a clean dataset check, a sensible preset and a strength that fits. The trainer handles the GPUs and the config - your job is the part no tool can do for you: choosing good photos.
New to the platform? Start with How to Run Custom Models on Raydance.ai to set up your account and your first pod, then come back and train your character.
Happy training!