- LoRA training
- Dataset size
- Flux
- SDXL
- SD1.5
How Many Images Do You Need to Train a LoRA? (Flux, SDXL, SD1.5)
Character LoRAs need 15–30 diverse images; styles need 50–200. Learn the ranges by LoRA type, the steps math, and the signs you have too few or too many.
“How many images do I need?” is the first question every LoRA trainer asks, and most answers online are either a single magic number or a shrug. The honest answer is a range that depends on what you are teaching the model, which base model you train on, and how diverse your images are. This guide gives you the ranges that hold up in practice, explains the math that connects image count to training steps, and shows how to tell when you have too few or too many.
Recommended image counts by LoRA type
The table below reflects the defaults that recur across the Kohya sd-scripts documentation, Civitai’s training guides, and thousands of published character LoRAs. Treat the low end as the minimum for a usable result and the high end as the point where adding more images stops helping.
| LoRA type | Images | Why this range |
|---|---|---|
| Character face and identity | 15–30 | Enough angles and expressions to generalize without memorizing one photo. |
| Character with fixed outfit | 25–40 | Outfit details need more full-body and mid-shot coverage than a face alone. |
| Object or product | 15–30 | Rigid subjects need fewer samples; cover every side and two lighting setups. |
| Art style | 50–200 | Styles are learned from variety of subject matter, so more images help. |
| Pose or concept | 30–60 | The model must see the concept across many different subjects. |
For a character LoRA the sweet spot is about 20 images. Below 12 images the model tends to reproduce your source backgrounds and poses. Above 40 images, gains flatten unless every extra image adds a new angle, lighting condition, or expression.
Why the number varies
Four variables move the target up or down.
- Base model. Flux.1 learns identities from fewer images than SDXL or SD1.5 because its 12-billion-parameter transformer and T5 text encoder capture detail faster. Many Flux character LoRAs train well on 10–20 images, while SD1.5 often wants 20–30.
- Subject complexity. A face is a small, consistent region. An outfit with patterns, armor, or accessories adds surface area the model must learn.
- Image diversity. Twenty near-duplicate selfies teach less than twelve images that vary angle, lighting, distance, and background. Diversity is the real lever, not count.
- Training budget. Image count, repeats, and epochs together determine total steps. You can compensate for a small set with more repeats, up to a point.
10,000×
fewer trainable parameters with LoRA versus full fine-tuning of GPT-3 175B
Source: Hu et al., LoRA paper (2021)12B
parameters in Flux.1, the largest open-weights image model commonly trained with LoRA
Source: Black Forest LabsHow image count relates to steps and epochs
Trainers do not think in images, they think in steps. One step is one optimizer update on one batch. The relationship is:
total steps = (images × repeats × epochs) ÷ batch sizeCommunity defaults for a character LoRA land around 1,500–3,000 total steps on SDXL and SD1.5, and often 1,000–2,500 steps on Flux. With 20 images, 10 repeats, 10 epochs, and a batch size of 1, you get 2,000 steps, which is why “20 images, 10 repeats” is such a common starting point in Kohya folder names like 10_ohwx person.
Rule of thumb
Signs you have too few images
These symptoms show up at generation time, usually within the first few test prompts:
- The background from your source photos appears even when the prompt asks for a different scene.
- The character’s pose or head angle barely changes between seeds.
- You cannot change the outfit, hairstyle, or expression without losing the face.
- Images look sharper or more saturated than the base model’s normal output, a sign of overfitting.
The fix is not more copies of the same photo. It is more different photos. If you only have one good reference, you can generate a diverse dataset from a single image rather than scraping low-quality extras.
Signs you have too many, or too many bad ones
- Training takes hours but the likeness is fuzzier than a smaller earlier run.
- The model averages features across inconsistent sources, producing a “cousin” of your character.
- Blurry, low-resolution, or heavily filtered images teach the model those artifacts.
Every image in the set should be one you would be happy to see the model reproduce. Cutting a 60-image set down to the best 25 frequently improves results. If the likeness wanders between sources, read our guide to diagnosing and fixing identity drift.
The diversity checklist that matters more than count
For a 20-image character set, aim for coverage across these axes:
| Axis | Target coverage in 20 images |
|---|---|
| Camera angle | Front, three-quarter left, three-quarter right, profile, slightly above, slightly below |
| Framing | 8–10 close-ups, 6–8 medium shots, 3–4 full body |
| Lighting | Soft daylight, hard sun, indoor warm, studio, one low-key or dramatic setup |
| Expression | Neutral, smiling, serious, at least one open-mouth or laughing frame |
| Background | No two images share the same background; mix plain and busy scenes |
| Wardrobe | At least 3 different outfits unless the outfit is part of the character |
Model-specific guidance
Flux.1
Flux picks up identity quickly. Start with 15–20 images and natural-language captions. Overshooting on Flux shows up as a plastic, over-smoothed skin look. See Flux vs SDXL LoRA training for a full comparison.
SDXL
SDXL handles 20–30 images comfortably at 1024×1024. Its two CLIP encoders respond well to both tag-style and short natural captions. Aim for 2,000–3,000 total steps.
SD1.5
SD1.5 usually needs the top of the range, 20–30 images at 512×512 or 768×768, and benefits from regularization images to keep the base class from collapsing into your subject.
How to reach 20 diverse images when you only have one
Most creators do not have 20 professionally lit photos of their character from six angles. LoRA Dataset solves this by taking one reference image, extracting the persistent identity traits, and generating 20 or more variations that change pose, lighting, framing, and background while keeping the face stable. Each image ships with a caption that starts with your trigger word, so the export drops straight into Kohya, OneTrainer, or Civitai’s on-site trainer. The tool is free and runs on your own Google Gemini API key.
Frequently asked questions
Can I train a LoRA with only 5 images?
You can, but the model will likely memorize those five images and reproduce their backgrounds and poses. Five images work best as a starting point for generating a larger, more diverse set of 15 to 30 images before training.
Is 100 images too many for a character LoRA?
Usually yes. Beyond 40 images, extra samples rarely add new information about a face, and inconsistent images start to blur the likeness. Curate down to the best 20 to 30 unless the character has complex outfits or many distinct looks.
How many images do I need for a Flux LoRA?
Flux character LoRAs typically train well on 10 to 20 images because the 12B model learns identity quickly. Use natural-language captions and watch for over-smoothed skin, which signals overtraining.
Do more images mean I need more training steps?
Not necessarily. Total steps equal images times repeats times epochs divided by batch size. If you add images, lower the repeats so total steps stay in the 1,500 to 3,000 range for SDXL and SD1.5.
Should I include regularization images?
Regularization images help SD1.5 and SDXL keep the general class (such as "woman" or "man") from collapsing into your subject. They are less commonly used for Flux LoRAs. They are separate from your subject images and do not count toward the 15 to 30 recommendation.