# How many images to train a character LoRA

LoRA training: how many images do you need for a face? See what fal, WaveSpeed, Black Forest Labs and Higgsfield publish, and the number that works.

- Canonical: https://aiofmcity.com/blog/how-many-images-to-train-a-lora
- Topic: Character consistency (https://aiofmcity.com/topics/character-consistency)
- Author: Gus
- Published: 2026-10-13

---
## Key points

- Hosted trainers accept as few as four images. That is the least the software needs to run. It is not a recommendation for a face.
- For a character LoRA on Flux, use twenty to forty distinct images.
- An image is distinct when it adds a new angle, light, distance or expression. A second frame from the same shot only adds repetition.
- Faces need more training than objects. In Hugging Face's DreamBooth tests, faces took about 1,200 steps and objects 400.
- Higgsfield's 20 to 80 photos for Soul ID is the only published number written for one person's identity.

---
Use **twenty to forty distinct images** for a character LoRA on Flux. Most trainer pages say four. That number means something else: it is the least the software accepts. A face needs more.

The numbers look confusing because sources publish three different things under the same name. Some give a technical minimum. Some give advice for a style. One gives a number written for a person. Put them side by side and they no longer contradict each other.

## The image numbers each source publishes

| Source | Number | What the number is for |
| --- | --- | --- |
| fal, Flux LoRA fast training | at least 4 | the minimum input; "the more the better" |
| WaveSpeedAI, Flux Dev LoRA Trainer | at least 4 | the minimum input; more where possible |
| Hugging Face, DreamBooth | "a few images" | a subject, with a warning about overfitting |
| Black Forest Labs, FLUX.2 [klein] | 20 to 40 | a style LoRA, not a face |
| Higgsfield, Soul ID | 20 to 80 photos | one person's identity, on a hosted tool |
| This site's [consistency guide](https://aiofmcity.com/guides/character-consistency) | 20 to 40 distinct | a face LoRA on Flux |

> The training input is a zip archive of images, and the parameter description reads: "Try to use at least 4 images in general the more the better." Segmentation masks are on by default, and "for people a face mask is used if possible", which weights the training towards the face.
>
> — Source: fal — Flux LoRA Fast Training API, read 2026-09-15 — https://fal.ai/models/fal-ai/flux-lora-fast-training/api

> The Flux Dev LoRA Trainer takes a URL to a zip archive of images, recommending at least 4 and more where possible, with training steps defaulting to 1,000.
>
> — Source: WaveSpeedAI docs — Flux Dev LoRA Trainer, read 2026-09-14 — https://wavespeed.ai/docs/docs-api/wavespeed-ai/flux-dev-lora-trainer

Both trainers set the minimum at four, and both say straight away that more is better. The minimum tells you when the training job will run. It does not tell you whether the face still holds when the persona turns her head.

> DreamBooth "is a training technique that updates the entire diffusion model by training on just a few images of a subject or style", and the documentation warns that it "is very sensitive to training hyperparameters, and it is easy to overfit."
>
> — Source: Hugging Face Diffusers — DreamBooth training, read 2026-09-15 — https://huggingface.co/docs/diffusers/training/dreambooth

> For a style LoRA, Black Forest Labs gives an "optimal dataset size: 20-40 images": below 20 lacks variation, and above 40 causes "style dilution through excessive variation". It pairs 20 to 40 images with 1,200 to 2,000 training steps.
>
> — Source: Black Forest Labs — FLUX.2 [klein] training example, read 2026-09-15 — https://docs.bfl.ml/flux_2/flux2_klein_training_example

That advice is for a style. A model learns a style from what the images have in common. A face works the other way round: the identity has to be the only thing the images share. The reason for the upper limit still applies to faces. Once new images stop adding information, they teach the model something you did not intend.

> A Soul ID is trained from 20 or more photos of the same person, up to 80, well lit, showing the face from different angles and expressions, without sunglasses, heavy shadows or cropped faces, and including at least one full-height photo.
>
> — Source: Higgsfield Help — Soul ID characters, read 2026-09-14 — https://higgsfield.ai/creator-hub/help-center/ai-models/how-do-i-create-and-use-a-soul-id-character

Higgsfield's number is the only published one written for one person's identity. Its lower limit is 20, the same as the range above. Higgsfield trains a hosted character, so you do not get a LoRA file that you own. [The Higgsfield review](https://aiofmcity.com/tools/higgsfield) explains this.

## Why a face needs more than the minimum

> In Hugging Face's experiments, "a learning rate of 2e-6 with 400 training steps works well for objects but faces required 1e-6 (or 2e-6) with ~1200 steps", and "prior preservation is important to avoid overfitting when training on faces. For other subjects, it doesn't seem to make a huge difference."
>
> — Source: Hugging Face blog — Training Stable Diffusion with DreamBooth, read 2026-09-15 — https://huggingface.co/blog/dreambooth

Those tests used DreamBooth on Stable Diffusion, not a Flux LoRA. So you cannot copy the step counts. The general lesson still applies. A face is harder to learn than an object, and it overfits more easily. Overfitting means the model copies its training images too closely. What helps is a varied set and careful training. A bigger pile of similar frames does not help.

## How to count distinct images

Keep an image only if it adds something that no other image in the set has:

- a new **angle**: front, three-quarter on each side, head tilted up or down;
- a new **light**: window light, direct sun, warm indoor, overcast;
- a new **camera distance**: head and shoulders, mid-shot, full body;
- a new **expression**: neutral, closed smile, open smile, mid-speech.

Four angles, four lights and three distances already give you far more combinations than a set needs. Pick twenty to forty images spread across them. Then everything varies except the face, and the face is what you want the trainer to learn. The [character consistency guide](https://aiofmcity.com/guides/character-consistency) covers the rest: captioning, the images that ruin a set, and the settings that decide how hard the training pushes.

## Signs that you have too many similar images

Three signs tell you that the set has too many similar images, and not too few images:

1. **Something that is not the face follows the persona everywhere.** The same earrings on a beach, the same grey backdrop in a kitchen. Remove the repeats. Do not add more images.
2. **The face only holds from the front.** The set has enough files but too few angles. Add three-quarter and tilted views, and remove some of the frontal ones.
3. **The light will not change.** You prompt for midday sun and still get the soft studio light. Every frame had the same light, so the LoRA learned it as part of her.

### What the sources do not say: Finding a published number meant for a face LoRA

This article was meant to start from the number a trainer vendor recommends for a character. No vendor publishes one.

Every trainer page read for it gives either a minimum (fal and WaveSpeed, both four) or advice for a style (Black Forest Labs, 20 to 40). Hugging Face's DreamBooth documentation says "a few images" and spends its detail on overfitting. The one number written for a person is Higgsfield's 20 to 80. It belongs to a hosted identity feature and not to a LoRA.

So the working range above comes from the [character consistency guide](https://aiofmcity.com/guides/character-consistency). The table says where each number comes from, so that no minimum is passed off as a recommendation.

The range in this article comes from [AI character consistency](https://aiofmcity.com/guides/character-consistency), the full guide behind the number. If you want to train a LoRA by API, without a GPU of your own, see [the WaveSpeedAI review](https://aiofmcity.com/tools/wavespeed).

## Frequently asked

### How many images do you need to train a character LoRA?

For a face on Flux, use twenty to forty distinct images. Hosted trainers accept as few as four. Four is only the least the software will run with. It does not give you a stable face.

### Can you train a LoRA with too many images?

Yes. At some point new images stop adding a new angle, light or distance. After that they only add repetition. The LoRA learns whatever repeats as part of the identity: the same earrings, the same backdrop, the same main light.

### Why do LoRA trainers say at least 4 images?

Because four is roughly what the training job needs to run. fal and WaveSpeed both give it as a minimum, and both add that more is better. Neither says four is enough for a consistent person.

### Is the image count different for a style LoRA?

Black Forest Labs publishes 20 to 40 images as the optimal range for a FLUX.2 style LoRA. It warns that fewer images give too little variation and more images dilute the style. That advice is for a style and not for a face, but what it says about variation also applies to faces.