How Many Images Do You Need to Train a Face LoRA? (What the Trainer Docs Say)

You need about 10 to 20 good images to train a face LoRA, and 20 to 40 when the same LoRA also has to carry the body, outfits and scenes. That is where the trainers' own documentation lands for FLUX.1, FLUX.2, SDXL, Z-Image and Qwen-Image: the lowest documented working number is 2, and most recommendations sit between 10 and 20.

Numbers checked in October 2026 against Replicate's FLUX fine-tuning guide, Black Forest Labs' FLUX.2 [klein] training guide and training example, Scenario's training help, Astria's docs, fal's trainer API reference and Hugging Face's SDXL LoRA write-up. This is a survey of published recommendations. We did not run a training benchmark for it, and nothing below is a measured result of ours.

The question gets a different answer in every forum thread because people mix up three jobs: a face-only LoRA, a full character LoRA and a style LoRA. They need different amounts of data, and the documentation says so once you read it side by side. This page separates the three, lists every number with its source and date, and ends with which number to start from. If you are still deciding between a LoRA and a reference-image method, our consistent character AI guide compares all five approaches first.

How many images to train a face LoRA: what the docs agree on

Nine documented figures, five findings. Each one is traced to its document in the table further down.

• 10 to 20 is the documented centre for a face. Replicate says at least 10 in its docs and 12 to 20 in its launch post. Black Forest Labs sizes a character LoRA at 10 to 15 images. Astria builds a person model from around 16. Scenario calls 5 to 15 the sweet spot.
• Very small sets are documented to work. Replicate says as few as two images, fal says at least 4, and Hugging Face trained SDXL face LoRAs on 6 to 10 photos, with good results from six well-chosen ones.
• 20 to 40 is the range for broader LoRAs. Black Forest Labs calls 20 to 40 images the optimal dataset size in its FLUX.2 [klein] training example, which trains a style. It is also a sensible ceiling for a full character.
• Past that, the docs start warning you. Above 40 images dilutes what the LoRA learns (Black Forest Labs), too many similar images cause overfitting (Scenario), and fewer curated images beat more average ones (Hugging Face).
• The number has barely moved in four years. DreamBooth asked for 3 to 5 images in 2022. The newest hosted trainers ask for 5 to 15 in 2026.

Method: where these numbers come from

We read the dataset guidance in the official documentation of LoRA trainers and hosted training services, plus two primary write-ups from the teams that built the methods. For each one we recorded the number it gives for training a person, face, character or single subject, the base model it applies to, and when it was published or last checked. Where a source gives a range, the table shows the range.

Two kinds of source were left out. Forum threads and third-party blog posts were excluded because we could not verify the runs behind their numbers. And the open-source trainers most people use locally publish settings, not dataset sizes: the sd-scripts training docs, the AI Toolkit README and the FluxGym README give no recommended image count at all. FluxGym's code only sets an upper limit of 150 images per dataset.

Face LoRA dataset size by base model and trainer

Here is every figure with the model it was written for. Read across a row before you compare down a column, because a number for a hosted FLUX.1 trainer with auto-captioning is not the same claim as a number for a local SDXL run.

Sources, in table order: Replicate docs, Replicate launch post, fal API reference, Black Forest Labs klein training example (both rows), Scenario character guide, Astria AI photoshoot guide, Hugging Face and the DreamBooth project page.

Two gaps are worth knowing about. First, nobody publishes a separate number for Illustrious or Pony. Both are SDXL fine-tunes, so the SDXL guidance is the closest thing on record; Civitai's trainer guide, which lists both as base models, states no minimum, and its worked examples use 15 to 30 images. Second, the Black Forest Labs "20 to 40" figure sits in a style-training example. The same page's step table is where the character figure of 10 to 15 images appears.

Has the number changed from 2022 to 2026?

Less than you would expect. Base models got far larger over these four years, and the documented dataset size stayed inside the same small band.

A reasonable reading is that a stronger base model already knows what a face looks like and only has to learn which face, so it does not need more examples than a weaker one did. That is our interpretation, not something the docs state.

The bigger change is that some jobs no longer need training at all. Reference-image models now take several photos of a character at generation time; Black Forest Labs' release notes list FLUX 3 Image, launched on October 1, 2026, with up to ten references per request. Our Flux consistent character guide covers when references are enough and when a LoRA still earns its training run.

How many images for a character LoRA?

Plan on 20 to 40. A face LoRA has one job, which is the face. A character LoRA has to hold the face plus body proportions, a few outfits and the way the character sits in a scene, so it needs coverage a face-only set does not have.

Astria's guide shows the shape of a mixed set: 4 to 8 portrait or waist-up photos, about 4 full-body shots, and the instruction to upload both kinds. Scale that up and you arrive in the 20 to 40 range that Black Forest Labs calls optimal. Our character LoRA dataset checklist turns that into a 40-shot capture plan: shoot or generate all 40, then keep the 20 to 40 that survive culling.

Face LoRA dataset size: why more images can hurt

More images help only while each one adds something new. The docs describe three ways a bigger set makes a worse LoRA.

• Near-duplicates multiply exposure. Scenario's character guide says too many similar images push the model toward overfitting. Twenty selfies from one afternoon teach the trainer one image twenty times.
• Too much variation dilutes. Black Forest Labs puts the limit at 40 images for a style and describes what happens past it as dilution. For a face the equivalent is a set where age, makeup or styling changes so much that the LoRA averages them.
• Weak images set the quality bar. Hugging Face's write-up says fewer well-curated images work better than more images of mid-to-low quality. The LoRA cannot tell your best photo from your blurriest.

LoRA training images needed: what counts as one

Every figure above assumes usable images. An image counts toward your 10 to 20 only if it passes the checks the docs repeat.

For a real person, such as your own AI twin, only train on a face with that person's written consent. For a persona who does not exist, you generate the set first; the LoRA training guide covers building references for a synthetic character and the trainer settings that follow.

Dataset size sets the step count

The image count is half of the exposure equation. The other half is how many times the trainer sees each image, and several docs express that per image.

• Astria: its portrait preset defaults to 27 steps per image with a 300-step minimum, and its High preset to 100 steps per image (FLUX fine-tuning docs).
• Hugging Face, SDXL: tested 75, 100 and 120 steps per image on a six-photo face set and reported good results at 120 when the set is diverse enough not to overfit.
• Black Forest Labs, FLUX.2 [klein]: 800 to 1,200 steps for a character LoRA of 10 to 15 images.

Worked through for 15 images: 405 steps on Astria's portrait preset, 1,800 at Hugging Face's 120 multiplier on SDXL, and 800 to 1,200 on klein. Different models and trainers, so the totals do not compare directly. The point is the direction: you pick the images first and derive the steps from them. The free LoRA Training Steps Calculator does the kohya arithmetic for images, repeats, epochs and batch size, and our guide to face LoRA training steps and learning rate lists the documented targets per trainer.

Which number to start with

Pick the row that matches the job, train once, and let the samples tell you what is missing. This table is our reading of the documents above, not a test result.

That loop is where most of the work is, and it is the part a number cannot do for you. The AI Influencers program picks up from here: designing the persona, generating the dataset when the character does not exist yet, testing checkpoints and turning the finished LoRA into a posting system.

Limitations of this compilation

• Documentation is not a benchmark. No source here publishes a controlled comparison of dataset sizes. The figures are vendor guidance.
• Defaults differ. Each number assumes that vendor's captioning, resolution, step count and learning rate. Move one and the best image count can move too.
• Some sources are old. The DreamBooth and Hugging Face figures predate FLUX. They are included to show the trend, not as current settings.
• Style and character get mixed. The 20 to 40 figure comes from a style example. We apply it to full characters as a ceiling, which is a judgment call.
• No runs of our own. We did not train LoRAs for this page. When a planner tool or our own run data is published, this page will link to it and say what was measured.

Face LoRA dataset size: FAQ