Train your first LoRA of your face, your style or your product, and test it properly

Decide what you are teaching, and teach one thing

A LoRA learns whatever your images have in common. Pick one concept: a face (you or a consenting person), a style (your illustration or photo look) or a product (one item, one colourway). A face and a style in one dataset gives you a LoRA that does neither well.

If you only need a look for a few images, try the tool's reference features first: Midjourney has no LoRA training, but Moodboards and Personalization profiles do a similar job for style. Train when you need the same face, style or product across dozens of images.

Choose where to train

No-code, inside one app: Krea trains a style, character or object on Krea 2 from a minimum of 3 images (10 to 30 recommended). Magnific (formerly Freepik) Custom Characters want 12 to 24 images and Custom Styles 10 to 50, on paid plans. Higgsfield Soul ID trains a person from 20 to 80 photos in a few minutes. You can't download any of these: they only work in that app.

A file you own: Replicate, fal and Civitai give you a LoRA file you can run elsewhere and mix with others. Replicate's fast-flux-trainer and fal's flux-lora-fast-training both take a zip of images and a trigger word and train FLUX.1 [dev]. Civitai trains SD 1.5, SDXL, Flux.1 and Flux.2 Dev and more. With a strong GPU you can train locally with Ostris AI Toolkit, which has a web UI; its FLUX example config is written for 24 GB of VRAM.

Gather 10 to 20 good images

Replicate's trainer pages ask for 10 to 20 images (12 to 18 for the ostris trainer) at around 1024 × 1024. fal recommends 1024 × 1024 or larger with no compression artefacts. Quality beats quantity: every flaw that repeats across images gets learned.

Face: different settings, expressions and backgrounds, with the same haircut and age; Replicate warns that lots of hands framing the face led to more hand errors. Style: varied subjects in a consistent style, and no single element dominating. Product: several angles and lights, on plain backgrounds and in context, with the same colourway throughout.

Use only images you own or have permission to use. For a face that isn't yours, get the person's consent first.

I'm training a [face / style / product] LoRA. Here are my candidate images. Check them as a dataset: flag near-duplicates, blur, heavy compression, crops that cut the subject, anything inconsistent (different haircut, colourway or style), and anything that repeats in too many images (same background, same pose, a watermark). Recommend which 10 to 20 to keep and why, and what kinds of image are missing.

Pick a trigger word and caption the images

The trigger word is the switch: it is attached to every training image, and you type it in prompts to call the LoRA up. Use a made-up string that isn't a real word, such as TOK, or a coded one like CYBRPNK for a style (Replicate's advice). For a person, pair it with a plain noun at generation time: "a photo of TOK woman".

Replicate and fal can caption automatically. To write your own, put a .txt file with the same name next to each image (img_0.jpg and img_0.txt). In captions, describe what changes from image to image (clothing, background, light) and leave out what you want learned, so the trigger word absorbs it. That is common practice rather than a documented rule, but it is why a good LoRA can put your face in new clothes.

Write a training caption for each of these images for a [face / style / product] LoRA with the trigger word [TOK]. Start each caption with the trigger word. Describe only what varies between images: pose, clothing, background, lighting, framing. Don't describe the [face / style / product] itself. One line each, under 30 words, as a list of filename: caption.

Start with the trainer's defaults

Don't tune on your first run. Replicate suggests 1,000 steps to start; its ostris trainer accepts 1,000 to 3,000 and says ranks of 16 to 32 work well (as low as 8 for styles, up to 64 for likeness). fal's trainers default to 1,000 steps; set is_style to true for a style, which skips auto-captioning. Replicate's fast trainer asks you to choose subject or style.

Replicate's 2024 FLUX guide put 1,000 steps at 20 to 30 minutes and under $2; its newer fast trainer runs on eight H200s at $0.0122 a second. fal's FLUX.1 fast trainer is a flat $2 a run and its FLUX.2 [dev] trainer $0.0064 a step. Civitai shows the exact Buzz cost before you submit.

Test it with a fixed grid

Run the same four or five prompts at three strengths (0.6, 0.8 and 1.0) with a locked seed. Include one prompt close to your training images and three that push away from them: a new setting, a new outfit or object, a different medium.

Too weak: the face isn't recognisable or the style barely shows, even at 1.0. Train more steps or improve the images. Too strong: every result has the training background, pose or lighting, and other styles won't take. Lower the strength (Replicate's trainer notes suggest trying 0.8 to 0.95 for style LoRAs), or retrain with fewer steps or more varied images.

Write a LoRA test grid for my [face / style / product] LoRA, trigger word [TOK]. Five prompts: one close to my training images ([DESCRIBE THEM]), and four that push it: a different setting, a different outfit or object, a different art medium, and a group or busy scene. Same seed for all. Tell me what a too-weak and a too-strong result would look like for each prompt.

Use it, mix it and mind the licence

On Replicate the training creates a model at your destination: run it with the trigger word in the prompt. In Automatic1111 and Forge, put the file in models/Lora and write <lora:filename:0.8> in the prompt (the number is the strength, usually 0 to 1). In ComfyUI, add a LoraLoader node (strength_model and strength_clip default to 1.0) and chain a second loader to mix two.

When you mix LoRAs, lower each one, for example to 0.6 or 0.7, so they don't fight. Your face plus your illustration style is a classic pair.

Check the base model's licence before commercial use. Replicate's note for FLUX.1 [dev] LoRAs: images generated on Replicate can be used commercially, but if you download the weights and generate on your own computer, you can't.

Trainers at a glance (as of October 2026)

Replicate fast-flux-trainer: zip of 10 to 20 images around 1024 px, trigger word, subject or style; auto-captions or your .txt files; 8x H200 at $0.0122 a second; result is a FLUX.1 [dev] model you run on Replicate, with downloadable weights.

Replicate ostris/flux-dev-lora-trainer: 12 to 18 images, steps 1,000 to 3,000, rank 16 to 32; 1x H200 at $0.001525 a second.

fal flux-lora-fast-training: zip of images, trigger word, steps default 1,000, is_style for styles; $2 a run; run it with fal's flux-lora endpoint and a strength.

fal flux-2-trainer: FLUX.2 [dev]; at least 10 images; steps default 1,000 (100 to 10,000); learning rate default 0.00005; $0.0064 a step; LoRA output in fal or ComfyUI format.

Civitai: Create > Train a model; SD 1.5, SDXL, Pony, Illustrious, Flux.1 Dev, Flux.2 Dev and Klein, Qwen-Image, Wan and others; from 500 Buzz for SD 1.5 and SDXL, more for Flux and video; exact cost shown before you submit.

Krea: at least 3 images, 10 to 30 recommended; style, character or object on Krea 2; usable across Krea's image, edit, enhancer and video tools.

Magnific: Custom Characters 12 to 24 images, 5,500 credits; Custom Styles 10 to 50 images (at least 2,000 px wide or tall), 2,300 credits on Medium or 5,500 on High or Ultra; paid plans for image-based training.

Higgsfield Soul ID: Soul > Train new character, 20 to 80 recent, well-lit photos of one person including a full-length shot; works in Soul models and as an Element in Seedance; not exportable as a file.