The training dataset is where LoRAs are made or broken

For a single subject (a person, a product, a specific character), 15 to 30 images is a sweet spot. Fewer than 15 makes it hard for the model to disentangle the subject from its surroundings, it doesn't have enough examples of the same thing in different contexts to figure out what stays constant. More than 40 starts to introduce more noise than signal, especially if your images are not perfectly consistent. For a style LoRAStyle LoRAA LoRA trained on images that share an aesthetic but vary in subject, so it can render any prompt in that art style, photographic look, or brand identity. (an art style, a photographic look, a particular aesthetic) the right count is usually 30 to 60, because styles are higher-dimensional than subjects and need more examples to characterize.

Variation matters more than quantity. For a person LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models., you want photos from multiple angles (front, three-quarter, profile, even some from behind), multiple lighting conditions (soft daylight, hard sun, indoor, dramatic), multiple backgrounds (so the model learns that the background is incidental), multiple expressions, and a mix of distances (close-up faces, medium shots, full body). For a style LoRA, you want a wide range of subject matter rendered in the same style, so the model learns that the style is what is consistent and the subject is what varies. The single most common mistake in LoRA training is uploading thirty selfies all taken in the same room with the same expression, the model learns to generate that exact pose in that exact room and treats the face as incidental detail.

Train at 1024x1024 for FLUX and SDXL. Train at 512x512 only for legacy SD 1.5 work. The resolutionResolutionThe pixel dimensions of an image or video, for example 1024x1024. Higher resolution shows more detail but costs more memory and time. of your training images sets a hard ceiling on the resolution at which the LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. will produce good outputs, a LoRA trained on 512px images will look soft and lossy when used at 1024px, even if the underlying base model handles 1024px fine. Use lossless formats (PNG, not JPEG) when possible, especially for fine details like hair, fabric texture, and subtle skin tone variations. JPEG compression artifacts are subtle to the human eye but the model will learn them and start producing them as a feature.

Bad source images cannot be rescued by training. Out-of-focus images teach the model to produce blurry outputs. Low-light images with heavy noise teach the model to produce noisy outputs. Heavily edited or filter-laden images teach the model to produce that specific edit/filter as part of the subject. The rule is: if you would not be happy with an output that looked exactly like your training image, do not include that image in your training set.

Each training image is accompanied by a text caption that describes what is in it. During training, the model learns to associate the caption with the image, when you later promptPromptThe text description you provide to a model to specify what you want it to generate. the model with similar text, it produces a similar image. The captions are how you tell the model what to learn from each image. Getting captioningCaptioningWriting the text descriptions that accompany training images. In LoRA training, whatever you caption becomes a controllable variable and whatever you omit becomes baked-in identity. right is harder than it sounds, because the captions interact with the LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. in a counterintuitive way.

The basic principle is that anything you put in the caption becomes a controllable variable, the model learns that this specific feature is associated with this specific word, and it will let you turn that feature on or off via the prompt. Anything you do not put in the caption becomes part of the implicit identity of the subject, the model learns that this feature is always present and bakes it into the LoRA itself. This is the rule that makes captioning subtle and important: you caption the things that should be variable, and you omit the things that should be constant.

Take a concrete example. You are training a character LoRACharacter LoRAA small model trained on images of one specific subject so that a fixed identity can be summoned consistently across an entire production. for a fictional character named Aria, who has long red hair and always wears a green coat. If you caption every image as 'aria, a woman with long red hair wearing a green coat, [background description],' the model learns that 'long red hair' and 'green coat' are independent attributes of women in general, not features of Aria specifically. When you later prompt 'aria standing in a forest,' you might get a generic woman in the forest, because the red hair and green coat were captured by the captions, not by the LoRA. The fix is to caption every image as just 'aria, [background description].' Now the model learns that aria is the implicit name for the entire visual identity, the hair, the coat, the face, all of it, because all of those things are present in every training image and none of them are mentioned in the captions.

Conversely, if Aria sometimes wears the green coat and sometimes wears a leather jacket, you should caption the coat-wearing images as 'aria wearing a green coat' and the jacket-wearing images as 'aria wearing a leather jacket.' Now the model learns that the coat and the jacket are variable attributes, and you can prompt 'aria wearing a red dress' and get a coherent result with the dress instead of either the coat or the jacket. The captioning structure encodes which features are part of the identity and which are part of the styling.

This is also why the choice of trigger wordTrigger wordA unique, deliberately unusual token used in training captions to name what a LoRA teaches, so that including the word in a prompt activates the learned subject or style. matters. A trigger word is the unique tokenTokenA single unit of input, a word fragment for text or a patch for images, that a model processes. you use in captions to refer to the subject, 'aria' in our example, but more often something deliberately weird like 'ohwx' or 'sks' or '0r4n93h31r' to make sure the word does not collide with anything the base model already knows. You want the trigger word to be something the model has not seen during pretraining, so that you are teaching the model a brand new association rather than fighting against an existing one.

Practical captioning workflow

In practice, most people use an automated captioning tool to generate initial captions and then edit them by hand. The two common tools are BLIP (which produces natural-language descriptions) and WD14 tagger (which produces tag-style descriptions like 'long_hair, red_hair, green_coat, outdoor, daylight'). Tag-style captioning works better for FLUX and SDXL, but you have to manually remove the tags that describe features you want to be part of the implicit identity. The auto-captioner does not know which features are supposed to be learned by the LoRA and which are supposed to be variable, that judgment is the artist's job.