Keeping one male character consistent across images
You generate a man you like. You ask for him again in a different room and get his cousin: similar hair, different nose, ten years younger. This is the most common frustration with image generators, and it is not a fault in your prompt. A diffusion model has no memory of previous images. Every generation starts from fresh noise, and “a man with a short beard” matches millions of possible faces. Consistency has to be added on purpose. Here are the methods, from simplest to most involved.
Level 1: a fixed character description
Write one paragraph that defines him and reuse it word for word at the start of every prompt. Only the scene changes.
The description needs to be more specific than feels natural. “Handsome man with dark hair” pins down nothing. Something like this pins down a lot:
Tomás, a 38-year-old man, olive skin, square jaw, deep-set dark eyes, thick eyebrows, short black hair greying at the temples, trimmed full beard, small scar through the left eyebrow, broad shoulders, stocky build, hairy forearms.
Distinctive marks (a scar, a gap in the teeth, a particular tattoo) help more than general attractiveness words, because they give the model something unusual to hold on to. Giving him an invented name can also help a little on some models, since the name ties the description together. Use a made-up name, never the name of a real or famous person.
This level costs nothing and works on any tool with a text prompt. Expect a family resemblance, not an identical face.
Level 2: fix the seed and the settings
The seed sets the starting noise. Where a tool lets you see and fix it, keeping the same seed, model and settings while changing only a few words produces images that stay close to each other. It is ideal for small variations: a different shirt, a different expression.
The limit is that large changes to the prompt (a new pose, a new camera angle) alter the image so much that the seed no longer holds the face. Seeds keep a picture stable; they do not keep a person stable.
Level 3: reference images
Many tools let you supply an image as a guide. The feature has different names, such as image-to-image, character reference or face reference, and different strengths. The model takes features from the reference and carries them into the new picture.
The workflow that tends to work:
- Generate until you have one clean, front-facing portrait of your character with neutral expression and even light.
- Save it as his reference sheet.
- Use it as the reference for every later image, together with your fixed description.
Only ever use a reference image that you generated yourself of a fictional character. Uploading a photo of a real man, whether a stranger, someone you know or a celebrity, to place him in sexual images is making a deepfake. It is harmful, it breaks the rules of any reputable service, and in a growing number of places it is a crime. See deepfakes and the law.
Level 4: a saved character in a builder
Character builders and companion apps do the bookkeeping for you. You design the man once through menus, and the service stores whatever it needs (a description, a seed, a reference, sometimes a small trained model) to bring him back. This is the least effort for the user and often gives the steadiest results inside that one service.
The cost is lock-in. The character lives on their servers in their format. If you cancel or the service closes, he usually does not come with you. Keep copies of your favourite images and of the text description so you can rebuild him elsewhere.
Level 5: training a character LoRA
The strongest method is to train a small add-on model, a LoRA, on a set of images of your character. Once trained, a trigger word summons him in any pose, outfit or style with high reliability.
It is also the most work. You need a set of varied, consistent images of the character to begin with (usually produced with levels 1 to 3 and carefully selected), plus either a local setup with a capable graphics card or a service that offers training. Captions, training settings and trial runs all take time.
The same rule applies with more force: train only on images of a fictional character that you generated. Training a model on photos of a real person without consent in order to make sexual images of them is exactly what deepfake laws are aimed at.
What breaks consistency
- Changing the art style. A character defined in photoreal will not survive a jump to anime without a reference or LoRA.
- Extreme angles. Profiles, views from behind and from above lose the face.
- Small faces. In a full-body shot the face occupies few pixels and the model improvises. Generate closer, or use a face-detail or upscale feature if the service has one.
- Conflicting words. Adding “younger” or “slimmer” to one scene changes the whole man.
- A model update. Hosted services change their models, and your character can shift with them.
A realistic standard
Even professional workflows produce drift. Aim for “clearly the same man” instead of pixel-identical, generate a few candidates per scene and keep the one that matches best. If the body type keeps slipping, repeat the build words in the scene part of the prompt as well; the wording in prompts for male characters helps here. Moving images raise the difficulty again, as described in AI video of male characters.