SeaArt Consistent Characters: Same Face, Many Images

Making one good character is easy. Making the same character twenty times, in different scenes and poses, is the problem every illustrator, storyboard artist and brand team hits on SeaArt (seaart.ai). The model has no memory between images, so consistency has to come from what you give it each time: the same words, the same model stack, a reference image where the platform supports one, and, for long-running characters, a LoRA. Here is a method that works from cheapest to most robust.

Level 1: a fixed character block

Write a short, specific description of the character and reuse it word for word. Specific beats pretty. "A woman in her sixties, silver bob haircut, round tortoiseshell glasses, deep green wool coat, small scar on left eyebrow" gives the model five anchors. "An elegant older woman" gives it none.

• Five to eight visual anchors. Hair (colour, length, style), eyes, one facial feature, skin tone, build, signature garment, one accessory. Choose things the model can actually draw at the size you generate.
• Put the block first, after any trigger word, then the scene and action. Keep the block identical across prompts; change only what follows it.
• Name the character with an invented word (for example "Marisol Vey") and use it consistently. A made-up name does little on its own but helps once you move to a LoRA, where it becomes the trigger word.
• Negative prompt the common drift: "different hairstyle, different face, multiple people, text, watermark".

Lock everything else too: same checkpoint and version, same LoRAs and weights, same sampler, same CFG, same size. A character that looks consistent on one checkpoint will look like a cousin on another.

Level 2: fixed-seed testing

The seed decides the random starting noise. With a fixed seed and the same settings, changing only the scene text tells you how much of the face survives the change. Generate a baseline, then produce three variations that change only the action or background. If the face holds across all three, your character block is strong enough. If it drifts, add anchors or raise the weight of the ones that drift most, for example (silver bob haircut:1.2).

Do not ship on a single seed. A character that only exists at one seed is fragile. Once the block passes the fixed-seed test, run it on three or four random seeds and check the face is still recognisably the same person.

Level 3: SeaArt’s character and reference tools

SeaArt offers tools for consistency beyond plain text, including character-focused features and reference-image controls that steer the face or whole figure from an uploaded image. Their exact names and placement change with updates, so look in the generator for options that accept a reference image and mention face, character or identity. The working principle is the same across tools of this kind:

1. Make a clean reference first. A front-facing, well-lit, head-and-shoulders image of your character generated with your character block. No hats, no heavy shadows, no second person.
2. Upload it as the reference, keep your character block in the prompt, and start with the tool’s default strength.
3. Raise strength for fidelity, lower it for freedom. High values copy the face but fight new poses and expressions; low values allow pose changes but drift.
4. Keep lighting close to the reference for your hero shots. Identity holds better when the light on the face is similar.

Only use references you have the right to use, and never a real person without their consent. The practical rules match those in Midjourney consistent characters, even though the controls differ.

Level 4: a character LoRA

For a character you will draw for months, a LoRA is the robust answer. You can search SeaArt’s community library for an existing one if the character is original to a creator who published it, or train your own from 15 to 40 images of your character generated with the methods above. Vary pose, expression, outfit and background while keeping the face identical, caption the variable parts, and use your invented name as the trigger word. Once trained, the LoRA plus the trigger word at a weight around 0.7 to 0.9 reproduces the character on any compatible checkpoint of the same family. Training runs on credits, so check the platform’s current pricing page.

Settings at a glance

Keep a character sheet

Write down the exact character block, checkpoint and version, LoRA names with weights, negative prompt, CFG, size and the seeds of your best outputs, alongside three or four reference images. When a client comes back after six weeks, this sheet is the difference between resuming in minutes and rebuilding the character from memory.

Try it

Model: an illustration SDXL checkpoint, locked
Prompt: Marisol Vey, a woman in her sixties, silver bob haircut, round tortoiseshell glasses, deep green wool coat, small scar on left eyebrow, standing on a railway platform at dusk, holding a paper cup, soft station lights, 50mm lens, storybook illustration
Negative prompt: different face, different hairstyle, multiple people, text, watermark
Size: 832x1216. Steps: 28. CFG: 6.5. Seed: 1234

Keep the seed and change only the scene after "left eyebrow": a kitchen, a rooftop, a bus. If her face survives all three, move to random seeds. If it drifts, add one more anchor or weight the one that slipped.