Reference Images & Character Consistency

8 min read

Your mascot looks different in every image? Learn the reference-image techniques that keep characters and products identical across an entire campaign.

You designed the perfect brand mascot. In image two she has different hair. In image three she is a different person entirely. Welcome to the single biggest gap between hobbyist and professional AI imagery: consistency. Campaigns, storyboards, product catalogs — none of them work if the star keeps shape-shifting. The good news: this is a solved problem, and the toolkit fits in one lesson.

Why Models Forget

A diffusion model has zero memory between generations. Every image starts from fresh noise, and the prompt "a red-haired barista" describes millions of possible red-haired baristas. The model picks a different one each time — not because it is broken, but because that is exactly what sampling from latent space means. Consistency never happens by accident; you have to engineer it.

The Consistency Toolkit

1. Reference image — attach a photo of your character or product to every generation; the model copies identity instead of inventing one
2. Character sheet description — write one exact paragraph (face, hair, outfit, build) and reuse it verbatim in every prompt
3. Fixed seed — reusing a seed with small prompt tweaks keeps composition and identity closer between variations
4. LoRA fine-tuning — for long-running brand characters, a custom-trained model locks identity at the weights level (advanced)

The reference image is your workhorse. Models like NanoBanana and Gemini Flash Image accept one or more reference images alongside the text prompt — upload your character, then describe only what *changes*: the pose, the setting, the action. Identity comes from the pixels, novelty comes from the words. In VAR2's Create Image, that is exactly what the reference slot is for.

Using the reference image, the exact same woman — mid-30s, copper-red bob haircut, round gold glasses, mustard yellow apron — now standing behind an espresso bar in a Tokyo cafe at night, laughing while pouring latte art, neon signage glowing through the window, 4:5 composition

Products Are Stricter Than Faces

Humans forgive a slightly different jawline; nobody forgives a misspelled label. Logos, label text, and exact proportions are where regeneration falls apart. The professional move: never regenerate the product — edit around it. Keep the original product pixels as your reference, and change scenes with image-to-image tools or targeted Quick Edit passes. For full product-scene workflows, the compositing lesson covers cutouts and background plates.

Related glossary terms: Reference Image, Character Consistency, Image-to-Image, Seed, LoRA

Upload a reference image in Create Image and generate the same character in three different scenes. Try Reference Images