Attach the same clear photo of the person to every image and change only the scene, outfit, or pose in the prompt — on ChatImg that means uploading the photo as a reference (Nano Banana 2.1 takes up to 14) instead of describing the face in words. A text-only prompt invents a new face on every run, however detailed it is.
Every consistent-character workflow — fine-tuning, adapters, or reference-based editing — works by feeding the model images of the person. For a handful of pictures you do not need training: attach the same photo to each request on ChatImg, say the face must not change, and vary one thing per image.
A text-to-image model has no memory of the last picture. Words like "same woman, brown hair, freckles" match thousands of faces, so each run picks a different one. A reference photo gives the model the one face to follow.
Reuse one sharp, front-facing photo for the whole set. Swapping references mid-set is the most common cause of drift.
"Same person as the reference photo. Do not change the face." Then describe only what is new.
Scene, outfit, or pose — not all three at once. Big jumps give the model room to reinterpret the face.
Copy it, keep line 1, change line 2.
Upload the same clear, front-facing photo for every picture in the set.
"Same person as the reference photo. Keep the face, hairline, and skin tone unchanged."
"Now standing in a rainy Tokyo street at night, wearing a navy raincoat."
Put the result next to the reference. If it drifted, send it back: "Match the face to the reference exactly."
Sets where a new face on picture three ruins the whole thing.
One person across a week of posts, in different places and outfits.
Generate one portrait you like, then reuse it as the reference for every panel.
Several headshots of the same person for LinkedIn, a team page, and a speaker bio.
The same model holding different products, so the catalog looks like one shoot.
A consistent face means the same recognizable person appears across several generated images while the scene, clothing, or pose changes. It is achieved by conditioning the model on reference images of that person — by fine-tuning (DreamBooth), by an image-prompt adapter (IP-Adapter, InstantID), or by an editing model that accepts reference photos — rather than by text alone.
Because without a reference image the model regenerates the whole picture, face included, from words that describe many possible people. Attaching the reference and stating that the face must not change turns the request into an edit of a known identity instead of a fresh draw.
Not for a small set. Training (for example DreamBooth with a few photos) pays off for hundreds of images in one style. For a few pictures, a model that takes reference photos at request time — such as Nano Banana 2.1 on ChatImg, up to 14 references — gives the same face with no training step.
DreamBooth (Ruiz et al., 2022) showed that a text-to-image model fine-tuned on 3–5 images of a subject can place that same subject in new scenes — identity comes from the images, not the text.
DreamBooth — arXiv:2208.12242 ↗IP-Adapter (Ye et al., 2023) adds an image prompt to diffusion models so a reference image steers the output alongside the text prompt, without fine-tuning the base model.
IP-Adapter — arXiv:2308.06721 ↗InstantID (Wang et al., 2024) preserves a person's identity from a single reference face image at generation time, with no per-person training.
InstantID — arXiv:2401.07519 ↗FAQ'S
Ask us anything!
Upload one photo of the person, then change only the scene.
