A step-by-step workflow for building an AI character sheet from a single reference photo. Get the exact prompts for front, three-quarter, side and back views, an expression row and a props panel, then fix drift panel by panel with Nano Banana Pro and Image Editor Pro.
You have one good picture of a character: a face you like, an outfit that works, a mood worth keeping. Then the second image comes out with different eyes, a new jacket and a hairline that moved two centimeters. That drift is the reason an AI character sheet generator workflow exists. It pins the character down before you build scenes around it. This article shows how to go from a single reference photo to a full sheet, with the exact prompts for front, three-quarter, side and back views, an expression row and a props panel. Everything below runs on Nano Banana Pro and PicassoIA Image Editor Pro, with one vision model to write the descriptions for you.
A character sheet is a reference page. Animators, game artists and comic illustrators have used them for decades so that ten people can draw the same person without arguing about the nose. In a photographic workflow the idea is identical: one subject, shot or generated in a controlled set of views, with the details written down beside it.
You need one whenever the same person has to appear more than once:
A useful sheet is not a pile of pretty portraits. It has a fixed list of panels, and each panel answers a different question about the character.
| Panel | What it records | Why it matters |
|---|---|---|
| Front view | Face, hair, outfit layout | The anchor for every later image |
| Three-quarter view | Nose, jaw, cheek depth | Most scenes show this angle |
| Side profile | Nose bridge, ear, hair volume | Exposes silhouette errors |
| Back view | Hair length, collar, bag straps | Needed for walking-away shots |
| Expression row | Smile, anger, surprise | Keeps the face stable when emotion changes |
| Props and palette | Objects, fabric, colors | Stops items from changing between shots |
A single image only shows what the model decided to draw. Anything not visible stays open for reinterpretation, so the back of the jacket, the shape of the ear and the scar on the eyebrow get invented fresh every time. A sheet closes those gaps. When you attach it later as a reference, the generator has a fact to follow instead of a guess to make.
The cost of skipping this step grows with the size of the project. Two images with a slightly different nose pass as a style choice. Forty images with a slightly different nose read as forty different people, and fixing them one by one takes far longer than building the sheet would have.
💡 Tip: Treat the sheet as a contract. Anything you want to survive across fifty images must appear on it at least once.
The trick behind this whole workflow is reading the image first and writing the prompt second. Do not describe the character from memory. Let a vision model describe the photo, then edit that text into a block you can reuse.
Open Gemini 3.5 Flash, attach the photo (it accepts up to 10 images per prompt) and ask for a plain inventory. A request like this works well:
Describe the person in this image for a character reference sheet.
List on separate lines: age range, face shape, eye color and shape,
eyebrows, nose, lips, skin tone and marks, hair color, length and texture,
clothing item by item with colors and materials, accessories.
No interpretation. No mood words. Only what is visible.
The "no mood words" line matters. Words like "mysterious" or "confident" push the generator toward a style instead of a face, and the style is the first thing to drift. If you have several photos of the same person, attach them all and ask the model to list only the traits that appear in every one. Those shared traits are the safest ones to freeze.
Take the answer and cut it down to the traits that must never change. Aim for 40 to 60 words. This is your character block:
CHARACTER: woman in her late twenties, oval face, light freckles across nose and cheeks,
copper-red wavy hair to the shoulders, green eyes, straight eyebrows, mustard waxed-cotton
field jacket with brass buttons, cream knit sweater, brown leather strap across one shoulder
Paste this block word for word at the start of every prompt. Changing one adjective between runs, say "copper" to "ginger", is enough to shift the face. Small marks such as freckles and scars deserve their own words, because they are exactly what a model smooths away first. Say which side of the face a mark sits on, and say it from the viewer's point of view, so the left eyebrow scar does not flip in a mirrored panel.
Each panel gets one template. The character block stays fixed and only the view line changes. Camera words matter here: name the lens, the height of the camera and the light, because those are the three things that change a face the most without touching the character itself.
[CHARACTER BLOCK]. Full-body photograph, standing straight, arms relaxed,
facing the camera, neutral expression, plain mid-grey seamless backdrop,
soft even studio light, 85mm lens, eye level, raw photo, natural skin texture.
Choose a neutral expression and flat light on purpose. Dramatic lighting hides the details the sheet exists to record, and a grey backdrop keeps the outfit colors honest. Generate four or five versions and pick the one that matches your reference photo best. That winner becomes the anchor for the rest of the sheet.
Attach the approved front view as a reference image and change only the view line. Start with the three-quarter angle, because it is the one you will use most:
Same person as the reference image. [CHARACTER BLOCK]. Three-quarter view
turned 45 degrees to the left, full body, same outfit, same grey backdrop,
same lighting.
For the side profile, swap the view line for "left side profile, looking straight ahead". For the back view, write "seen from directly behind, hair and jacket back visible". Back views are where drift shows up first, because the front reference says nothing about the rear of the jacket. Add one line describing it, for example "plain back, no logo, stitched seam across the shoulders", so the model has nothing to invent.
Generate six expressions as separate images: neutral, smile, laugh, anger, surprise and worry. Keep the same backdrop and crop, head and shoulders only.
Same person as the reference image. Head and shoulders, [expression] expression,
mouth and eyebrows clearly showing the emotion, same grey backdrop, same light.
Six small portraits side by side let you spot a slipped face in one second. If the freckles vanish on the angry one, you see it right there on the grid. Name the expression with a physical description when the plain word fails: "eyebrows pulled together, jaw tight, lips pressed" gives you anger without handing the model a mood to reinterpret.
Nano Banana Pro fits this job because it accepts up to 14 reference images in one prompt and renders up to 4K. The original photo, the front view and the side view can all travel with every new request.
A filled-in prompt for the side profile looks like this:
Use the first image for the face and hair, the second image for the outfit and pose.
CHARACTER: woman in her late twenties, oval face, light freckles, copper-red wavy hair
to the shoulders, green eyes, mustard waxed-cotton field jacket, cream knit sweater.
Left side profile, full body, plain mid-grey seamless backdrop, soft even studio light,
85mm lens, eye level, raw photo.
| Setting | Recommended value | Reason |
|---|---|---|
| prompt | Character block + view line | Same words every run |
| image_input | Up to 14 images | Original photo plus approved panels |
| aspect_ratio | 16:9 for full-body rows, 1:1 for expression crops | Default is match_input_image |
| resolution | 2K while drafting, 4K for final panels | Saves time during iteration |
| output_format | jpg for drafts, png for final files | PNG avoids extra compression |
| safety_filter_level | Leave at the default | Portraits rarely trigger it |
💡 Tip: Change one thing per run. If you alter the view, the lighting and the outfit at once, you cannot tell which change broke the face.
Even a good run produces one panel that is wrong. Do not regenerate the whole set. Edit that single panel.
A simple rule decides which tool to reach for. If the face itself is wrong, go back to the generator and run that view again with the approved front view attached. If the face is right but one detail is off, such as a missing button, a different collar or a wrong sleeve color, edit the panel instead. Editing keeps everything you already approved and changes only what you name.
PicassoIA Image Editor Pro takes up to three reference images and lets you address each one by number inside the prompt. It is unlimited, with no per-edit cap, so you can iterate on one panel as many times as you need. A typical correction looks like this:
Keep the face, hair and pose from image 1. Replace the jacket with the
mustard waxed-cotton jacket from image 2, same brass buttons, same fit.
Keep the background and lighting unchanged.
Set aspect_ratio to match_input_image so the panel keeps its size, lock the seed when you find a result you like, and set num_outputs to 2 when you want to compare two versions of the fix in one run. Export as PNG for panels that go on the final sheet.
Run through this short checklist after every batch:
Name each approved panel with the character and the view, for example mara_front.png and mara_back.png, and keep them in one folder. When a new project starts, that folder is the only thing you attach.
A sheet that shows only the body leaves the objects to chance. Add a props panel: a flat-lay photograph of every item, prompted on a neutral surface with the materials spelled out.
Overhead flat-lay on a weathered oak table: folded mustard waxed-cotton jacket,
worn brown leather satchel with brass buckles, olive wool scarf, cream linen shirt,
generous spacing, soft window light, 35mm lens, raw photo.
Then write a palette of three to five colors using plain names such as mustard, olive and cream. Image models usually respond more reliably to color words than to hex codes. If you want clean cutouts of the props for the final sheet, Bria Remove Background does it in one pass, and Clarity Pro Upscaler sharpens the panels you plan to print.
Give each prop a single defining detail, such as a scuffed brass compass or a satchel with one broken buckle. A character with one worn object feels lived in, and that object is easy for a model to repeat.
A finished sheet also feeds video. Seedance 2.0 accepts up to nine reference images for character consistency, so the front, side and back panels can go in together. Prompt a slow quarter turn on a grey backdrop and the clip keeps the same face and outfit across the movement. It renders up to 720p with synchronized audio, which is plenty for a turnaround reel or a pitch deck.
Paraphrasing feels harmless and it is the most common cause of drift. Keep the character block in a notes file and paste it. Never retype it.
A face lit from the left looks like a different face lit from above. Keep the studio setup identical across the view panels and save dramatic light for the scenes you build later.
Most people stop after the front and the profile. Then the first walking-away shot shows a coat with a hood the character never had. Generate the back, even if you think you will never use it.
Pick one photo you already like and run the three steps: describe it with Gemini 3.5 Flash, freeze the character block, then generate the front, side and back views in Nano Banana Pro. Fix any wrong panel in PicassoIA Image Editor Pro, and your character will look the same in image fifty as it did in image one.
Open Picasso IA, upload your reference photo and try the front view prompt first. Do not aim for perfect on the first pass. Aim for six panels that agree with each other: front, three-quarter, side, back, one expression row and one props flat-lay. Once those match, you can lay them out on a single board, share it with a teammate or a client, and ask them to circle anything that looks off before a single scene is built. One hour of work on the sheet saves days of regenerating faces later. Browse every available model at picassoia.com/en/all-models and start with the sheet.
