Character consistency

Consistent characters stem from a clear set of references. Make sure you cover the angles that the model can't infer and keep each reference focused on its intended purpose. This will ensure that the same character is consistent throughout multiple shots.

The model interpolates your character from the references. The closer your shot matches the existing references, the closer the output will be. However, if you move further away from these, for example by using a new angle, expression, or hand position that wasn't in the frame, the model will have to guess more.

This guess comes from the training data. Basically It's filling the gaps that you didn't provide. The richer the reference set, the less the model has to make up.

The model encodes each reference into a compressed token representation, with a fixed budget per image regardless of source resolution. It then attends across those tokens while generating. Anything not strongly determined by the references gets sampled from the model's learned priors over what such images tend to look like.

A single front-facing shot tells the model nothing about the side or back of the head or the jawline in profile. The model fills in these gaps using training data, resulting in a different face each time.

The four angles that cover most production needs:

• Front
• Three-quarter
• Profile
• Back

The cleanest way to present these is as a single sheet with one image showing all four views side by side, with the same lighting and scale. This allows the model to read it as one character seen from different sides rather than as four separate references that may or may not depict the same person.

If consistency of body shape matters across shots, extend the sheet to a full figure from every angle.

The reference image must contain enough detail for the model to distinguish freckles, wrinkles, birthmarks and skin texture. Low resolution flattens everything. The resulting character will be a smoothed average of the face you provided.

Use high-resolution images without visible artefacts, but don't expect a 24-megapixel reference image to produce better results than a clean 2K one. Modern models downsample every reference to a fixed token budget. A resolution of roughly 2048 pixels, with the face occupying 30–50% of the frame, is more than enough for good results.

Modern reasoning models are better at separating geometry from lighting than older diffusion models. They can extract information from real-world photos that earlier tools could not handle.

A character sheet is an image that you'll reuse across multiple shots in combination with different references, and its job is simply to teach the model who the character is. Flat, even lighting removes one variable from the re-parsing job on every shot.

Once you have a clean sheet, the model can render the character in any lighting you want during production.

In a reference-based panel workflow, the character sheet is rarely used on its own. It is used alongside wardrobe, lighting and pose references. Each reference is designed to contribute something different.

However, models don't always respect these boundaries. Even with our system prompt, which harnesses each reference to its panel, signals may spill over. The cleaner each reference is, the less spillover there is.

The anchor sheet's role in this setup is to establish the character's identity, not their style. For example, if the character is wearing a tailored navy coat in the anchor sheet, that coat may appear in every shot, conflicting with the wardrobe choices selected in the wardrobe panel. Neutral clothing in plain, skin-adjacent tones with no detail keeps the anchor sheet from interfering with the rest of your panel stack.

Deprompt's Character preset produces a sheet with four angles — front, three-quarter, profile and back — with lighting, framing and wardrobe that can be used across models without clashing with the rest of your panel stack. You can build it from parameters or from a single reference image.

Select the desired crop: head and shoulders for face-led work or full figure for maintaining body shape consistency across shots.

The preset incorporates the above practices so you can produce a sheet that works with all modern models.

• Gemini API: Image generation — official model IDs, input limits, reference image handling
• Gemini 3 Pro Image — Vertex AI documentation — token budget, media_resolution, output specs
• Gemini 3 Pro Image Model Card (PDF) — character editing benchmarks, documented limitations
• Nano Banana Pro: Gemini 3 Pro Image model from Google DeepMind — launch announcement
• Nano Banana 2: Combining Pro capabilities with lightning-fast speed — NB2 launch and positioning
• Ultimate prompting guide for Nano Banana — Google Cloud — official prompting recommendations
• Nano Banana Pro is the best AI image generator, with caveats — Max Woolf — independent testing, prior-override behaviour, token economy
• Ultimate Nano Banana Pro Character Consistency Guide — Prompting.systems — six-image sweet spot, reference composition
• Nano Banana Pro Face Consistency: The Complete 2026 Guide — LaoZhang AI — resolution thresholds, face-frame ratio
• Nano Banana Pro Face Consistency Complete Guide — APIYI — re-encoding behaviour across calls