Keep a character consistent across shots
How to keep a character, vehicle, or product looking the same across every shot in your sequence. The single highest-leverage technique for pro-quality output.
The recurring failure mode in pure prompt tools: the hero looks different in shot 2 than in shot 1. A different jacket. A different face. The Lambo is now red. Re-prompting won't fix it because the model is being asked to invent the look every time, and small variations compound.
The fix has one moving part. Bind a reference image to the 3D object once, then reuse the same object across every shot. The visualizer locks the appearance to the reference for every render that includes the object.
Steps
Place the object in the scene. Use the Asset Library for stock geometry or import your own model for a client product. The object lives in the scene; shots are framings of it.
Select the object and open Object Details. Click the object in the viewport. The Details panel docks on the left. Scroll to the Reference images section.
Attach a reference image. Click Upload or Add from Media Library. JPEG and PNG accepted. Up to four images per object for multi-view references.
Reuse the same object across shots. Don't duplicate the object per shot. Move the same instance, or let the camera reframe it. The reference binding follows the object, so every shot that includes it inherits the locked appearance.
Render. In Visualize mode, generate. The reference travels into the auto-prompt's Subjects block and into the model alongside the prompt. The character, vehicle, or product comes back as the reference, not as a generic interpretation of one.
Faster path: the AI Composer
Instead of opening the Details panel, type into the AI Composer prompt input at the bottom of Build mode:
Attach this as a reference image to the chair: <image URL>
The composer finds the matching object and attaches the reference. The Build viewport shows a base color at most, not the reference itself, which resolves at render time. Faster when you can describe the object in a sentence. See AI Composer.
Why this works
The model receives a structured 3D scene plus a prompt, not just text. The reference image attaches to a specific object in the scene, so the model knows exactly which thing in the rendered output should match the reference. Pure prompt tools can't do this because they have no concept of "this object" – they only have the description.
For the deeper version, see Image reference and How the visualizer thinks.
Tips that level up the result
Vanilla object names beat descriptive ones for character consistency. "Mark" and "Jeff" hold their look across renders better than "pirate captain" and "British sailor". Descriptive names pull the model toward generic interpretation and away from the specific reference.
Multi-view references improve mesh-aware results. Front, three-quarter, and side photographs produce more consistent renders than a single angle.
Reference-tuned image models honor references most tightly. See Models for which models in the current working set are tuned for reference-image conditioning. Older video models honor references less tightly – if a reference seems ignored on video, try first-and-last-frame interpolation from a reference-locked still.
Limits
Reference fidelity varies by model. If a reference seems ignored, switch to a reference-tuned image model (see Models), then use first-and-last-frame interpolation for video.
Reference doesn't override composition. A reference image tells the model what the object looks like, not where it goes. Composition is a Compose-mode concern.
One reference per object at a time. You can attach up to four images for multi-view, but they're treated as multiple angles of the same object, not different alternates.
References are downsampled and composited internally. Every reference attached to an object lands on a single 720 × 1024 sheet before the model sees it. Multi-view (front, three-quarter, side) works well at this resolution; oversized source files give the model nothing extra. Sub-megapixel inputs are fine.
Related
LoRAs – the alternative when a fine-tuned model is required.
Last updated
Was this helpful?