Reference inputs: photos, scenes and prompts

Three references, three jobs

A generated clip needs answers to three questions: who is in it, what happens, and any detail the template could not guess. The photos answer the first question, the scene preset answers the second, and the optional prompt answers the third.

Keeping the jobs separate is the single most useful mental model for better results. When a clip disappoints, the fix is almost always in exactly one of the three — a likeness problem belongs to the photos, a motion problem belongs to the scene, and a detail problem belongs to the prompt.

Diagnose before you regenerate: ask which of the three references let you down, and change only that one.

Photos carry the likeness

The model reads faces, bodies and clothing from your photos and keeps them consistent through the clip. Everything identity-related — who appears, how old they look, what they wear — comes from the photos and cannot be corrected later by a prompt.

The rules that matter most: faces large in frame and looking at the camera, torsos visible, even front-facing light, and no sunglasses or heavy filters. For two-photo templates the two photos should roughly match in light and camera distance, because the generator has to place both people in one shared scene.

Sharpness beats resolution. A crisp 2-megapixel portrait produces a better likeness than a soft 12-megapixel group shot, and a photo where the face fills a quarter of the frame beats a distant full-body where it fills a twentieth.

Scene presets carry the motion

A scene preset is a pre-tuned description of the moment: the location, the light, the camera distance and the emotional register of the movement — a reunion hug rushes, a lakeside hug settles. Picking a scene is how you direct the action now, in the same way a reference clip used to be.

Scenes beat freeform prompting for motion because they are tested. Each one ships with motion and lighting that the model renders reliably; a hand-written action description is a guess by comparison. Browse the scene library, watch the sample clips where they exist, and pick by the story you are telling rather than the prettiest thumbnail.

If the motion in your result feels wrong — too fast, too still, too wide — change the scene, not the prompt. The scene is where movement lives.

The prompt carries the details

The optional written direction exists for what the scene cannot know: a specific place, an outfit, the reason for the moment. It is appended to the scene you picked, so the two should agree rather than compete.

• Describe facts the scene cannot know: a city, a season, a piece of clothing, the occasion.
• Keep it under 200 characters. One or two concrete details beat a paragraph.
• Do not re-describe the action — the scene already handles it, and conflicting directions blur the motion.
• Do not ask for camera moves or editing effects. Framing and pacing are tuned into the scene.
• If a detail matters, name it plainly. Poetic phrasing gives the model nothing to render.

A good test: if you deleted your prompt, would the clip still make sense? If not, the prompt is doing a scene’s job — move that information into the scene choice instead.

How the three combine — and how they conflict

The generator treats the scene as the backbone and the photos as the cast. Your prompt is a small addition on top, so it can refine but not override. A prompt that contradicts the scene — an indoor scene with an on-the-beach prompt — produces a compromise that serves neither: the model averages the two settings and the result feels nowhere.

The photos can conflict with the scene too. A sharp flash-lit indoor portrait dropped into a golden-hour street scene will keep its harsh light longer than you expect, because likeness is preserved before atmosphere is applied. When a result looks cut-and-pasted, soften the photo’s light before blaming the scene.

• Scene and prompt disagree: the scene wins — pick a scene that already points the right way.
• Photo and scene disagree: the photo wins on the person, the scene wins on the world around them.
• Two photos disagree with each other: the result splits the difference, which is why matching light between photos matters more than either photo’s absolute quality.

If you filmed reference clips before

The clips are not wasted. Pause a clip at the moment with the best posture and expression — both people facing each other, arms roughly where a hug would land — and save that frame as an image. A clean still from a video is often a better input than a rushed photo, because you chose it from hundreds of frames.

Grab a frame per person if the clip shows them separately, or one shared frame if they appear together — a shared frame works as the single-photo input. Crop so faces are large and backgrounds are plain, then treat the still exactly like a photo: same light rules, same readiness check, same scene choice.

A phone screen recording of a video call works too — pause where both faces are clear and screenshot. It is a still photo by the time the generator sees it.

A five-point input checklist

If a result misses the mark, run the checklist top to bottom before re-rendering. Most regenerations fail for the same reason the first one did, because only the prompt was changed.

• Faces large, front-on and sharp in every photo you upload.
• Two photos match in light and camera distance, or one photo shows both faces clearly.
• A scene chosen for the story, not the thumbnail — sample clips watched where they exist.
• A prompt only for details the scene cannot know, under 200 characters, never contradicting the scene.
• One variable changed per regeneration, so you know what actually fixed it.