Image Prompts

Every Actor, Prop, Set, and Frame in Ozu starts as an image prompt. Good prompts make the difference between a frame that works on the first try and one you reshot four times. This page covers what actually matters when you write them. You rarely write the raw prompt yourself — Ozu composes it from your script and asset details. But the words you choose in beat descriptions, asset notes, and reshot instructions all land in that prompt. The more deliberate you are, the better the frame.

​

The structure that works

Image models pay the most attention to the first few words of a prompt. Lead with what matters most. A good order: Subject → Composition → Lighting → Environment → Mood → Style “A weary detective in a rumpled trench coat leans against a rain-slicked lamppost under a single pool of amber light, noir atmosphere, shot on 35mm film” beats “detective, trench coat, rain, lamppost, noir, cinematic.” Write in flowing natural language — not a tag list.

​

Be specific about the subject

“A woman in a coat” gives the model nothing to work with. “A tall woman in her thirties with short auburn hair and angular features, wearing a rumpled navy peacoat” gives it a character. For Actors especially, specify:

• Age range (early thirties, mid fifties)
• Ethnicity or features (East Asian, high cheekbones, freckled)
• Build (tall and lean, stocky, petite)
• Hair (colour, length, style)
• Clothing (material, cut, colour — “rumpled wool overcoat” not “coat”)
• Expression and pose (slight smirk, arms crossed, shoulders forward)

The same specificity helps for Props and Sets. “A cluttered rolltop desk with scattered papers and an open leather journal” beats “desk with stuff on it.”

​

Composition and camera

Use filmmaking vocabulary — the models are trained on it. Shot size: extreme close-up, close-up, medium shot, wide shot, establishing shot. Be explicit about framing. Angle: eye level, low angle, high angle, Dutch angle, over-the-shoulder. Depth of field: “Shallow depth of field, subject sharp, background bokeh” or “deep depth of field, everything in focus.” Lens references: “Shot on Canon EOS R5 with an 85mm lens,” “wide-angle 24mm,” “anamorphic lens.” Lens choice bundles sharpness, compression, and depth-of-field behaviour into one phrase.

​

Lighting

Lighting shapes mood more than almost anything else. Always specify it — unspecified lighting defaults to flat. Direction: key light from camera-left, backlit, side-lit, top-lit. Named patterns: Rembrandt lighting, butterfly lighting, split lighting. These produce distinct, recognisable results. Natural light: golden hour, blue hour, overcast, harsh midday sun, cold moonlight. Practicals: neon glow, single bare bulb, candlelight, screen glow, fireplace light. Atmosphere: film grain, haze, fog, dust motes in light beams, rain-slicked surfaces. These add depth.

​

Style

A few words of style direction pull the whole image toward the look you want. Embed them in the prompt naturally.

​

Aspect ratio

Pick the ratio before you generate — reshaping afterwards loses quality.

• 16:9 — widescreen, most cinematic scenes
• 4:3 — classic film, TV, documentary feel
• 9:16 — vertical, mobile-first, hero portraits
• 1:1 — square, social media

Reinforce the ratio inside the prompt text too (“wide 16:9 cinematic composition”) — some models drift if you only set the parameter.

​

What to avoid

• Quality keyword stacking. “Masterpiece, best quality, 8k, ultra-detailed, award-winning” does nothing. Modern models default to high quality. Every word should carry visual information.
• Meta-language. “A beautiful image of,” “an amazing photograph of,” “stunning.” Zero visual content. Cut it.
• Contradictions. “Bright well-lit room with deep shadows” confuses the model. Pick one.
• Text in images. Most image models render text as gibberish. If you need a sign or label, generate the image without it and composite text later.
• Tight close-ups on hands. Hands are the classic AI failure mode. Frame wider, or pose hands in pockets, behind the back, or holding a simple object.
• Over-complex compositions. Four-plus characters in one frame drift fast. Keep it to two or three, and use spatial blocking: “A on the left facing right; B on the right facing left.”

​

Keeping characters consistent

If you reshot an Actor in scene three and it doesn’t match the one from scene one, the problem is usually wording drift. Lock the description. “Grey wool overcoat” must not become “dark coat” four frames later. Ozu stores your Actor’s hero look and wardrobe variants and feeds them as references into downstream frames. Keep hero looks clean and well-lit — they set the baseline for every frame that uses that Actor.

​

Related

​

FAQ

Do I need to write prompts by hand? No. Ozu composes them from your script analysis, Asset data, and Scene/Shot/Beat notes. The words you write in those places shape the final prompt. Will the same prompt give me the same image twice? No — image generation is stochastic. Two generations from identical prompts will be visually similar but not pixel-identical. Reshot until you get a version you like. Why does my character look different in each frame? Usually drifting wording. Ozu’s Actor references help, but if your descriptions vary (“grey coat” → “dark overcoat”) the model will drift with them. Lock terms exactly. What aspect ratio should I use? Match the delivery format. 16:9 for most films and streaming. 9:16 for mobile-first. 1:1 for social posts. Set it per project at onboarding. Can I use my own photos as references? Yes. Drop any image on the Flow canvas to create an Imported Asset that slots into the pipeline. It can anchor a Set or feed into an Actor. Why does text in my frame look like gibberish? Most image models can’t render readable text. If a frame needs a sign, a label, or a title card, export the frame clean and composite text in your video editor.