Fix a weak AI image or video prompt, symptom by symptom

Symptom: the text on the image is misspelled

Unquoted words get read as description, so the model writes its own. Quote the exact words, give them a size and a place, and run it on Nano Banana Pro, the model we tag for lettering.

1. Poster headline in quotes

   A charity 10K poster with the headline "RUN THE RIVER" in huge white capitals across the top third, "Sunday 6 April · 8am start" in small type at the bottom, runners crossing a stone bridge between them.

   Weak version: "river run poster with the date". Nothing was quoted.

2. Painted florist sign

   A hand-painted wooden sign reading "Salt Street Florist" in cream brush script on dark green, above a shop doorway on a bright morning, buckets of tulips below it, the sign sharp.

3. Price board, one item per line

   A barbershop felt letter board with exactly three rows: "CUT 25", "BEARD 15", "CUT + BEARD 35". White plastic letters, even spacing, shot straight on.

4. Candle label

   A matte black candle jar with a white label reading "CEDAR & SMOKE" in spaced sans-serif capitals and "8 oz" below it, light from the left, the label facing the camera.

5. Timetable with three headings

   A flat weekly timetable titled "Studio Classes" with three rows headed "Mon · Yoga 7pm", "Wed · Pilates 6pm" and "Fri · Spin 7am". Navy on off-white, one small icon per row, wide margins.

Symptom: the product or person drifts from your reference

Give what must not change its own sentence, and attach more than one angle. Both Nano Banana models take up to 14 references in upload order, with no @ tags. If none of them load, the image renders from text alone and still bills.

1. A separate preserve sentence

   Keep the sneaker from the attached photo exactly as it is: same shape, colours, stitching and logo placement. Place it on wet city asphalt at night, neon reflections in the puddles, low angle.

   Weak version: "my sneaker on a street at night". Nothing said what must stay.

2. Name each image by its order

   Use the first image for the jacket and the second image for the woman. Show her wearing that jacket on a windy seaside promenade in overcast light, the jacket's colour and zips unchanged.

3. Sketch as layout, photo as object

   Using the attached sketch as the layout and the attached product photo as the object, render a finished shop shelf with the bottle on the middle shelf, warm retail lighting.

4. Several angles of one product

   Four attached photos show the same backpack: front, left side, back and a logo close-up. Show it at a three-quarter front angle on a mossy rock by a stream, every strap and pocket matching the photos.

5. Change the background only

   Change only the background of the attached photo to a sunny park lawn. Keep the dog's markings, collar, pose and the light on its fur exactly as they are.

Symptom: the video model ignores your reference images

Each video model we run has its own reference tag, so a prompt pasted from another model points at the wrong image. Seedance 2.0 uses @Image1, HappyHorse 1.1 [Image 1], Wan 2.7 Image 1 and Gemini Omni Flash <IMAGE_REF_0>, counted from zero. Name every reference and say what it shows.

1. Seedance 2.0: @Image tags

   Seedance 2.0
   The paper kite from @Image1 lifts off the dune in @Image2 and climbs into a windy sky, the camera tilting up to follow it, wind and flapping paper on the soundtrack.

   Seedance rejects input images with faces, even AI-generated ones.

2. HappyHorse 1.1: bracketed tags

   HappyHorse 1.1
   The woman in the red raincoat from [Image 1] walks through the night market from [Image 2], turns to the camera and says, "Best dumplings in town." Handheld tracking shot, lantern light.

3. Wan 2.7: tags without brackets

   Wan 2.7
   The chef in the white apron from Image 1 plates a dessert on the marble counter from Image 2, the camera pushing in slowly as he adds a mint leaf.

4. Gemini Omni Flash: zero-based tags

   Gemini Omni Flash
   The cyclist in the yellow jacket from <IMAGE_REF_0> rides the coastal road from <IMAGE_REF_1> at sunrise, then stops and looks at the camera, saying: "Twenty kilometres before breakfast."

   Omni Flash speaks quoted dialogue itself and picks the voice.

5. Nano Banana: refer by order, not by tag

   Nano Banana Pro
   Paint the mural from the first image across the brick wall in the second image, following the bricks' texture and the wall's perspective, in the same afternoon light.

Symptom: the request fails, or a "no" gets ignored

HappyHorse 1.1 and Gemini Omni Flash reject a separate negative prompt, and "no cars" still puts cars in the prompt. Google's Nano Banana guide gives the fix: describe what you want, such as an empty street.

1. An empty street, described

   HappyHorse 1.1
   An empty cobblestone street at dawn, shutters closed, one street lamp still glowing, mist low over the stones, pale pink sky above the rooftops, a slow push-in.

   Weak version: "street at dawn, no cars, no people, no text".

2. One object on a bare counter

   Gemini Omni Flash
   A bare white kitchen counter with only a glass of orange juice on it, plain walls, soft window light. A woman lifts the glass and smiles, saying: "Morning."

3. A blank T-shirt

   Nano Banana Pro
   A plain, unbranded white T-shirt laid flat on light oak, smooth fabric, a blank chest, soft overhead daylight.

   Weak version: "white t-shirt, no logo, no wrinkles, no text".

4. A still lake

   HappyHorse 1.1
   A calm lake at noon with a mirror-still surface under a clear sky, one wooden rowing boat moored at the jetty, pine trees reflected on the water, a static wide shot.

5. Ask for the sound you want

   Veo 3.1
   Close-up of hands kneading dough on a floured board, a locked-off camera, soft side light. The only sounds are the dough on the wood and a ticking kitchen clock.

   Veo 3.1 always generates audio, so name the sounds you want instead of writing "no music".

Symptom: the light is flat and the image looks generic

Phrases like "good lighting" and "high quality" give a model nothing to act on. Name where the light comes from, how hard it is, its colour and the lens.

Build the lighting phrase with our cinematic lighting prompt generator

1. Raking low sun

   Nano Banana Pro
   Portrait of an old fisherman mending a net on a harbour wall, hard low sun from the left raking across his face, deep shadow on the right, 85mm lens look at f/2, warm afternoon colour.

2. One warm practical lamp

   Nano Banana 2 512 draft
   A bowl of ramen on a dark wooden counter, one warm lamp above and behind it, the steam backlit, the rest of the bar in shadow, 35mm lens look.

   Draft the light on Nano Banana 2 at 512 for 20 credits, then rerun on Pro.

3. Two colour temperatures

   Nano Banana Pro
   A modern living room at blue hour, cool daylight from a big window on the right, a warm floor lamp on the left, the two colours meeting across the sofa, 24mm lens look.

4. Soft overhead flat lay

   Nano Banana Pro
   Overhead flat lay of a leather notebook, fountain pen and brass compass on navy felt, soft shadowless light from a large diffuser directly above.

5. Spotlight and haze in video

   Veo 3.1
   Medium shot of a guitarist on a small club stage, one magenta spotlight from above, haze, warm amber backlight, the camera slowly circling him. His guitar over a murmuring crowd.

Symptom: the camera or the sound in the clip is wrong

Veo 3.1 always generates sound, so a prompt that leaves out the camera and the audio leaves both to the model. Name one move with a speed and an end frame, and describe audio in separate sentences, as Google's Veo prompt guide recommends.

Pick the move with our camera movement prompt generator

1. Dolly-in with an end frame

   A slow dolly-in from a wide shot of a lighthouse on a cliff to a close-up of its lamp, stormy dusk. The only sounds are waves breaking below and gusts of wind.

2. Locked-off tripod

   A locked-off tripod shot at eye level of a potter shaping a bowl on a wheel; the camera stays perfectly still. The hum of the wheel and wet clay under her hands.

3. Rising drone reveal

   A drone shot rising straight up from a red tent in a pine forest to reveal a mountain lake at sunrise. Birdsong and a light breeze.

4. Handheld follow with one line

   A handheld shot following a courier cycling through a busy crossing at a steady distance. Traffic, a bike bell, a bus braking. The courier calls out: "Two minutes away."

5. Pan that ends on a speaker

   A slow pan across a bakery counter of fresh loaves, ending on the baker, who looks up and says: "First batch is out." Warm morning light, the crackle of cooling crusts.