Seven image models, seven jobs: the wrong model beats a perfect prompt

These models aren't interchangeable. Each was built for different work. Picking the right model matters as much as picking the right words—so the routing table sits on page one of chapter one, ahead of every writing tip.

Which Nano Banana to pick

Nano Banana 2 landed in February 2026 and can own most day-to-day work: quality close to Pro, much faster, solid text, and the best consistency across multiple references. Default pick for edits, product, and brand. Nano Banana Pro is the November 2025 generation, positioned as the premium option—use it when you need precision, brand-level fidelity, or to polish the hardest prompt. Slower, higher ceiling.

The others have their own personalities. Flux.2 was built for control: ten references at once, exact hex colors, structured storyboard lists—great for brand and product pipelines locked end to end. Hand it a vague prompt and you get a polished, mediocre frame. Krea 2 lives at the other pole: built for style, working from style refs and mood boards to dodge the default AI retouch look. The tradeoff is it interprets more than it obeys, and on-image text is weak. Recraft earns its row because it handles layout-heavy posters and art-school illustration—and it's the only one that exports real vector files.

Recraft V4.1 specialty showcase: layered display type, trained brand styles, and the only real path to true vector. From handbook p. 11.

Six things every image prompt should cover: subject, action, scene, light, framing, style

Today's image models read sentences, not keyword piles. Describe a coherent scene and they build it. Reasoning models like Nano Banana Pro and Nano Banana 2, GPT Image 2, and Seedream parse a dense prompt in chunks before they render, so busy frames with many subjects and details still hold up.

The structure is six parts. The diagram below breaks a real cologne-ad prompt so you can see which part of the frame each segment owns:

Join the six parts and you have the full prompt. The result is the image below. Bottle on the right, a crown splash beside it, hard side light across wet slate—everything in the frame was named in the prompt. Nothing was left for the model to invent:

Nano Banana 2 output. The client wanted "energy on the shelf," not a pose, so the whole prompt only set three things: the instant, the light, the frame. From handbook p. 6.

Same structure, different job. Left: treat the prompt like a layout brief—what the copy says, where each element sits, where to leave air. Right: chase a cinematic feel with a few "looks like" words.

Layout · GPT Image 2. Title "MORNINGS, UPGRADED" in heavy ink-black sans, two lines filling the top half; subhead bottom left; stainless tumbler bottom right with soft shadow; one thin cobalt line as the only accent. Handbook p. 6.

Cinematic feel · Nano Banana 2. A woman in a red coat mid-stride across a wet crosswalk at dusk, motion blur on her body, a bus streaking light trails behind, sodium lamps and shop-window glow. Subject on the left third. Handbook p. 8.

35mm film still: a woman in a red coat dashes across a rain-wet crossing at dusk, mid-stride with motion blur, a bus streaking past behind, sodium streetlight and shop-window glow, muted Kodak palette, natural grain, subject on the left third, full frame edge to edge, no letterboxing, no black bars.

The AI look is the model's default: blanks get filled with average answers

This section is the handbook's foundation. Start with the observation: leave any model alone and it runs toward the same default—subject dead center staring at camera, posed, flat light, everything sharp, plus a coat of processed sheen.

There's a word for that now: slop—the look you clock as AI-generated in half a second. Once you see it, the first reaction is that the brand didn't care. The handbook's stance is simple: never ship it.

Slop is the default. Every parameter you set—where the frame sits, which instant you catch, which lens, what time of day, what light—nudges the image one step away from that default. Every blank you leave, the model fills with its most average answer. The Prompting Handbook · Field notes: killing the slop

The set below is the same model, same person—the cleanest illustration of that line.

Left is true default: one line—"a smiling woman at a farmers market stall holding a basket of vegetables"—and nothing else. Clean, competent, and still dead center, face to camera, staged smile, sharp from front to back: stock-AI energy. Right is the same woman, directed: right third, three-quarter away from camera, caught mid-arrange smiling at someone off-frame, a soft foreground basket framing the shot, warm afternoon side light from the left, a gentle rim on the hat brim, pores and flyaways left on the skin. Count the extra choices on the right—seven or eight. The gap is those choices, not the model. Handbook p. 15; both Nano Banana 2.

Moving an image off default is two moves: first direct the shot (bias the frame, catch a real instant, build foreground/background depth), then pull the grade toward a soft cinematic look (gentle contrast, creamy focus roll-off, fine even grain). The second move has paste-ready copy. The first you write yourself.

soft cinematic lens, shallow depth of field with creamy focus roll-off, gentle contrast with soft highlight roll-off, fine uniform film grain, no oversharpening, no digital look

Natural skin texture: realistic pores, fine hairs, subtle veining, a faint highlight. Tactile and real, not retouched plastic.

Of film artifacts, only one is worth keeping: grain. Dust, scratches, "old damaged film"—leave those out.

Three field notes for holding the frame

On-image type: quote the words, then brief the layout like a designer brief

This is where most prompts break, especially for beginners. Models place type the way designers do, not the way typists do. You're briefing a layout; the copy is just what fills the slots.

Layered type · Nano Banana Pro. Green caps "NATURE" fill the top two-thirds; dark vine leaves sit behind; hand-drawn lilies, marigolds, and roses sit in front—some occluding the letters, some behind. The prompt stated that front/back relationship outright. Handbook p. 9.

Full sentence as subject · Nano Banana 2. "A USER INTERFACE IS LIKE A JOKE. IF YOU HAVE TO EXPLAIN IT, IT'S NOT THAT GOOD." in two blocks filling the frame, a black stick figure mid-knee-slap overprinted in the middle—halftone, distressed ink, two-color print. Every letter and mark lands. Handbook p. 9.

Vintage screenprint design poster, giant green condensed text filling the canvas: "A USER INTERFACE IS LIKE A JOKE." across the top half and "IF YOU HAVE TO EXPLAIN IT, IT'S NOT THAT GOOD." across the bottom half, simple black doodle figure centered in overprint, laughing with head thrown back, one hand slapping a knee, the other holding a computer mouse, halftone dot texture, risograph print effect, distressed ink, off-white paper background, bold graphic design, 2-color print, retro illustration.

No more masking to edit: name what changes, and what must not

Edits are conversational now: you say the change out loud instead of painting a mask. One-line formula—say what changes and what stays put, and only name the one thing that moves.

Swap the background: product and shadow move together

Left: original—a bottle of "LUMIÈRE NATURELLE" serum on a wood kitchen counter. Right: same bottle, same shadow, grey studio sweep behind. The whole prompt does two jobs only: what changes, what must not. Handbook p. 13, Nano Banana 2.

@image1 Keep the product and its shadow exactly; replace the background with a soft, sunlit grey studio sweep.

One original, two different one-line edits

Original

Edit 1 · midday

Edit 2 · person under tree

Composite two images into a moment that never happened

@image1 is a model shot; @image2 is a stainless tumbler product shot. One sentence puts both into the same sunlit moment and keeps her face and the tumbler label. Number the refs, then name them in the sentence—that's the key move for this kind of composite. Handbook p. 13, Nano Banana 2.

@image1 @image2 The model from @image1 leans against a sunlit concrete wall raising the tumbler from @image2 toward camera, laughing off-camera, golden side light, shallow focus. Keep her face and the tumbler's label exactly.

Edit strengths differ by model: Nano Banana 2 is the daily workhorse—regional fine edits, camera moves, focus and grade changes, day-to-night relights. Seedream can merge up to ten references while holding face, light, and palette. Nano Banana Pro can hold five characters and fourteen objects in one frame without cross-talk. Recraft outputs real editable vector, not just PNG.

When a character has to show up again: build a three-panel sheet first, then carry it into every image

For any recurring character, nothing pays off more than this: before they appear a second time, make a three-panel character sheet and treat that sheet as the master for every generation after.

Two rules. First, attach that sheet as a reference image on every later run. Second, write every prompt that describes the character from the sheet—not from memory.

Far left, marked @reference, is the original street photo. The three panels on the right are the generated sheet: head-and-shoulders front view, exact side profile, full-body standing pose, pure white studio backdrop. Same person throughout—dark hair slicked back, magenta acetate sunglasses, blue technical shell with orange webbing, matching trousers, clean sneakers. Handbook p. 17, Nano Banana 2.

reference: her portrait · Character sheet: one 16:9 frame divided into three equal panels on a pure white studio background. Left panel: straight-on head-and-shoulders portrait. Centre panel: exact side profile portrait. Right panel: full body, standing relaxed. The same woman in all three panels, the woman from the reference image, keeping her identity exactly: dark hair slicked back, magenta acetate sunglasses, natural lips, blue technical shell jacket with orange trim details, matching blue trousers and clean sneakers. Soft even studio light, no props, no text, fine film grain, photorealistic.

If you can't name it, you can't get it: a short word list that actually works

The handbook's read is simple: a lot of teams never land the image they want because they don't have the word for the look they're chasing. So it gives a short list.

One hard rule for use: two or three terms per prompt is enough. Don't dump a whole group. Stack too many and they dilute each other—the strongest one stops landing.

What changed on video: one pass, picture and sound together

The biggest shift on the video side in 2026 is the move to "omni" architectures—text, image, video, and audio into one model, out as a finished clip with synced sound, in a single run. Last generation took your text and gave you silent picture; sound was a separate job.

The handbook is concrete on Gemini Omni Flash: sketch direction and try angles there, then finish in Seedance. It generates its own voice and speech, but it won't lip-sync to an audio file you provide. Sora 2 and the Wan line round out a wider map; the handbook doesn't rank them in detail.

One piece of context the handbook skips: Happy Horse showed up anonymously atop Artificial Analysis's text-to-video leaderboard in early April 2026; Alibaba claimed it on April 10. Version 1.1 came from Taotian Group's Future Life Lab and shipped June 23.

Two controls people mix up: start/end frames vs. reference images

You can use both in the same generation: pin a start frame to lock the opening composition, and hang a character reference so that face doesn't drift for the whole clip.

Eight beats of a video prompt, and six rules

A video prompt is a shot description, not a bag of tags. Write it like a director calling a setup for the crew. The eight beats, in order: shot size, camera move, subject, action, scene, light, style, sound.

Social ad · Seedance 2.5. A hand lifts a flask off a sunlit kitchen counter, quick push-in, steam rising. Still from the handbook; the moving version lives on Magnific's in-product canvas. Handbook p. 24.

Cinematic · Seedance 2.5. Low side track: a rider glides the coast at dusk, foreground grass bent by wind. Also a handbook still. Handbook p. 24.

Vertical 9:16 medium shot. A hand lifts a coffee flask off a sunlit kitchen counter, quick push-in, steam rising. Bright natural morning light, clean and airy. SFX: gentle pour, ambient kitchen. No music. No subtitles.

Order varies a little by vendor. Seedance wants shot count, duration, and aspect first, then subject → action → environment → camera → style, with refs as @image1, @video1, @audio1. Kling runs subject + motion + scene + camera + light + mood. Happy Horse follows the same shot-description path. Gemini Omni Flash is the outlier—tuned for conversational direction, so you can talk loosely and revise round by round.

The strongest video edit isn't in the video model: pull the first frame, edit it as an image, send it back

This is the handbook's most counterintuitive move—and the one you can steal whole. It starts from a plain fact: a clip is driven by its first frame.

Follow that out. If the first frame decides how the shot looks, export any clip's first frame, treat it as a still, edit it in an image model (that's where instruction-following is strongest), then hand the frame back to the video model with the original clip attached as a motion reference. You get the same shot, same move, same pacing—only the change you asked for.

1

Start with a clip

Example: a 5-second Seedance 2.5 shot—white sneaker on a pedestal, slow 180° orbit, a rim light sweeping the sole, one continuous take. Generated or live-action both work.

2

Export the first frame as a still

No extra tools required. In Magnific, every generation already stores the first frame with the clip—grab it and treat it like any image. That's the frame above.

3

Edit that frame in Nano Banana Pro

Do the change in an image model—that's where instruction-following is strongest. Same rule as always: name what changes, pin what must not.

@image1 Change the white sneaker to a matte-black colorway with black laces and a black sole; keep the pedestal, the lighting, the camera angle and everything else exactly.

Same pedestal, same light, same angle—only the shoe changed color.

4

Send the edited frame back into Seedance

Set it as the start frame and attach the original as a motion reference. You get the same orbit, same light sweep, same pacing—only the shoe is different.

@image1 @video1 Repeat the exact same slow 180° orbit and rim-light sweep as @video1, now around the matte-black sneaker from the start frame. Same pacing, same framing, one continuous take. No cuts.

First frame of the new clip. Handbook p. 27—the four images are the real outputs of this pipeline in order.

A related move: turn one still into full camera coverage of a set

Same logic, different use. Drop a scene still into Seedance as the start frame (Kling works at 4K too) and ask for a slow walkthrough that rotates and holds a beat at each angle. What comes out feels like a real tech scout: every hold is a frame you can pull, with consistent light and set, ready to use as a later start frame, reference, or plate.

The starting point is only this: a stadium that exists only as an image. The crew needs full camera coverage of the venue. Handbook p. 28, Nano Banana 2.

Eight angles pulled from that single fly-through. Because they all come from one continuous render, crowd, lights, and architecture stay consistent frame to frame—something you can't get generating each angle alone. Handbook p. 28, Seedance 2.5.

start frame: stadium still · Real-estate style drone fly-through shot from the stands of a packed football stadium during an evening match. The camera slowly moves through the tier, rotating and stopping for a beat facing each angle: the camera rotates right and slowly pushes in over the crowd toward the floodlit pitch, then rotates left and slowly pushes in along the rows of cheering fans close to the lens, then rotates 180 degrees and slowly pushes in toward the upper tier and roofline behind.

Eight hard rules of the engine: why three characters max, and why you don't shoot the mirror

The handbook calls this section the physics of generated scenes—and adds a warning: they sound like superstition, treat them like law. Each one was written after a render broke it.

How multi-shot pieces join: five kinds of seam

A multi-shot piece is a chain of seams. Each seam answers a different version of the same question: does the next shot pick up this frame, this action, pass through something, or just break clean?

Twelve camera moves, and two rules that cover all of them

Two rules that cover the lot: never stack multiple moves in one prompt—one move per clip; add "slow" to almost every move. Slow motion hides the places the model can't render cleanly. Fast motion exposes every break.

One last piece: audition the character

The character sheet locks the face. An audition locks the performance. Before a character plays a real scene, shoot a short self-tape and watch how they move, speak, and answer direction. Generate two or three takes with different "voice recipes," pick one, then hang the winning clip as a video reference on every later scene—the same way the sheet carries look. The handbook's math: lock the performance once. It's cheaper than re-rolling scenes until the acting happens to land.

Dialogue · Kling 3.0 Omni

Beat sheet · Seedance 2.5

Audition · Seedance 2.5

Cinematic self-tape audition, one continuous take. Medium close-up, locked-off eye-level camera, shallow depth of field: a woman in her 30s against a plain grey wall, soft window light from the left, listening to an off-camera reader. A beat of silence; she looks down, then back up. She says, quietly: "You knew. The whole time, you knew." Voice: controlled, hurt, almost warm; it catches on the second "knew." Jaw tightens, one slow blink, the trace of a smile that is not one. She holds eye contact and says nothing more. Natural skin texture, fine film grain, muted palette. No music. No subtitles.

6-second product shot, one continuous take. 00:00 the watch lies flat on black stone, still. 00:02 it rises and rotates slowly to face camera. 00:04 it settles as a rim light sweeps across the dial. SFX: low hum, soft click. No music. No subtitles.

That new prompt has a skeleton—six parts: subject, action, scene, light, framing, style. Laid out sideways, each part's job in the frame is obvious.

String the six parts together and you have a full prompt. In the resulting still, the bottle sits on the right, the splash forms a crown, hard light rakes wet slate—each detail was named in advance. Nothing was left for the model to invent.

Same model, same person. Prompt only "a smiling woman holding a basket of vegetables at a farmers market stall," and you get a clean, competent image that is also pure slop: centered subject, facing the camera, posed smile, everything sharp from foreground to back. Move her to the right third, three-quarter turn away, catch the smile aimed off-frame, soft-focus veggie crate in the foreground, warm afternoon side light from the left, pores and flyaways left on the skin—those seven or eight extra choices are the whole gap. The handbook calls the first kind slop, and it hands you two ready-to-paste blocks.

Video got a new architecture this year: text, image, video, and audio feed one model and come back as a finished clip with synced sound. The handbook calls it omni (all-modal); the last generation still needed a separate audio pass. The most counterintuitive move lives here too—the best tool for fixing an already-generated shot is a detour through an image model.

That moving square is the pulled first frame. The first frame pins where the shot starts; the motion reference pins how the camera moves. Half the control is the image model's ability to follow edit instructions; half is the video model holding continuity. Same trick can restyle a whole clip, frame by frame.

The 38 pages pack twenty-plus real original prompts, each paired with the still it produced. Everything below has full copy-paste text in the main piece.