The AI Video Prompting Guide

Soul 2.0 + Soul ID

Character-consistency backbone

• Preset-driven, not prompt-driven A: pick one of 20+ aesthetic presets (Mystique City, Warm Ambient, Editorial Street Style, Nature Light, Theatrical Light…) — the preset carries lighting, camera and style. Your prompt supplies subject, action, setting, outfit. It understands fashion and editorial language natively.
• Soul ID training A: upload 20–80 photos of the person — well lit, varied angles and expressions, at least one full-height shot, from the last 4–5 months. Avoid sunglasses, heavy shadows, crops and group shots. Quality beats quantity. Trains in about 3–5 minutes.
• Cross-model bridge A: the trained character appears under Elements — type @ in the prompt field of Nano Banana, Seedream, Kling, Seedance or Cinema Studio and select the character. This is how one face survives a whole film.
• Don't re-describe the face in the prompt when the character is selected — the ID handles it.
• Failure modes: extreme style shifts cause small drift (expected); a weak training set gives a weak lock; one identity per Soul ID, so multi-character scenes need multiple Elements.

Popcorn

Storyboards

• Manual mode is the filmmaker mode A: per-frame prompts — set shot type (close-up, wide, over-the-shoulder), action and lighting per cell. Auto mode takes one world-prompt and returns a 4/6/8-frame narrative arc.
• Up to 4 reference images, roles assigned by ordinal in prose: "The man from image one standing in the forest from image two, holding the object from image three."
• Subject first — the first-named subject becomes the tracked subject. Concrete action verbs. Lighting and camera early.
• Face, clothing and proportions are tracked across frames automatically; you can fix a single frame without breaking the rest.
• Failure modes: low-res references degrade every frame; mismatched aspect ratios; abstract prompts.

Nano Banana / 2 / Pro

Google — reference control and text rendering

• Narrative prose, not tag lists A: [Subject] + [Action] + [Location] + [Composition] + [Style]. Positive framing only. Edits start with a strong verb.
• Up to 14 reference images — the largest reference budget anywhere here. Assign roles in prose ("the attached sketch as structure, the attached fabric as texture").
• Responds strongly to photographic vocabulary A: camera hardware ("shot on Fujifilm"), lens and aperture ("low-angle, shallow depth of field f/1.8"), lighting design ("chiaroscuro," "golden-hour backlighting"), film stock ("1980s colour film, slightly grainy"). Name materials precisely — "navy blue tweed," not "nice fabric."
• Text rendering is a headline strength (especially Pro): put literal words in quotes, name a font style per line, multilingual is fine.
• Edits: conversational instructions plus an explicit "keep the rest identical." Iterate in follow-ups rather than one giant prompt.
• 1K/2K/4K; ratios 21:9 → 9:16. Soul ID attaches via @Elements.

Seedream 4.5 / 5.0

ByteDance

• 4.5: a five-part order — subject → style → composition → lighting → technical. Front-load the character description. Sweet spot is 30–100 words; under 15 goes generic, over 150 and the instructions start competing. D
• 5.0: six elements (format, subject, composition, lighting, in-image text, style). Order-tolerant but completeness-rewarded; stay under ~600 words. Up to 10 references, named by content ("the tan leather-bound journal") rather than "image 3." Pin the invariants: "Keep the exact silhouette, the exact camera angle, the exact studio lighting from the reference."
• Very responsive to lighting vocabulary (golden hour, low-key, dramatic side light) and camera terms ("85mm lens," "rule of thirds").
• Text: 4.5 renders it well; 5.0 garbles long text — acknowledged — so keep 5.0 text short.
• Failure fixes: inconsistent subject → lead with the subject and add specific traits; style miss → explicit "not cartoon-like" plus reinforcing synonyms. Watch for 5.0's over-smooth skin and action overshoot on dynamic poses.

FLUX.2 Pro

Black Forest Labs

• Subject + Action + Style + Context, most important first. 30–80 words ideal. A
• Two unique powers. Literal JSON prompting for multi-element precision — {scene, subjects:[{description, position, action}], style, color_palette:["#hex"], lighting, camera:{lens, angle, depth_of_field}} — and hex-colour binding: "The sofa is deep teal hex #1B6B6F" holds far better than naming a colour.
• Multi-reference 8–10 images, roles by index ("clothing from image 1, style from image 2").
• Loves hardware realism: "Shot on Hasselblad X2D, 80mm, f/2.8," and era emulation ("2000s digicam").
• No negative prompts. It's guidance-distilled — write "sharp focus throughout," never "no blur." This is the single most common mistake people bring over from SD-era habits.

GPT Image 2

OpenAI

• Order: background/scene → subject → key details → constraints. Labelled segments beat dense paragraphs. A
• The core skill is change/preserve editing: Change: [exactly what]. Preserve: [face, identity, pose, lighting, framing, background]. Constraints: [no extra objects]. Restate the preserve list on every iteration — drift accumulates silently.
• Best-in-class small text (quotes plus typography; spell hard words letter by letter).
• Photorealism: say "photorealistic," add real-texture cues ("pores, fabric wear"), and prefer "35mm film photograph" over polish words.
• Character template: "Same outfit, features and proportions; new scene and action. Do not redesign the character" plus an anchor image.

Recraft

Titles, posters, prop graphics

• Structured order: concept → environment → framing/pose → attributes → spatial relations → lighting → camera → mood. A
• Typographic systems are the specialty: quoted text plus format type, hierarchy, placement and colour blocking. For vector and logo work: "no gradients, no shadows," strict palette.
• Avoid stacked evaluative adjectives — "stunning," "gorgeous" — they measurably degrade output. Style-reference copies layout, angle and composition strictly, which is what makes uniform title-card sets possible.

Draw-to-Video / Draw-to-Edit

Input mode, not a model

• Sketch the motion directly on the start frame A: arrows for direction and speed (longer = stronger, curved = arcs), short text labels for actions ("man walks"), numbers for sequence order.
• Draw-to-Edit takes up to 8 references — drop an object in and size it to match perspective.
• One redraw beats five prompt rewrites.

Upscale

Finishing

• Image presets: Standard, High Fidelity v2, Text Refine (signs and titles), Art & CG, Low Res. Video: Proteus (standard), Iris (faces and portraits — use it for character shots), Gaia, Theia. A
• Settings: Face Enhancement 70–85%; denoise ≤50; sharpness above 70 causes artifacts. AD
• Upscaling will not fix composition or heavy motion blur. Fix those upstream.

Seedance 2.5 / 2.0

ByteDance — the workhorse

• Official formula A: Subject + Action + Setting + Visual style + Camera work + Sound — only subject and action are mandatory. For multi-shot, declare shot count, total duration and aspect ratio upfront, then baseline (film stock, aesthetic) → environment and character → numbered shots.
• 2.5 A: up to 30s per generation. The UI selectors (era, genre, lighting, physics, lens, tone, pacing) do heavy lifting, so the prompt can stay short — around 350 words for a 30s clip is the official sample scale. Up to 50 references (practically ~8–12 image subjects). Audio is generated in the same pass. Region edit fixes one object, face or background without re-rolling the clip — use it instead of retrying.
• References: "@image1 provides the character's face. Do not use its background." For multi-angle references of one object: "all three images define one single [object]."
• Camera: name the real move, and state exclusions to lock POV — "No cuts, no zoom, natural head movement." Handheld: "unstabilized, micro-jitters, abrupt jerks."
• Audio AD: specify a sound list or "NO MUSIC / NO SFX" — silence left unstated gets a score. 2.5 delimiters D: music ( ), SFX < >, dialogue { }, subtitles 【 】. Declare language and accent before the lines.
• Failure fixes: drift on long clips → timed beats plus per-segment camera; feature dilution → fewer references; background leakage → explicit exclusions; random cuts → "no cuts."

Kling 3.0 / 2.6

Kuaishou — multi-shot, cheap drafts, lipsync

• Write directions to a scene, not a list of objects. D The pattern is a master prompt (style anchor, 2–3 traits per character, environment, palette) plus per-shot prompts (framing, camera, beats, dialogue, duration).
• Multi-shot AD: up to 6 shots / 15s in one generation. Label each shot with framing and duration; characters and grade carry across automatically.
• Start and end frame supported (or end-frame only) — the continuity power tool between scenes.
• Dialogue D: [Name: role, tone]: "Line." Action before the line; unique names, never pronouns; "Immediately" between speakers kills dead air. Line budget is roughly 8–12 words per 5s before lip sync falls apart. Voice Binding (3.0 Omni, official) binds a voice from 5–30s of clean audio or a 3–8s single-speaker video.
• Negatives: list concrete artifacts — "frozen lips, warping fingers, jittery eyes, character drift."
• Kling 2.6 is the budget single-shot B-roll and cheapest lipsync option; use the simple subject + action + environment + camera formula there.

Veo 3 / 3.1

Google — atmosphere and native audio

• Formula A: [Cinematography] + [Subject] + [Action] + [Context] + [Style & Ambiance], written in prose. Clips 4/6/8s, 16:9 or 9:16.
• Image-to-video: the image carries subject and scene; the prompt carries motion and audio. Ingredients-to-Video supplies references by role ("Using the provided images for the detective and the office…"). First+last frame: describe the camera path bridging the two stills.
• Audio syntax A: dialogue in quotes with delivery in prose (She says, "We have to leave now," in a weary voice); SFX: thunder cracks; Ambient noise: quiet hum of the bridge; music as a described line ("a swelling orchestral score begins").
• Timestamp prompting gives you mini-edits in one generation: [00:00-00:02] Medium shot… [00:02-00:04] Reverse shot…
• The gotcha: quoted dialogue tends to burn in captions. Append "no subtitles, no captions, no on-screen text." D Describe exclusions positively where you can — "desolate landscape with no buildings" style.

Wan 2.5 / 2.6 / 2.7

Alibaba — frame control and Speak

• Formula A: Entity + Scene + Motion (plus aesthetic control, stylization and sound).
• Image-to-video: the image owns entity, scene and style; the prompt is motion and camera only, with tempo words ("slowly," "quickly"). Re-describing the image makes it fight you.
• Camera: verb phrases (push-in, pull-out, tracking, tilt up, fixed camera). Keep orbit arcs under ~45°. "Generate single shot" prevents unwanted cuts.
• Audio A: voice = line + emotion + tone + speed + timbre + accent; speaker-labelled lines (Explorer: "…") with stage directions ("trembling voice, calling out over wind"). Suppress with "No dialogue" / "No background music." Wan Speak (LipSync Studio) maps speech onto footage while preserving identity — the fix when word-exact lipsync matters.
• 2.7 multi-shot: Shot 1 [0–3 s]: …, references cited as "Image 1" / "Video 1". 2.6 takes up to 3 reference characters ("character1…").
• Known limits A: real people's names are rejected; one clip is one continuous shot outside explicit multi-shot; exact on-screen text won't render; long choreography breaks, so keep actions simple.

MiniMax Hailuo 02 / 2.3 / H3

Bracket director syntax and physics

• Hailuo 02 A: the unique bracket director syntax — [Push in], [Truck left], [Pedestal up], [Static shot], [Tracking shot]… up to three combined: [Truck left, Pan right, Zoom in]. More than three degrades. First and last frame supported.
• Hailuo 2.3 A: short, physical prompts. Verbs that translate to physics — "dissolve," "ignite," "shift into light." Strengths are transformations, elemental FX and anime line stability. Camera comes from preset categories in the UI, not inline. Image-to-video rule: describe only what changes, and never ask for a wide shot from a close-up input. Fixes: morphing → one action only; static output → add motion verbs; "deep-fried" look → delete "8k, masterpiece" quality tags. No native audio documented.
• H3 D — no official prompt documentation exists yet. Community formula: reference description + core idea (1–2 sentences) + a time-stamped visual process; 50–120 words typical; camera in natural language tied to time windows ("0–3s: camera slowly pushes in"). Write dialogue verbatim — paraphrase leaves no lipsync reference. Specify the ending state or the last frame is unusable. For identity, enumerate every feature; never "keep consistent."

The 65 camera presets (DOP)

Named moves as selectable controls

Zooms: Zoom / Rapid Zoom / Crash Zoom In & Out, YoYo Zoom, Eating Zoom, Eyes In, Mouth In · Dolly: Dolly In/Out/Left/Right, Super Dolly, Double Dolly, Dolly Zoom (the Vertigo warp) · Orbit: 360 Orbit, Arc L/R, Lazy Susan, 3D Rotation · Crane & aerial: Crane Up/Down/Over-The-Head, Jib, Aerial Pullback, Overhead, FPV Drone, Flying Cam · Pan & tilt: Pan L/R, Whip Pan, Tilt, Dutch Angle, Incline · Handheld & rig: Handheld, Static, Wiggle, Snorricam (body-locked), Robo Arm, BTS, Hero Cam, Glam, Head Tracking, Object POV · Vehicle: Car Chasing, Car Grip, Buckle Up, Road Rush · FX: Through Object In/Out, Bullet Time, Focus Change, Fisheye, Low Shutter · Time: Hyperlapse, Timelapse variants.

Cinema Studio

The assembly environment

• Layered prompting A: (1) scene setup — location, time, weather; (2) numbered shot breakdown with camera type per shot ("Shot 1: telephoto… Shot 2: handheld tracking…"); (3) character action, blocking and dialogue; (4) technical specs — lens, movement, lighting quality.
• Hardware selectors: 5 camera profiles (ARRI/RED/IMAX), 11 lenses from 8–75mm. Focal-length choice, not adjectives, is what makes it cinematic — telephoto compresses, handheld shakes, wide gives context. D
• Lighting written contextually ("dusk sky," "warm tungsten," "silhouetted by sunset"), not as parameters. Pacing inline ("0–4 seconds: …").
• References via @image / @video tags in every prompt. Character drift → character sheets plus tagged references every time. One camera move per shot.