Midjourney Photorealism Prompts: A Guide to AI Video Porting

Why photorealism prompts change meaning when they start moving

Photorealistic stills and photorealistic video are frequently treated as one skill with two export options. They are different disciplines. A still is judged as a single frozen decision: one camera position, one lighting moment, one expression. A clip is judged as a chain of decisions that has to stay believable for several seconds, and motion exposes inconsistencies that a still can hide behind a well-chosen frame.

When you paste a carefully tuned text-to-image prompt into a video generator, the first frame is usually fine. The composition holds. Then the model starts improvising. Skin texture drifts, background details warp, a hand becomes an impression of a hand, and the slow dolly you asked for turns into a gentle smear. Nothing in the prompt was wrong; the prompt simply described a moment instead of a movement.

That is the core translation problem. A still prompt answers the question "what should this one frame look like?" A video prompt has to answer three more: what changes, how fast, and what must not change. Write with those three answers in mind and photorealism stops being a lottery and becomes a repeatable process.

This guide is a practical workflow for porting strong photorealistic still prompts into video models. It covers prompt anatomy, the camera language generators actually honor, subject consistency, syntax differences between model families, a step-by-step porting loop, troubleshooting, and the templates worth keeping in your library.

Deconstructing a still prompt into portable parts

Most strong photorealistic prompts are dense bundles of five or six ideas squeezed into one sentence. That density is a feature for image models, which weigh many tokens at once, and a liability for video models, which must distribute attention across time.

The five components worth keeping

Strip any prompt down and label each clause:

1. Subject and state — who or what, plus wardrobe, expression, material, or condition.
2. Environment — location, time of day, weather, background density.
3. Lens and format — focal length, aperture, film stock, sensor character.
4. Lighting — direction, quality, color temperature, practical sources.
5. Grade and mood — contrast curve, palette, era, reference feel.

Everything else is usually noise. Brand names, decades, and photographer references can stay, but only when they map to a visible decision. "Shot on a medium-format camera with shallow depth of field" gives a video model something to preserve. "Award-winning" gives it nothing at all.

Clauses that break in motion

Some phrases are stable across stills but degrade instantly in video:

• Static framing words such as "perfectly centered" or "symmetrical" fight any camera movement you later request. Pick one or the other.
• Texture micro-detail like "every pore visible" tends to flicker frame to frame because the model has no persistent texture memory.
• Composite language such as "double exposure" or "collage" confuses temporal models that assume one continuous space.
• Vague motion adverbs like "dynamic" or "cinematic energy" produce camera shake rather than intentional movement.

A useful rule: if a clause cannot be drawn as a change between the first frame and the sixtieth, it belongs in the still prompt, not the video prompt.

Rewriting stills language as shot language

The translation step is mostly grammatical. Stills prompts lean on adjectives; video prompts lean on verbs and camera nouns. Rewrite each aesthetic adjective into an action or a movement that produces the same feeling.

"A moody portrait" becomes "she turns her head slowly toward the window light." "Epic landscape" becomes "the camera rises above the ridge as fog rolls down the valley." The mood survives, and the model now has something to animate.

Camera moves that survive generation

Models vary widely in how well they execute camera instructions, but a few move types are consistently reliable:

• Slow push in — the safest way to add intention without exposing artifacts.
• Lateral truck — excellent for revealing environment; keep the speed low.
• Static frame with internal motion — the highest-fidelity option when the subject moves and the camera does not.
• Handheld drift — forgiving, because imperfection reads as style rather than error.

More ambitious moves — crane shots, orbits around a subject, rack focus between two planes — work, but they consume a larger share of a clip's runtime to look right. If your shot is only four seconds long, an orbit will not complete. Match the move to the duration, not to your ambition.

Lighting and color continuity

Video models often re-decide lighting on every frame. To reduce drift, state the light source physically and keep the number of sources small: "one warm practical lamp behind the subject, cool window light from camera left." Avoid stacking five light descriptions. The model averages them into mush and then changes its mind halfway through the clip.

For color, name the grade rather than the mood. "Low-contrast grade with lifted blacks and warm highlights" outperforms "beautiful color." If you are generating a sequence, keep the grade sentence identical across every shot. A shared clause is the cheapest continuity tool available, and it costs nothing but discipline.

Keeping faces, products, and places consistent across shots

Consistency is where ambitious projects fail. A single beautiful clip is easy; four clips that read as the same scene are hard. Order your options by reliability:

• Image conditioning — generate or pick a locked keyframe and drive the video from it.
• Multi-image fusion — upload several angles of a subject so the model averages a stable identity. Best for products.
• Subject or character locks — some tools offer dedicated features; use them before relying on prose.
• Seed reuse — reproducible noise gives modest stability, not identity.
• Textual identity anchor — a repeated description in exact words across prompts.

That last one is underrated. Write an "identity block" of 30 to 50 words describing your subject in precise, unchanging language, then paste it verbatim into every prompt in the sequence. Vary only the action and camera clauses. Models attend to repeated tokens, so a consistent paragraph does more work than scattered synonyms.

For products, use multi-image fusion whenever it is available and keep the background clause identical across shots. For locations, generate a wide establishing shot first and then use it as the reference for closer coverage. This mirrors how a real crew works: establish geography, then move in.

Model-specific syntax: what each family expects

No single prompt string ports cleanly to every generator. Treat each family as a dialect with its own grammar.

Concise cinematic models

Some engines respond best to short, present-tense, sentence-like prompts with explicit camera terms. They handle human motion and physical interaction well, and they reward the phrase "one continuous shot." Skip the comma-separated tag stack you would use for a still image. Write prose. If you want a specific move, name it in its own short sentence.

Others in this family favor stylized realism and quick physical actions. They do better with a clear subject-action-object structure and shorter prompts overall. When results feel chaotic, cut length rather than adding detail.

Temporal control models

Engines built around motion control, especially in image-to-video mode, benefit from separating the camera directive from the scene description. Put the scene in one sentence and the camera in another, using an imperative: "Camera pushes in slowly." This prevents the model from blending the move into the subject description.

Narrative-style engines favor natural phrasing and long descriptive sentences. They like the pattern "the camera does X as the subject does Y," which binds motion of the frame to motion inside the frame. Use that construction when you want both to happen at once.

Open-weight and effect-focused pipelines

Open-weight models let you steer look through settings, motion strength, and style adapters rather than words. Your prompt can be simpler because much of the aesthetic is handled outside it. Effect-focused tools favor punchy short prompts and keywords tied to their own effect vocabulary. They are weaker at literal photorealism and stronger at motion effects, so use them for transitions and accents rather than hero shots.

A general rule that holds across families: three sentences maximum. Sentence one is subject plus action. Sentence two is camera. Sentence three is lighting and grade.

A step-by-step porting workflow

Here is the loop that keeps quality high and wasted attempts low.

Step 1: Lock the keyframe. Generate the still in your image model, upscale it, and pick one frame you would be happy to publish. Video models inherit the flaws of their source frame, including any weird hand or soft corner.

Step 2: Label the prompt. Split your still prompt into the five components from earlier. This takes two minutes and saves twenty.

Step 3: Write the identity block. Thirty to fifty words, unchanged across the whole sequence.

Step 4: Convert adjectives to actions. Every mood word becomes a movement or a gesture.

Step 5: Add exactly one camera move with an explicit pacing word: slow, gradual, gentle, drifting.

Step 6: Add the lighting and grade sentence. One light setup, one grade.

Step 7: Generate two or three short takes. Short means four to six seconds. Long clips hide errors less well and consume far more compute.

Step 8: Diagnose before rewriting. Name the failure precisely: subject drift, motion quality, camera adherence, lighting shift, or texture flicker.

Step 9: Change one variable per iteration. If you change three things and the result improves, you have learned nothing reusable.

Step 10: Extend only after the base looks right. Lengthening a flawed clip multiplies the flaw.

Troubleshooting common failures

Faces warp during movement. The motion is too large for the model to hold identity. Reduce the amplitude, add image conditioning, and describe stable facial features explicitly.

Textures flicker. Remove pore-level and fabric-weave detail. Ask for softer, less granular texture, and accept that the model will invent some surface noise anyway.

The background melts. Simplify the environment. Fewer distinct objects means fewer things that can morph. Replace a busy street with three named elements.

The camera ignores your request. Move the camera directive into its own short imperative sentence and name the speed. If it still fails, the move is probably too complex for the runtime.

Motion is too fast. Add pacing words and reduce the number of simultaneous actions. One subject doing one thing always looks better than three.

Two subjects merge into one. Describe their positions relative to each other and the frame, then keep only one of them moving.

Color shifts mid-clip. You are stacking mood, grade, and film stock. Keep the grade clause and delete the rest.

Text and logos look wrong. Generate the shot without text and composite the real thing afterward.

Everything looks plastic. You have over-smoothed. Request film grain, imperfect skin, and micro-movement in the frame.

Quality control and iteration discipline

Generating is fast; judging well is the bottleneck. Watch each take at full speed first, the way an audience would. If it holds up, then step through frame by frame to find the defects you missed. Defects that only appear on pause are usually acceptable; defects that appear at speed are not.

Judge at the target resolution and aspect ratio. An image that looks sharp in a square preview can fall apart in a tall vertical crop, especially around the edges of the frame where models tend to lose coherence.

Keep a small log with five columns: prompt, model, seed, settings, verdict. After a week you will have a personal map of which phrasings work in which system, and that map is worth more than any generic prompt list. Set an attempt budget per shot, perhaps ten generations, and a stop rule. When you hit the limit, change the approach rather than the wording.

Adaptable prompt templates

Use these as skeletons rather than finished prompts.

Dialogue portrait shot. Identity block. Subject turns slightly toward camera and holds eye contact. Camera slowly pushes in. Soft window light from camera left, low-contrast grade with warm highlights, fine grain.

Product hero shot. Identity block describing the object and its material. Object rotates a quarter turn on a matte surface. Camera drifts laterally at a slow constant speed. Single soft key light above, dark neutral background, clean grade with deep blacks.

Cinematic landscape. Identity block describing terrain and weather. Fog moves across the valley and grass bends in the wind. Camera rises slowly above the ridge. Overcast directional light, cool palette, low saturation, subtle grain.

Notice what these templates avoid: no stacked adjectives, no more than one camera move, no mood words that cannot be drawn.

FAQ

Can I use the same prompt across different video models? No. Port the intent, not the string. Restructure it into the three-sentence format and adjust for each model family's preferences.

How long should a video prompt be? For most models, 40 to 80 words is the sweet spot. Shorter loses control; longer dilutes attention and increases the chance of contradictions.

Do seeds transfer between tools? No. A seed only means something inside the system that produced it. Port the prompt, not the random number.

Why does my output look artificial? Usually because the prompt asks for perfection. Request grain, slight asymmetry, imperfect surfaces, and a handheld quality. Realism lives in small flaws.

Should I generate video from text or from an image? Image-to-video gives you far more compositional control, and it is the right default whenever you already have a frame you like.

How do I handle dialogue or lip sync? Keep video generation and speech synthesis separate. Generate the shot with a neutral performance, then sync audio in an editing tool.

What about vertical formats? Decide the aspect ratio before you generate anything. Compositions built for widescreen rarely survive a crop to vertical without losing their subject.

How many takes should a shot need? A well-ported prompt usually needs three to six attempts. If you are past fifteen, the problem is in the prompt structure or the keyframe, not in luck.

The through-line in all of this is simple: still prompts describe a look, and video prompts describe a change. Once you start writing the change explicitly, keeping the look stable, and letting the model handle only what it is good at, photorealistic video stops feeling like gambling and starts behaving like craft.