Orelon logoOrelon
料金

ChatGPT Image Prompts for Cinematic AI Video Workflows

2026年11月6日 · Orelon Team 著

AI動画テンプレートを見る

着想のためにコミュニティ作品をいくつか閲覧し、任意のテンプレートを開いて Orelon で作成を続けましょう。

Turn ChatGPT image prompts into cinematic AI video with a repeatable workflow: shot cards, camera language, motion tests, review checks, and FAQs.

Most disappointing AI video starts with a prompt that sounds good but directs nothing. It says cinematic café at night, and the model supplies a generic café, a generic night, and a camera that drifts because no one told it what to do. ChatGPT is useful in this situation, but not as a machine that turns vague taste into perfect frames. It is useful as a writing partner that forces decisions: who is in the shot, what the audience should notice, where the light comes from, how the camera behaves, and what changes during the clip. This guide gives you a practical workflow for using ChatGPT to write image prompts, convert those prompts into motion prompts, test a short scene, and review the output before you commit to a final render. Orelon is where that workflow becomes a moving image once the decisions are clear.

Why the Prompt Is a Directing Document

An image prompt is not a wish. It is a compact directing document. It tells a visual model which details matter and which details can be invented. A strong prompt usually answers seven questions: who or what is the subject, what is the subject doing, where the scene takes place, what time and weather apply, what lens and framing shape the view, what light and palette set the mood, and what should be avoided. If you leave three of those unanswered, the model fills the gaps with averages. Averages are why so many generations look polished and forgettable.

Video raises the stakes. A still prompt describes a visible moment. A motion prompt also describes change: a hand moves, steam rises, a curtain lifts, a camera pushes forward, a face turns from doubt to recognition. The still has to be strong before motion can help it. If the frame is vague, motion will expose the vagueness. The subject will wobble, the background will crawl, and the viewer will have nothing to follow.

Consider two prompts. Weak: a cinematic coffee shop, moody lighting, beautiful, highly detailed. Strong: a barista in a small coffee shop at dawn, medium close-up from across the counter, steam rising from a cup, window light from camera left, warm brown and pale green palette, shallow depth of field, 16:9, no text. The second prompt is not better because it is longer. It is better because it gives the model a subject, a position, a light source, a palette, and a delivery format. It also gives the viewer a place to look.

When you work with ChatGPT, ask it to preserve that level of decision-making. Do not ask for a beautiful prompt. Ask for a prompt that a stranger could sketch, a camera operator could frame, and a colorist could grade. That standard is more useful than any list of magic keywords. It also makes feedback actionable. Instead of saying make it more cinematic, you can point to the lens line, the light direction, or the motion instruction and change one variable.

A directing document also protects you from a common trap: prompting a genre instead of a moment. A cyberpunk street gives the model a costume. A courier steps over a puddle while a neon sign reflects in the water, medium wide, low angle gives it a shot. Specificity is not decoration. It is the difference between describing a world and describing a frame inside that world.

What ChatGPT Can and Cannot Do

ChatGPT is a language tool. It can hold a brief in view, ask clarifying questions, generate variations, criticize weak wording, and organize a sequence. It can translate a mood board into shot language, turn a product brief into camera angles, and rewrite the same idea for a vertical format. It can also catch contradictions before you spend time generating: daylight on a street at night, a close-up that also needs to show a full room, a character who turns left while the camera moves right.

It cannot render the image or guarantee what a video model will do with your words. It does not know the final model's training data, motion biases, or failure modes unless you test and report back. It also cannot replace taste. If you ask for a masterpiece, it may give you a prompt full of empty praise words. Your job is to keep the conversation grounded in visible evidence.

Use it as a translator and a critic

The most productive relationship is simple. You bring the intent and the constraints. ChatGPT turns that into a structured description. Then it switches roles and critiques the description as if it were a skeptical director. A useful prompt for that second role is: act as a director of photography and list every visual choice that is missing, contradictory, or impossible to film. That request produces more value than asking for ten more adjectives.

You can also use ChatGPT as a storyboard planner. Give it a scene goal and ask for three shots: an establishing shot, a detail shot, and an emotional close. For each shot, require one subject, one action, one camera move, and one continuity anchor. This keeps the sequence coherent without turning every prompt into a novel.

What ChatGPT should not do is invent production facts you have not decided. If you do not know the aspect ratio, the delivery platform, or the character's wardrobe, say so. Ask for options with tradeoffs. A prompt written on top of hidden assumptions is harder to fix than a prompt with an explicit gap.

A Prompt Framework That Survives Model Changes

Different visual models respond to different vocabulary. Some are sensitive to lens language. Some respond better to plain scene description. Some treat camera movement terms as strong suggestions, while others ignore them. A framework helps because it separates the idea from the wording. You can change the phrasing without losing the scene.

Five layers that carry the idea

Use five layers for stills and motion. Keep each layer to one sentence unless the shot truly needs more detail.

  • Subject and action: who or what, doing what, with what expression or posture.
  • Environment and time: place, weather, era, time of day, and one environmental detail that matters.
  • Lens and framing: wide, medium, close, overhead, focal length, depth of field, and subject position.
  • Light and palette: source, direction, hardness, contrast, two or three colors, and one accent.
  • Constraints: aspect ratio, duration, text policy, negative space, motion limits, and things to avoid.

A reusable skeleton looks like this:

Subject: ...
Action: ...
Setting: ...
Camera: ...
Light: ...
Palette: ...
Style: ...
Avoid: ...
Output: 16:9 still, clean edges, no text.

Ask ChatGPT to fill the skeleton in plain language. If a layer needs two competing sentences, the idea may be two shots. That is not a failure. It is an edit decision made early.

Write the still before the motion

Start with a frame that works without movement. A still prompt for a glassblower might read: a glassblower shaping a vase in a dark workshop, medium shot from the side, orange furnace light from camera right, deep shadows, amber and charcoal palette, 35mm lens, shallow depth of field, 16:9, no text. That frame has a subject, an action, a light source, and a palette. Motion can now be added without rewriting the scene.

A motion version adds only what changes: the glassblower rotates the pipe, the glass glows and softens, sparks drift through the furnace light, the camera makes a slow fifteen-degree arc, five seconds, gentle natural motion. Notice that the motion prompt does not introduce a new room, a new time of day, or a new wardrobe. It protects continuity by repeating the anchors.

Ask for assumptions, not applause

After ChatGPT writes a prompt, ask it to list the five most important visual choices it made and what it would change if the scene were night instead of day, or close-up instead of wide. This exposes hidden decisions. You may discover that the model assumed a modern setting when you wanted a period workshop, or that it placed the subject in the center when you needed negative space for a title.

Assumptions are not mistakes if they are visible. They become mistakes when they stay hidden until the final render. Make ChatGPT surface them, then decide.

Build a Shot Card Before You Generate

A shot card is a short block of text that travels between ChatGPT, an image tool, a video tool, and an editor. It is more portable than a long paragraph because it separates intent from execution. Use six lines.

  • Intent: what the audience should notice.
  • Frame: subject position, shot size, and aspect ratio.
  • Light: source, direction, and quality.
  • Palette: two or three colors plus one accent.
  • Motion: one primary action and one camera move.
  • Guardrails: what to avoid.

Example for a letterpress studio: Intent, reveal the precision of hand-set type. Frame, medium close-up of hands placing a metal letter, centered, 16:9. Light, soft window light from camera left with a warm bounce from paper. Palette, ink black, cream, faded red. Motion, fingers press the letter into the form, slow push in. Guardrails, no extra hands, no text overlays, no fast movement.

Why the card beats a long prompt

A long prompt is hard to revise. When something fails, you do not know which sentence caused it. A shot card gives you named lines. If the hands look wrong, change the frame or guardrail. If the mood is flat, change the light or palette. If the clip feels chaotic, reduce the motion line. This is faster than rewriting the whole paragraph and hoping.

The card also helps a team. A client can approve the intent without understanding lens language. An editor can read the frame and motion notes. A second generation can keep the same card while changing only the aspect ratio for a vertical cut. Store the card next to the final output so you can rebuild the look later.

Turn a Still Prompt Into a Motion Prompt

Start with a still that already works. Then add motion in four places: subject, camera, environment, and light. Limit yourself to one primary motion and one secondary motion. Too many movements compete, and the model averages them into a soft, drifting clip that does not read as a deliberate shot.

Still prompt: a bookbinder in a narrow workshop, seated at a bench, hands stitching a leather spine, medium close-up, soft window light from the left, warm brown and deep green palette, shallow depth of field, 16:9. This is a clear frame. The viewer knows where to look.

Motion prompt: a bookbinder stitches a leather spine at a narrow bench. Medium close-up, hands remain in frame, needle moves steadily through the leather, dust motes drift in the window light. Camera holds mostly still with a slow five-degree push in. Warm brown and deep green palette, shallow depth of field, 16:9, five seconds, gentle natural motion. This version adds a subject action, a small camera move, an atmospheric detail, and a duration. It also sets a motion limit with gentle natural motion. That phrase is not a guarantee, but it steers the model away from dramatic flourishes.

Keep one primary motion and one secondary motion

The primary motion is the thing the audience must see: the needle moving, the glass rotating, the runner entering frame. The secondary motion is support: dust, steam, fabric, or a slight camera drift. If you add a third strong motion, such as a moving background crowd plus a moving camera plus a gesturing subject, the model has to choose what to stabilize. The result is often warping or flicker.

For continuity, describe the first and last frame when it matters. Starts with hands centered on the spine, ends with the needle pulled through and the camera slightly closer. This gives the model a small story arc instead of a looping texture. It also gives you a clear review target. You can compare the actual first and last frame against the description and decide whether to shorten the shot.

A Test Scene With Three Shots

Suppose you need a short opener for a documentary-style piece about a letterpress studio. Three shots are enough: establish, detail, and emotional close. Ask ChatGPT for a shot list with one job per shot, then write each prompt with shared anchors.

Shot Job Prompt core Motion
One Establish the space Wide shot of a letterpress studio, presses in rows, window light from the left, ink black and cream palette Slow push in
Two Show craft Macro of hands locking a metal letter into a form, dust in the air, same warm palette Fingers press, camera static
Three Give the human reason Close-up of the printer's eyes as they inspect a proof, faint window reflection Slight head lift, light shift

Continuity anchors that survive editing

Use the same window light from camera left in all three prompts. Keep the same ink black, cream, and faded red palette. Keep the same shallow depth of field. Do not introduce a new room, a new time of day, or a new wardrobe in shot two. The audience reads continuity from small repeats: the same light direction, the same paper texture, the same machine color.

Ask ChatGPT: which details in these three prompts could break continuity? It will often catch changes in lens, light direction, or subject position that you missed. Then generate each shot as a short test. Review the first and last frame before extending the strongest one. If shot two does not match shot one, fix the prompt before generating a longer version. A cheap test is more useful than an expensive rescue.

The prompt library is useful here because it lets you compare reusable structures rather than starting from a blank page. Keep the structure, change the scene. That is how a sequence stays coherent while each shot does a different job.

Use ChatGPT as an Editor, Not an Oracle

ChatGPT is strongest when you give it a job with constraints. A useful loop looks like this. Write a forty-word brief with the audience, subject, mood, and delivery format. Ask for six prompts across three visual directions, each under eighty words. Score each prompt for clarity, specificity, feasibility, and continuity. Ask for revisions on the top two, removing any word that does not change the image. Generate small tests, compare, and keep a winning prompt as a template.

Scoring is more useful than taste alone. Clarity asks whether a stranger could sketch the frame. Specificity asks whether the image has a particular idea rather than a stock genre. Feasibility asks whether the model can render the hands, text, crowd, or reflection. Continuity asks whether the prompt can survive a second shot in the same scene.

A revision pass that saves generations

Before generating, revise in passes. First pass: remove abstract praise words such as beautiful, stunning, epic, or masterpiece. They add no visual instruction. Second pass: check that every noun could be photographed. Third pass: read the prompt backward and confirm that light, lens, and subject do not contradict each other. Fourth pass: check the delivery format. If the final use is vertical, prompt vertical from the start rather than cropping a wide frame later.

Use four questions as a final gate. Is there one subject? Is there one action? Is there one main light source? Is there one reason for the camera position? If any answer is no, simplify. A strong prompt is not a longer prompt. It is a set of decisions that survive translation into pixels.

Keep a prompt log that teaches you

Record the final prompt, the aspect ratio, the model, the duration, and a short note about what failed. After a few sessions, patterns appear. Certain phrases create soft edges. Certain camera terms create unwanted movement. Certain color words create a plastic look. A log turns random testing into a personal playbook.

When a generation fails, change one variable at a time: light first, then lens, then motion. Changing everything at once teaches you nothing. If the hands fail, simplify the action and shorten the clip. If the background warps, reduce camera movement. If the mood is wrong, change the palette before adding more style words.

Choose the Right Workflow for the Shot

Use an image-first workflow when composition, character likeness, or product detail matters. Generate a keyframe with the AI image generator, refine it until the frame is right, then describe the motion. This is the safest path for branded products, recurring characters, precise layouts, and thumbnail-like frames that must read at a glance.

Use a video-first workflow when the idea depends on movement: weather, dance, abstract light, a camera move through a space, or a gesture that cannot be judged from a still. Start with a short motion test in the AI video generator, then extract a frame as a visual reference if you need consistency. Video-first is faster for discovery, but less controlled for exact composition.

Image-first, video-first, or hybrid

Need Start with Why
Product detail Image You can inspect edges, labels, and reflections before adding motion
Character likeness Image A keyframe gives you a stable face and wardrobe to repeat
Weather or atmosphere Video The idea is the movement of rain, fog, or light
Abstract transition Video The value is in the change, not the still
Narrative sequence Hybrid Build a master frame, then write shot cards for each beat

Hybrid is usually best for narrative work. Build one master frame, write a shot card for each beat, and use the same lens, palette, and wardrobe notes across prompts. ChatGPT can turn those notes into a shot list and flag continuity gaps before you generate. The video templates can help you test formats without rebuilding the same frame for every aspect ratio.

A practical decision rule: if you would storyboard the shot as a still, start with an image. If you would describe the shot as a movement, start with video. If you need both control and motion, build the still first and then animate the strongest frame.

Quality Control and Common Fixes

Review at full size, not in a grid. Check hands, teeth, eyes, jewelry, reflections, and the edges of the frame. In video, watch for flicker, warping, texture crawling, and limbs that change shape between frames. Pause at the first and last frame. If the composition collapses at either end, shorten the shot or simplify the motion.

Check whether the motion supports the story. A slow push should reveal something. A parallax move should create depth. A handheld drift should feel intentional. If the movement is decorative, remove it. Also check the delivery format. A vertical crop can destroy a carefully placed subject, so prompt with the final aspect ratio from the start.

A five-minute review pass

Watch once for story. Does the shot do its job in the sequence? Watch again for artifacts. Do the hands, teeth, and edges hold? Watch on a phone. Does the subject read at a glance? If you need titles, is there deliberate negative space? Finally, compare the result to the shot card. If the intent is missing, do not polish the render. Fix the prompt.

Common mistakes and fixes

  • Style stacking: cinematic, hyperreal, eight-kay, masterpiece. Choose one visual direction and describe its evidence.
  • Contradictory light: soft noon sunlight and dramatic neon shadows. Pick one source and one direction.
  • Too many actions: she turns, smiles, lifts a cup, and walks away. Use one action per shot.
  • No aspect ratio: add 16:9, 9:16, or 1:1 before generating.
  • Text in frame: avoid unless short and necessary; add text later in an editor.
  • No negative space: specify where the subject sits if you need titles.
  • Generic genre prompts: replace the genre with a specific moment, camera position, and light source.
  • No review loop: generate one test, inspect it, change one variable, and repeat.

Another common error is asking ChatGPT for a final prompt too early. The first draft is for discovery. The final prompt is for production. Treat the first three drafts as sketches, then tighten the best one until every line has a visual job.

FAQ

Can ChatGPT generate the image itself?

ChatGPT can help write, critique, and revise prompts, but the rendering happens in an image or video model. Treat it as a director's assistant, not the camera. The value is in clearer decisions, not in a hidden render button.

How long should a prompt be?

For a still, forty to eighty words is often enough. For motion, sixty to one hundred words gives room for one action, one camera move, and a duration. Longer prompts are useful only when every sentence adds a visual decision. If a sentence only adds praise, delete it.

How do I keep a character consistent?

Write a character sheet: age, face shape, hair, wardrobe, distinguishing marks, posture, and palette. Paste the sheet into every prompt and change only the action and camera. If the model supports reference images, use the same keyframe as the anchor. Keep the light direction consistent too, because a face can look like a different person when the light moves.

Why does my video look wobbly?

Usually because the prompt contains competing motion: a moving camera, a moving subject, and moving background elements all at once. Keep one primary motion, simplify the environment, and shorten the clip. If the subject is small in frame, move closer before adding camera movement.

Should I use the same prompt across different models?

Keep the five layers, but adjust the vocabulary. Some models respond better to lens and lighting language. Others respond to plain scene description. Test two versions and keep the one that preserves the subject. A prompt is not a universal file format; it is a set of instructions tuned to a tool.

How many variations should I generate?

Six to ten prompt variations are enough to find a direction. Generate two or three visual tests, choose one, and refine that. Endless variation usually hides a weak brief. If every variation fails, the problem is probably the subject, the light, or the action, not the wording.

What is a good first prompt for a product video?

Start with the product in a clean environment, one light source, one camera angle, and one motion. For example: a matte steel watch on a dark stone surface, medium close-up, soft window light from camera left, charcoal and silver palette, slow push in, four seconds, no text, 16:9. Test the still first, then add motion. If the product has reflective edges, simplify the background before increasing the light contrast.

How do I avoid text artifacts?

Do not ask the model to render important text. Generate the scene with clean negative space where the text will go, then add typography in an editor. If a sign or label must appear, keep it short, avoid small lettering, and expect to fix it in post. Text is one of the fastest ways to expose a weak render.

ChatGPT will not make the film for you. It will help you ask for the frame you actually want. Start with one scene, write the five layers, build a shot card, and test a short motion. Keep what works, log what fails, and refine one variable at a time. When you are ready to see the idea move, take it into the AI video generator and give the model a clear subject, a single action, and a reason for the camera. That is how a prompt becomes a shot, and a shot becomes a story.