How to prompt different image and video models

Keep the creative goal consistent while adapting the instructions to the selected model and its inputs. A new image, an image edit and a moving shot need different briefs. This guide explains the decisions first, then shows the model guidance used in OpenModels.

In short

Choose the task and input mode before refining the wording.
For new images, describe the result; for edits, describe the change and what must stay.
For video, describe the sequence and camera, and use audio instructions only where the workflow supports sound.
Select the model before using Enhance prompt, then review the draft and generation quote.

Who can run this: Anyone can read this guide. Generating requires the selected model's account access and credits; the composer shows the available workflow and current quote.

Which plan: Reading and drafting do not require a paid plan. Generation access and pricing depend on the selected model, workflow and account. Some runs use GPU credits; provider-billed runs use Premium model credits. Check the composer before submitting.

Start with the task, then choose the wording

A useful brief answers a practical question: what should the finished asset let someone see? For a catalogue image, the product and its materials might matter most. For an edit, it might be replacing a background without changing the subject. For a video, it might be showing how a mechanism opens. Write that goal in one sentence and use it to decide which details belong in the prompt.

Task	Put in the prompt	Set with the available controls
Create an image	Subject, arrangement, material, light, style and any exact visible text.	Model, aspect ratio and supported output options.
Edit an image	The specific change and the parts of the input to preserve.	The actual source image and the edit-capable workflow.
Create video from text	Opening composition, ordered action, camera and supported audio.	Model, duration and aspect ratio.
Animate a starting image	Movement developing from the supplied composition.	A model with a start-frame input and the attached frame.
Connect two frames	A plausible visible transition between the endpoints.	A model with first/last-frame inputs and both images.
Use subject references	The role of each reference and the new scene to create.	A compatible reference workflow and its supported attachments.

These choices cannot be replaced with magic words. Writing a duration inside an image prompt does not create a video, and asking a reference model to use an exact final frame does not add an endpoint control. Treat the input slots and generation settings as part of the brief. When the workflow changes, review the prompt again instead of carrying every sentence across unchanged.

Build a brief you can evaluate

Write the essential result Describe the subject and purpose in ordinary language. A transparent amber bottle on a light grey background is a more useful starting point than a string of quality adjectives. For video, add the main change: the bottle turns slowly to reveal its label. Keep the deliverable small enough to judge.
Declare the constraints Choose the composition, exact text and details that cannot change. If you are editing, separate the requested change from what must remain. If you are supplying several references, give each one a role so the model does not have to guess which provides the person, object or visual style.
Match the selected model Read its short prompting guide and check the actual controls. Image models do not all take the same reference inputs, and video workflows differ in whether they accept start frames, endpoints, motion sources or audio. Preserve your creative goal while adapting only the instructions that depend on that workflow.
Enhance, inspect and generate Choose the model and inputs before using Enhance prompt. Check that the rewrite keeps your important words and requested changes, then review the quote. Treat enhancement as an editable draft. Decide whether to generate only after the resulting prompt describes the asset you actually want.

Model-specific guides and examples

The following short guides come from the same model guidance shown in OpenModels model cards. Each includes an original example and links to its sources. They are starting points rather than a claim that a model will reproduce every detail perfectly. Read the linked source when you need the complete upstream recommendations, and use the controls exposed in OpenModels for the actual run.

GPT Image 2

Make the layout, wording and preservation requirements explicit.

State the intended image, its subject and the arrangement of important elements.
Quote exact visible text and say where it belongs.
For edits, specify both the change and the identity, branding or layout details to preserve.

Example: Create a clean product photo of a glass water bottle on pale stone. The label reads "Spring 01" in black. Keep the whole bottle visible with space above the cap.

Nano Banana

Use clear scene descriptions and precise editing constraints.

Describe a complete scene with subject, composition, setting and lighting.
For edits, name the change and explicitly preserve the other details.
Put exact visible wording in quotation marks and specify its position in the layout.

Example: Create a square bakery poster with a single croissant on cream paper. Place the exact heading "Fresh at dawn" above it in dark blue, with generous space around the lettering.

Qwen Image Edit

Describe the finished picture without an image, or a precise change with one.

With no reference attached, describe the finished picture: subject, composition, setting and light.
For an edit, image 1 is the source. Identify what changes and preserve the identity, layout, wording and background details that must remain.
Additional images provide reference context; explain each image's role using its attachment order.

Example: Change the mug in image 1 to the blue glaze in image 2. Keep the mug's shape, handle, printed lettering, table and lighting unchanged.

LTX 2.3 text to video

Write a flowing description of one coherent shot.

Use present-tense sentences: subject and setting, then action, camera movement and lighting.
Keep events in the order they should happen and fit them within the selected duration.
Choose a clear camera move or a steady frame; describe sounds only when the selected workflow supports and enables them.

Example: A cyclist coasts into a quiet courtyard and stops beside a fountain. The camera follows at shoulder height, then holds as sunlight passes across the stone wall.

LTX 2.3 image to video

Use the image for appearance and the prompt for change.

The first frame establishes the composition. Describe what starts moving and how it develops.
Keep the subject's identity and the scene consistent with the supplied frame.
Write one continuous action with a deliberate camera movement; avoid redundant descriptions of the still.

Example: The cat lifts its head toward the window and slowly blinks. The camera gently pushes closer while the curtain sways behind it.

LTX 2.3 first and last frames

Explain the physical path connecting the two frames.

Treat the supplied frames as the beginning and end of one continuous transition.
Describe intermediate movement, changes in pose and camera travel that connect those endpoints.
Keep timing plausible and preserve identity, clothing and objects through the transition.

Example: The closed flower gradually opens as the camera arcs slightly to the right, ending on the open petals and composition of the final frame.

Wan 2.2

Direct motion from the supplied still.

The image fixes the subject and scene. Focus on one main action and its visible result.
Describe the camera separately from the subject: a held frame, a slow push or a clear pan.
Keep the sequence short and physically continuous; avoid unnecessary static detail.

Example: The fox turns toward the falling leaf and takes two steps forward. A steady camera holds the woodland composition.

Google Veo

Direct the shot and its sound as separate parts of the scene.

Describe subject, action, environment, framing and camera movement in concrete language.
With first and last frames, describe the journey between them; with Ingredients, explain each reference's role.
When sound is enabled, specify ambience and any exact dialogue separately from visual action.

Example: A ceramic artist places a blue bowl on a wooden shelf. The camera tracks the hands in close-up. With sound enabled: a soft ceramic tap and quiet room ambience.

Improve a draft without losing the idea

Consider this rough request: Make a nice bottle advert. It leaves the product, composition, use and action undecided. A more testable image brief would be: A small amber glass dropper bottle centred on a pale grey surface, with soft window light from the left and enough clear space above for a headline. The bottle has a plain cream label reading AMBER. Show the whole bottle and its shadow. This is our own example, and the important improvement is that the result now has things you can check.

For an edit, the same creative goal becomes a different instruction: Keep the bottle, label and camera angle from the attached image. Replace the patterned background with a pale grey surface and soften the shadow. For a video that begins from that product image, focus on the change: The camera moves slowly closer while a narrow band of light travels across the glass. The bottle remains upright and the label stays facing the viewer. Choose an appropriate starting-image workflow before using that version.

Notice what stayed consistent: the product, clean presentation and readable label. What changed was the information the model needed. The edit prompt did not invent a replacement bottle; the video prompt did not ask for five new scenes. This approach makes it easier to adapt an idea across models while keeping control of what you are trying to produce.

Compare models with a stable brief

When comparing models, keep a copy of the same core brief and note any necessary workflow changes. A text-only request and an edit with a detailed source image are not equivalent inputs. Review the output against the deliverable: subject accuracy, composition, text, continuity or motion. The model that produces the most dramatic image may not be the one that best preserves a product's details.

Revise one important issue at a time. If the output is too busy, simplify the set rather than adding a long list of exclusions. If text is essential, state the exact wording and placement, then check the result at its intended display size. If movement is unclear, describe a shorter sequence with a visible beginning and end. Keep successful details intact so you can learn what the revision actually changed.

Questions about prompting different models

Should I use the same prompt for every model?

Keep the same creative brief when comparing results, but adapt workflow-specific instructions. An image edit needs a clear change and preservation request; a video needs temporal action. Native H3 section labels and reference notation should stay with the compatible H3 workflow.

Does a longer prompt always work better?

No. Add details that affect a result you can evaluate. Repeating quality words or combining several competing compositions makes a brief harder to review. Start with a clear subject, task and constraints, then add detail where a result shows that it is needed.

Why select a model before using Enhance prompt?

The selected model and inputs tell the enhancer which guidance applies. A prompt for a supplied starting frame should focus on movement from that frame, while a prompt for new-image generation must describe the image itself. Check the rewrite after changing models or attachments.

Will opening a tutorial or an example generate an asset?

Reading a guide does not submit a generation. Review the prompt in the composer, check the selected model, inputs and quote, and use Generate when you are ready. Each submitted render is governed by that workflow's current access and credit requirements.

Open the image composer — Opens the studio. Nothing runs, and nothing is charged, until you press Generate.

Last checked 2026-09-20. Prices and plan limits change; the pricing page is the authority.