1. Confirm the source and generation mode
A text-to-video request asks a system to create the scene and its movement. An image-to-video request starts with a particular photograph, so the image already supplies important visual evidence: composition, architecture, finishes, furnishings, lighting, and view. For an active listing, that approved photo should remain the reference—not a loose description of a similar home.
Runway's current image-to-video prompting guide explains that the image defines the visual starting point and the text prompt should focus on what happens over time, especially camera and subject motion. Interfaces and models differ, but that separation is a useful property-marketing discipline: source image for facts; prompt for controlled movement.
| Desired result | Best starting path | Important limit |
|---|---|---|
| Motion derived from an approved listing photo | Image-to-video with the original photo | Generated frames can still alter the property |
| True path from one room into another | Continuous captured walkthrough footage | One still cannot prove space beyond its frame |
| Simple movement with maximum factual control | Conventional pan, zoom, or parallax edit | Depth separation must still avoid false geometry |
| Imagined design or future concept | Separately labeled visualization workflow | Do not present it as current property evidence |
The video and virtual-tour comparison helps choose the medium when the buyer's question is about layout or continuous navigation rather than atmosphere and attention.
2. Create a source record before prompting
Keep creative instructions and factual evidence separate. The prompt can change many times; the approved source, permitted use, room identity, and review standard should not. Create one record per source image before uploading it to a generation service.
Property and listing ID: [address / internal ID]
Source image: [original filename / photographer / capture date]
Room or feature: [approved plain-language identification]
Rights: [editing / AI processing / video / social / web / MLS]
Source status: [original / retouched / virtually staged / aerial / other]
Must remain fixed: [architecture / fixtures / finishes / view / object positions]
Allowed motion: [camera move / supported environmental motion]
Forbidden inference: [unseen space / boundary / view / condition / inclusion]
Destination and crop: [horizontal / vertical / clean / branded]
Reviewer: [agent / broker / seller / editor as appropriate]
Confirm that the controlling media agreement permits the planned processing. The listing photography usage-rights checklist separates possession of a file from authority to edit, upload, distribute, and reuse it.
3. Write the motion prompt with a small formula
Start with the minimum direction needed to produce a useful shot. A short prompt makes the creative decision visible and the next iteration easier to diagnose. Add detail only to fix an observed problem.
Camera: The camera [pushes in / pulls back / pans / slides / tilts] [direction] at [slow / restrained / steady] pace.
Supported scene motion: [curtains move slightly / water ripples / foliage moves gently], only if visible and appropriate.
Continuity: Preserve the exact [architecture / finishes / fixtures / view / lighting direction / room scale / object placement] in the source.
Exclusions: No [new objects / people / doors / windows / rooms / landscaping / text / weather / structural changes].
Finish: Natural, stable movement with no warping, morphing, flicker, or exposure shift.
Some tools expose a separate negative-prompt control; others expect all direction in one field. Use the interface as documented, but do not mistake an exclusion list for a safety system. A model may still create a plausible-looking detail that is wrong.
4. Adapt the formula room by room
These examples are starting prompts. Replace each bracketed choice with an instruction that fits the actual photograph, then remove anything the source cannot support.
| Source image | Prompt starter | Review closely |
|---|---|---|
| Exterior hero | The camera makes a slow, level push toward the home. Preserve the facade, roofline, windows, driveway, landscaping, neighboring context, sky direction, and visible condition. | Roof edges, windows, property boundary, vegetation, vehicles, sky, neighboring structures |
| Wide living room | The camera slides gently from left to right at eye level. Preserve the exact walls, openings, flooring, fireplace, ceiling, windows, furnishings, and proportions shown. | Doorways, fireplace, built-ins, ceiling fixtures, floor pattern, room width |
| Kitchen | The camera makes a restrained diagonal push toward the island. Keep every cabinet, appliance, counter edge, light fixture, window, outlet, finish, and object in its original position. | Cabinet counts, appliance brands and shapes, handles, counters, outlets, pendants |
| Bedroom | The camera slowly pans across the photographed room without revealing space beyond the frame. Preserve all doors, windows, walls, flooring, furniture, and visible views. | Closet assumptions, added windows, bed scale, exterior view, door destinations |
| Bathroom or detail | The camera makes a subtle push toward the photographed feature. Preserve tile lines, mirror edges, plumbing fixtures, glass, hardware, reflections, and surface condition. | Mirror reflections, doubled fixtures, water, glass edges, grout, surface repairs |
| Patio, pool, or yard | The camera moves forward slowly while only existing foliage and water move gently. Preserve the pool shape, fence, grade, hardscape, landscaping, structures, view, and weather shown. | Fence and lot lines, water edges, terrain, added plants, neighboring property, clouds |
| Approved aerial photo | The camera makes a minimal forward drift within the existing aerial viewpoint. Preserve roads, parcels, roofs, terrain, water, structures, and horizon exactly as photographed. | Boundary implications, invented drone path, removed utility features, changed surroundings |
A source photo does not establish a room-to-room route. Cut between separately approved clips in the property-video storyboard instead of prompting one image to fly through a doorway into an unseen space.
5. Remove requests that invent property evidence
The most cinematic instruction is not always the most useful one. Reject a prompt before generation when it asks the system to supply information the listing record does not contain.
- Do not ask to reveal an adjoining room, closet, view, yard, floor, or amenity that the source image does not show.
- Do not ask for a drone rise, orbit, or property-boundary reveal from a ground-level photo.
- Do not add fire, running water, moving appliances, opened doors, people, pets, or vehicles merely to create action.
- Do not repair surfaces, change finishes, remove permanent features, improve landscaping, replace the sky, or alter visible neighboring context.
- Do not request a wider field of view that requires the model to create pixels beyond the photographed frame.
- Do not describe a neighborhood, school, commute, buyer, lifestyle, or inclusion through generated property imagery.
- Do not animate a virtually staged image until the staged source and the additional generated motion have separate approvals.
Use the virtual-staging video workflow when the source itself is altered. A label may add context, but it does not make an inaccurate permanent feature, view, dimension, or condition acceptable.
6. Change one variable at a time
Save the prompt and result together. If one version warps a doorway, changing the camera, speed, duration, style, crop, and exclusions at once makes it impossible to know what helped. Begin with the simplest useful move, record the failure, and revise one material instruction.
Clip ID and source: [scene / filename]
Generation tool and model: [service / current model or mode]
Settings: [duration / aspect ratio / seed or variation if available]
Prompt version: [exact prompt]
Observed defect: [frame time / changed detail / motion issue]
One change: [instruction added, removed, or simplified]
Result: [reject / revise / candidate / approved]
Reviewer and date: [name / timestamp]
Keep the generated candidate, rejection reason, and approved version. A final clip cannot be responsibly reviewed later if only the last prompt survives.
7. Review the generated clip frame by frame
Watch once for the overall shot, again at reduced speed, and again by scrubbing frame by frame. Compare the beginning, middle, and end with the full-resolution source. Review the crop that will actually be published, because a vertical conversion can hide the reference point that made a movement understandable.
| Review layer | Questions | Action when wrong |
|---|---|---|
| Structure and geometry | Do walls, openings, ceilings, stairs, counters, rooflines, fences, and proportions remain fixed? | Reject the clip; do not conceal the change with a faster cut |
| Condition and finishes | Did flooring, paint, tile, landscaping, damage, light, weather, or visible wear change? | Reject or return to conventional motion |
| Objects and fixtures | Were appliances, lights, handles, furniture, decor, vehicles, or personal items added, removed, or morphed? | Reject and simplify the motion request |
| Views and relationships | Did the clip invent space outside the frame, a room connection, lot boundary, window view, or nearby feature? | Reject; use a cut to another approved source |
| Playback quality | Is there flicker, wobble, exposure pumping, texture crawl, blur, or an unstable ending? | Revise or choose a different source and movement |
Record any AI-assisted motion and resulting review in the AI property-media checklist. The responsible professional must separately check current brokerage, MLS, platform, licensing, advertising, and disclosure requirements for the intended placement.
8. Hand approved clips into the complete edit
Approval belongs to a specific clip, source, crop, prompt version, and destination. Put only approved clips in the edit bin, keep rejected outputs separate, and preserve the source-image name in the scene record. Then judge the assembled sequence for new implied claims: two individually accurate shots can still suggest a false room connection when cut together.
- The opening frame and final frame remain faithful and stable.
- Every scene links to an approved source image and prompt record.
- The sequence does not imply continuous footage, a measured route, or unphotographed space.
- Narration and on-screen text describe only approved facts visible or documented elsewhere.
- Horizontal and vertical crops received separate review.
- The clean, branded, social, and direct versions use the correct clips and identity treatment.
- The final asset has a disclosure, source-access, update, and takedown decision for each destination.
If the manual generation loop is the bottleneck, Westby creates a listing-specific video preview from existing property photos and facts. You still review the property representation, source rights, and publishing rules, but you can evaluate an assembled direction before buying the finished video. Create a free preview from a listing address or link.
Frequently asked questions
What should an AI real estate video prompt include?
For image-to-video work, identify one camera movement, its direction and pace, any limited motion already supported by the scene, and the property details that must remain fixed. Keep the source filename, intended crop, duration, and approval status outside the creative prompt in a production record.
Should an image-to-video prompt describe the room again?
Usually the source image already supplies the room, composition, lighting, and style. Start by describing motion clearly. Add visual language only when the model or task requires it, and do not prompt for an unverified feature simply because it sounds cinematic.
Can a negative prompt guarantee that a property stays accurate?
No. Negative instructions may reduce some unwanted changes, but they do not prove fidelity. Review the complete generated clip frame by frame against the approved source image and reject any version that changes condition, geometry, fixtures, views, scale, or room relationships.
How many motions should one listing-photo prompt request?
Start with one simple camera move and one pace. Multiple camera moves, object actions, lighting changes, and transitions in one request make the result harder to control and review. Build the property story by sequencing short approved clips in the editor.