Controlled test specification · Adaptable guidance
AI Dunk Prompt for Photo-to-Video Basketball Clips
Copy an adaptable AI dunk video prompt for an image-to-video model, understand what this basketball dunk AI prompt can and cannot control, and compare it with AI Dunk’s guided preset workflow.
Copy the 8-second photo prompt
Attach one clear photo of an authorized adult through a reference image, appearance image, first-frame, or image-to-video input. The prompt describes one solo dunk in a continuous vertical shot.
Use the adult in my uploaded photo as the featured basketball player. Keep their recognizable face, hairstyle, body proportions, and visible clothing consistent throughout.
Create a fictional 8-second vertical 9:16 basketball video. Show one continuous run-up, gather, leap, and one-handed finish at a clearly visible hoop. Follow the player with a smooth medium tracking camera and keep the ball, hands, rim, and landing readable.
Use natural basketball movement and coherent court lighting. Avoid sudden scene changes, duplicate balls, extra fingers, changing clothes, unreadable sponsor text, real team logos, captions, and the likeness of any public figure.Test status: one controlled APIMart Wan 3.0 output was completed on October 27, 2026. The text is free to copy, but cost, duration, watermarks, audio, and input behavior depend on the selected video service. See the input and actual output.
A compatible model interprets the scene and action; results vary.
The public form uses a play, photo roles, court, and quality instead of a written prompt.
02 · Controlled test record
What happened in the controlled AI dunk prompt test?
This AI dunk prompt was submitted once through AI Dunk’s APIMart adapter on October 27, 2026. It used the exact eight-second prompt above and the AI-generated fictional adult chef portrait shown here. The request used 9:16, Standard 480p, no reference video, and model-generated audio turned off. The video below is the actual downloaded result from that task.
| Video service | APIMart Wan 3.0 official model page · model wan3.0-video |
|---|---|
| Input and prompt | One fictional chef portrait plus the exact eight-second prompt shown above; no reference video |
| Output settings | 9:16 · 8 seconds · 480p · model-generated audio off · one submission |
| Provider task | task_01M5YM3AZPHMTE7XHF6M44446Z · submitted 12:06:59 UTC · completed 12:09:12 UTC |
| Provider charge | 2.6304 APIMart credits, reported as $0.26304 for this run. These are provider billing units, not AI Dunk product credits. |
| Public result status | Completed and accepted for this single test. This does not establish a success rate, cross-model compatibility, or a “best prompt” claim. |
What we observed
- Identity: one chef remains visible throughout; the face, mustache, white jacket, and black trousers stay recognizable.
- Action: the continuous shot shows a run-up, gather, jump, one-handed rim finish, ball drop, and landing.
- Frame: the ball and hoop remain readable, with no extra person, duplicate ball, visible brand, caption, or watermark.
- Limits: one accepted result cannot prove repeatability. Hand, ball, rim, face, clothing, and court details can change in another generation.
The task completed in about two minutes and fourteen seconds, but that single elapsed time is not a generation-speed promise. The six-second and reference-video prompt variants were not tested.
03 · A shorter experiment
Try one compact dunk move in six seconds
If the external tool does not offer eight seconds, reduce the job to one continuous approach, gather, jump, rim finish, and landing. This shorter version removes extra action so the model has fewer moving parts to interpret.
Use the adult in my uploaded photo as the only featured basketball player. Keep their recognizable face and visible clothing consistent.
Create one continuous 6-second vertical 9:16 shot on a fictional basketball court. The player takes two quick steps, gathers one clearly visible ball, jumps, and finishes at the rim before landing in frame. Use one smooth tracking camera. Avoid cuts, duplicate balls, extra limbs, logos, captions, and public-figure likenesses.Not separately tested. Reducing the number of actions may make the request easier to interpret, but it does not guarantee a more stable result. Check the service’s duration and reference-image behavior before spending credits.
04 · Start with your photo
How do you use an AI dunk prompt?
- 01
Choose a compatible input mode
The documented route is APIMart Wan 3.0. Other services may call the image input a reference image, appearance image, first frame, or image-to-video input. A text-only mode receives no photo.
- 02
Attach one clear, authorized photo
Choose one adult with a visible face and body. Keep the image attached as an actual input; writing “my uploaded photo” cannot provide an image by itself.
- 03
Paste the prompt and set the format
Select vertical 9:16 and the requested duration when available. Do not assume every model supports the same resolution, length, audio, or reference controls.
- 04
Review the whole output
Check the face, clothing, hands, ball, rim, landing, logos, and captions. Another render may differ and may cost credits in the external tool.
AI Dunk’s own public form uses guided choices rather than a written prompt: play, photo roles, court, and quality.
05 · What the inputs control
Why can two AI dunk prompts produce different videos?
A reference photo can guide appearance while text describes action and camera. Without a video reference, the model must invent timing, body contact, ball motion, hoop geometry, and transitions. More detail can narrow the request, but it cannot lock every frame.
The guided product route separates those decisions into explicit controls. Read the AI dunk video tutorial for the photo, play, court, and quality workflow.
| Input | Photo + text elsewhere | Guided creator on this site |
|---|---|---|
| Featured people | One attached appearance reference unless the tool supports more | One dunker, or a dunker and defender depending on the play |
| Action and camera | Described in words and interpreted by the model | A selected play defines the intended movement and duration |
| Setting | Described in words | Auto or a named court option |
| Quality | Depends on the external tool | Standard 480p or HD 720p |
| Written prompt | Required for this route | Not required in the public form |
| Reference video | Not used in the documented photo + text setup | Not uploaded by the user |
| Duration | Requested in text and settings; support depends on the service | Fixed at 8 or 12 seconds by the selected play |
| Sound | Depends on the external model; the documented setup turns model-generated audio off | Authorized template audio is included in the final MP4 |
| Cost evidence | External service pricing applies | 50 credits for 480p or 100 credits for 720p |
| Evidence | One controlled APIMart Wan 3.0 input-to-output result, with its exact prompt and settings shown above | Accepted product samples and the current product contract |
06 · If your tool supports two references
What if you can upload a reference video too?
Some tools can accept a video for motion and camera plus a photo for the lead subject’s appearance. Confirm that the selected model supports both inputs together. A video upload alone does not mean the model can replace its subject.
Use the uploaded video only as the action and camera reference, and the uploaded photo as the appearance reference for the lead basketball player. Replace only the lead with the authorized adult from the photo. Keep their face, hairstyle, body proportions, and visible clothing consistent during the run-up, jump, and landing.
Follow the reference video's order of movement and camera framing without copying logos, captions, audio, or identifiable people. Keep the ball visible, preserve believable hoop geometry, and generate new court ambience. Results may vary.What the action reference contributes
A suitable reference can show the order of the approach, gather, jump, finish, and landing, plus the camera distance. The prompt clarifies which subject to replace and which visible features to preserve.
Not tested in this review. Use only a performance video you are authorized to upload. This is adaptable wording, not confirmation that any particular service supports appearance replacement or that a public trend video may be downloaded and reused.
07 · Judge the whole clip
What should you check before another attempt?
Review the full result before changing this AI dunk prompt. Identify one visible problem, then make the smallest relevant adjustment. These changes may help a later attempt, but they do not guarantee a fix.
The face or outfit changes
Try reducing turns and requests that hide the face. Remove conflicting clothing descriptions. These changes cannot lock identity through occlusion.
The hands or ball deform
Try reducing simultaneous action and requesting one visible ball. Dense hand-ball-body contact can still create artifacts.
The camera cuts or jumps
Try reinforcing “one continuous shot” and remove extra angle or montage instructions. The model may still introduce a transition.
The hoop or finish is unclear
Try adding “keep the rim visible throughout the approach and finish.” Another render is not guaranteed to correct the framing.
Badges or sponsor text appear
Try removing real team, venue, and branded uniform descriptions. Do not publish an output with unwanted marks or misleading text.
The second person is wrong
Return to the solo prompt or use a model that explicitly supports multiple reference people. Do not infer multi-person identity control from a generic image-to-video input.
08 · Direct answers
AI dunk prompt questions
What is an AI dunk prompt?
It is written guidance for an AI video model covering the subject, basketball action, camera, format, and unwanted artifacts. Attach a photo separately when the person’s appearance matters.
Is this an AI dunk video prompt?
Yes. The main version is an eight-second photo-to-video prompt for one fictional basketball dunk shot. It is adaptable guidance, not a promise of identical results.
Which tool was this prompt tested with?
It was tested once with APIMart Wan 3.0 using one AI-generated fictional chef portrait, the exact eight-second prompt on this page, 9:16, 480p, no reference video, and model-generated audio off. The actual output is shown here; one result does not establish repeatability or cross-model compatibility.
Can I use it in another model?
You can adapt the wording when the selected model accepts both a reference image and text. Compatibility, duration, resolution, audio, cost, and reference behavior must be checked in that service.
Do I need to attach a photo?
Yes when you want the output to resemble a particular authorized adult. Writing ‘my uploaded photo’ does not provide an image unless the file is attached through a supported reference-image input.
Does the prompt generate sound?
The text does not guarantee sound. Audio behavior depends on the external model and its settings. The documented Wan 3.0 setup keeps model-generated audio off so the visual result can be reviewed separately.
Is copying the prompt free?
Yes. Copying the text is free, but an external service may charge for generation, retries, higher resolution, or longer duration.
Can it preserve my face exactly?
No exact identity match is promised. The photo, reference mode, turns, occlusion, body contact, and video model can all change facial or clothing details.
Can I use a reference video?
Only if the chosen service supports photo and video references together and you are authorized to upload both files. A video upload alone does not prove that subject replacement is supported.
What should I change when hands or the ball look wrong?
Reduce simultaneous motion, keep one ball and the rim visible, and simplify the action. These are input adjustments to try, not guaranteed fixes.
About this page
What is verified here?
The action reference verifies the target solo-dunk sequence. The separate controlled test verifies one actual input, exact prompt, APIMart Wan 3.0 task, and output. The downloaded result passed the visible review described above. It does not prove repeatability, compatibility with another model, a success rate, or a “best prompt.”
Only upload photos or videos you are authorized to use. The action clip is product reference material and does not establish an input-to-output relationship with any uploaded portrait.
Prefer guided controls?
Choose a play.
Skip the prompt.
Use one or two authorized photos, select the intended dunk play, choose the court and quality, then review the setup before submitting.
Open the AI Dunk generator