GUIDE / AI VIDEO PRODUCTION
How to make an advertising video with AI
From the first product photo to a finished ad: a practical process, sample prompts, a model comparison and a way to calculate costs.
01 / BEFORE YOU START
AI does not replace the process. It speeds up specific stages.
A good ad still needs an idea, a shot list, a consistent product and deliberate editing. A generator is a production tool: the more clearly you describe the result and prepare the source material, the less random each attempt will be.
Prompt
An instruction describing the scene, camera movement, lighting, pacing and constraints.
Text-to-video
Video generated from text alone. It offers freedom, but less control over the product.
Image-to-video
Animating a prepared image. The best starting point for a product ad.
Video-to-video
Changing existing movement or style while preserving the dynamics of the source footage.
02 / BRIEF AND SHOT PLAN
Start by answering six questions.
- ProductWhat exactly are we advertising?
- AudienceWho are we speaking to?
- PromiseWhat should the viewer remember?
- Format9:16, 1:1 or 16:9?
- DurationHow many seconds do we actually need?
- ActionWhat should the viewer do after watching?
SAMPLE BRIEF
Artisan coffee
- Goal
- show the product’s freshness and character
- Format
- 30 s / 9:16 / social media
- Tone
- warm, calm, premium
- CTA
- discover the coffee and order online
FILM PLAN
Six shots are enough to tell a story.
Guide version18.09.2026
| Time | Frame | Movement | Purpose |
|---|---|---|---|
| 0–3 s | Close-up of the cup and rising steam | Slow dolly-in | Capture attention |
| 3–8 s | Packaging beside the brewed coffee | Subtle camera arc | Introduce the product |
| 8–13 s | Beans and a clearly readable label | Light moving across the surface | Detail and quality |
| 13–20 s | Wide shot of the kitchen | Gentle tracking shot | Context of use |
| 20–25 s | Return to the cup | Push-in towards the steam | Emotion and appetite |
| 25–30 s | Product, logo and address | Stable final frame | Brand recall |
03 / SOURCE MATERIAL
Three types of image, three different roles.
Shows what the product or character should look like.
The model uses it to identify shape, colour, material and brand character. It does not have to be the first frame. Prepare a sharp, neutral photo without unnecessary background elements.
- consistent packaging
- colour and texture
- logo and proportions
The first frame of the generated shot.
The start frame’s composition affects the whole movement. Leave room in the direction the camera or subject will move, and avoid elements likely to distort.
- sets the composition
- controls the starting point
- reduces randomness
Defines where the shot should end.
Useful for transformations, changes in framing or a precise end to a movement. Both frames must belong to the same visual world: similar lighting and compatible perspective produce better interpolation.
- controls the end of the movement
- helps with transformations
- makes scenes easier to connect
CASE STUDY / TADOS FIREK
Turning photos of a housing development into moving advertising shots.
The starting point was a folder of photographs showing the buildings, surroundings and architectural details. I planned the virtual camera movement, wrote prompts for MiniMax H3, selected the best generations and edited them so that successive shots maintained a calm, even pace.
TOOLS AND MATERIALS
source materialProject photographs
planningChatGPT
generationMiniMax H3
editingPremiere Pro- Input
- folder of development photos
- Mode
- reference photo
- Generation
- clips with extra editing handles
- Result
- selected, trimmed shots
The photographs established how the buildings looked. The key work was deciding how to show them and treating the generations as raw footage, with selection, pacing adjustments and deliberate cuts.
PRODUCTION WORKFLOW
From a still photograph to a clip ready for editing.
- 01
Choosing the frame
Assessing the perspective, visible detail and available room for camera movement.
- 02
Planning the movement
A description of the intended movement in Polish, turned into a precise prompt using cinematography terminology.
- 03
Generation
The same photograph is sent to MiniMax H3 as a reference photo, alongside the prompt.
- 04
Selecting the middle section
Removing the camera’s acceleration and deceleration leaves a section with a more even pace.
- 05
Editing
Adjusting speed, rhythm and shot order, while checking architectural consistency.
THE LANGUAGE OF MOVEMENT
Make the cinematography decision first, then write the prompt.
Dolly-in
The camera physically moves closer to the building. This is not a lens zoom.
Lateral tracking
Movement parallel to the façade, useful for showing the rhythm of windows and balconies.
Vertical movement
The whole camera rises. This should not be confused with simply tilting upwards.
Gentle arc
A small move around a corner instead of a full orbit that would require unseen visual information.
EDITING THE PACE
Ten seconds of generation does not mean ten seconds in the ad.
These values illustrate the process; they are not a fixed recipe. I first choose a section with good movement, then adjust its speed.
Setting up the conversation for a specific model
I will use MiniMax H3 to create advertising shots from photographs of architectural projects. First, find current documentation and guides on writing prompts for this model. Check how to describe camera movement, pacing, shot duration and the use of reference photos. Cite your sources. Do not confuse H3 with Hailuo or H3 Max. I will then upload photos and describe the intended movement in English. Use these to prepare ready-to-use prompts in English. The image will be used as a reference photo, not a first frame. Do not generate a video; provide only the text and settings.
01Reference photo, not first framewhy I chose this mode
The generator received exactly the same photo that ChatGPT had analysed. It established the appearance of the building and its surroundings, rather than instructing the model to make the first frame identical to the file.
I would choose first frame when the project requires an exact opening image. Here, the reference for the building and the freedom to create movement mattered more.
02Specify a constant speed separately“smooth” does not always mean “constant”
The model often accelerated the camera at the start and slowed it down at the end, so I added this to the prompts: The camera moves at a steady speed throughout the shot. It is already in motion at the beginning and continues moving at the same pace through the final frame.
The instruction alone did not guarantee the result, so selecting the best section remained part of the edit.
03Trim before speeding upthe order of adjustments matters
Setting the speed to 150% also speeds up the acceleration and deceleration. I therefore trimmed the variable-speed sections first, then matched the middle section’s length and speed to the surrounding shots.
When the speed also changes in the middle, Time Remapping is an option. For a small adjustment, though, I start with the simpler Speed/Duration control and check the result in the full sequence.
04Checking architectural accuracyappealing movement is not enough
I compared the generated clip side by side with the source photo, checking the number and arrangement of windows, balcony widths, corner shapes, façade materials and background details.
When advertising a real project, we animate the way it is shown, not random changes to the building itself. An error at the end does not necessarily rule out a usable section at the beginning or in the middle.
05Interpolation after changing speedchecking edges and details
After speeding up the clip, I checked window frames, railings, the roof and foreground objects. Frame Sampling is a good first test; Frame Blending and Optical Flow can help, but they can also introduce artefacts.
I did not apply the same speed percentage to every shot. The deciding factor was how the rhythm of the whole ad felt.
04 / CHOOSING A MODEL
Choose the model for the shot, not the shot for the model.
One project can use several generators. The choice depends on the source material, intended movement, scene length, audio and budget for iterations.
Data updated18.09.2026
| Model | Best starting point | Strength | Check before generating |
|---|---|---|---|
| MiniMax H3 Max TurboPRICE / CAPABILITIES | text or first frame | a very low-cost starting point for repeated advertising tests | platform, resolution, prompt expansion and promotion end date |
| MiniMax H3 Max | text, image or references | better instruction following; 480p, 768p and, through fal, 1080p | this is a cloud offering; do not promise a local Max installation |
| MiniMax H3 | text, first and last frames, references | stereo, 4–15 s clips and open base weights for 768p | EU licence, hardware, workflow, and the difference between local 768p and 2K through the API |
 Seedance 2.5 | image, video, audio | longer scenes and reference-based generation | duration limit, resolution and feature availability on the platform |
 Kling 3.0 | product or character | controlled movement and shots featuring people | model variant, audio mode, 3–15 s duration |
| Kling O3 | existing video | modifying and extending movement | whether the mode is for editing rather than generation from scratch |
 Veo 3.1 | image + description + audio | cinematic visuals and generated audio | Fast/Lite version, duration, resolution and reference support |
| Runway Gen-4.5 | text or image | precise camera control | in-app and API costs, and available controls |
| Gen-4 Turbo | prepared image | fast, lower-cost animations | required input image and 5- or 10-second limit |
Highlighting H3 Max Turbo is a recommendation for low-cost tests based on price and features, not the result of an independent quality comparison of all models.
Visual consistency first
In a product ad, consistent packaging, labels and proportions matter more than spectacular movement.
Natural movement first
For human subjects, check hands, facial expressions, gaze direction and continuity of movement between frames.
Camera control first
For architecture and interiors, choose a model that responds well to specific camera-movement instructions.
MINIMAX H3 LOCALLY
Open weights do not automatically mean a low-cost deployment in Poland.
H3 in ComfyUI
ComfyUI supports H3 from version 0.30.0, with ready-made text-to-video workflows, FL2VA for first and last frames, and Ref2VA for reference material. The base model generates at 768p; the official Regenerate-2K module remains an API service.
Turbo LoRA reduces the number of steps, but does not turn local H3 into the cloud-based H3 Max or H3 Max Turbo.
HARDWARE FOR TESTING
A sensible starting point, not a speed guarantee.
Data updated18.09.2026
| Component | Cautious testing | More frequent use |
|---|---|---|
| GPU | NVIDIA 12 GB VRAM, memory-efficient settings and offloading | NVIDIA 24 GB VRAM with suitable weights |
| RAM | 64 GB | 64–128 GB depending on the workflow |
| Storage | NVMe SSD and 100–150 GB of free space | extra space for models and working files |
| First test | short clip, lower resolution, one generation | 768p and longer shots only after measuring memory use |
A total memory requirement of 42.5 GB does not mean a graphics card with 42.5 GB of VRAM: ComfyUI uses offloading. Generation time must be measured for the specific card, weights and number of steps.
Safely preparing a local H3 test
Prepare a local installation of MiniMax H3 in ComfyUI on this computer. First, detect the operating system, GPU, VRAM, RAM, driver and available disk space. Check the current MiniMax and ComfyUI documentation. Verify the licence terms for a user in Poland. If local use requires separate permission, tell me and do not download any weights until I confirm that I have the necessary rights. Once confirmed, create an isolated installation, choose compatible versions of PyTorch and the libraries, download only the official files needed for one image-to-video workflow, configure suitable quantisation and offloading, and run a short test while measuring time and memory usage. Do not equate local H3 or Turbo LoRA with H3 Max. Do not use a paid API, change drivers or disable security features without my permission.
05 / PROMPTS
A good prompt describes what happens in the scene.
Do not write a list of adjectives. Describe the subject or product, the action, camera movement, lighting, pacing and what the model should avoid, in that order.
- 01What do we see?product, surroundings, framing
- 02What happens?movement of the subject and surroundings
- 03How does the camera move?push-in, orbit, pan, locked-off
- 04What is the lighting like?direction, softness, time of day
- 05What should be avoided?distortions, text, changes to the product
Animating a product shot
A vertical 9:16 advertising shot. A cup of freshly brewed coffee and the product packaging sit on a wooden countertop. Delicate steam rises from the cup. The camera makes a very slow, steady push-in towards the product. Warm, soft side lighting, shallow depth of field and natural reflections on the packaging. Preserve the exact appearance of the label, the packaging proportions and realistic steam physics. Calm pacing, a premium feel, no additional objects and no text.
English prompt for the generator
Vertical 9:16 premium coffee commercial. A cup of freshly brewed coffee and the product package stand on a wooden countertop. Subtle realistic steam rises from the cup. The camera performs a very slow, stable push-in toward the product. Warm soft side lighting, shallow depth of field and natural reflections on the packaging. Preserve the exact label, package proportions and product colors. Calm pacing, cinematic but realistic, no extra objects and no on-screen text.
Common errors to rule out
no distorted packaging, no changed logo, no misspelled text, no duplicated objects, no floating elements, no warped cup, no camera shake, no flicker, no sudden movement, no artificial steam, no oversaturated colors
06 / PROMPT LIBRARY
One method does not fit every model.
Choose a model family and open the instructions for the specific variant. The library brings together guidelines, examples and ready-to-use ChatGPT starter prompts without mixing up features from different models.
- 19
- model-specific guides
- 6
- model families
- 1
- consistent workflow
Distinguish between text-to-video, first-frame animation, reference-based generation and video editing.
Camera commands, negative prompts and audio support depend on the specific variant.
Describe camera movement separately from subject movement, and do not include features that the mode does not support.
MODEL GUIDE
Choose a model from the list
After selecting a model, you will see prompting guidelines, examples and a ready-to-use instruction for starting a conversation with ChatGPT.
07 / PRODUCTION
Generate short shots. Build the film in the edit.
Build a consistent set of frames
Prepare the product photo, start frames and, where needed, end frames. Keep the colour palette, perspective and lighting style consistent.
- check the logo and packaging text
- save frames in the target aspect ratio
- remove elements that are difficult to animate
Test one variable at a time
Test a simple movement first. Only make the camera or surroundings more dynamic once the product remains stable. Keep a record of the model, mode, prompt and attempt number.
- start with shorter generations
- compare 2–4 variants
- do not try to fix everything with one prompt
Shape the pacing, sound and message
Combine the best sections, trim the beginnings and endings, and add sound design, music, voiceover and a closing card with a call to action.
- do not let the length of generated clips dictate the edit
- match colour and contrast between shots
- check the advert with the sound off too
08 / PLATFORMS AND PLANS
Compare the cost of the result, not the subscription.
The figures below come from material prepared on 18 September 2026. Pricing, limits and model availability change, so check the provider's current terms before buying.
Higgsfieldvideo models
Magnificimages and video
Artlistmodels and assets
Runwaygeneration and editing
OpenArtimages and video
Adobe FireflyAdobe generation tools
Data updated18.09.2026
| Platform / plan | Monthly | With annual billing | Credits |
|---|---|---|---|
| Higgsfield | 49 USD ≈ 186.19 PLN | no confirmed figure | 1 000 / month |
| Magnific Premium | 16 EUR ≈ 69.81 PLN | 12 EUR ≈ 52.36 PLN / month | 240 000 / year |
| Magnific Premium+ | 36 EUR ≈ 157.08 PLN | 27 EUR ≈ 117.81 PLN / month | 600 000 / year |
| Artlist Starter | 19.99 EUR ≈ 87.22 PLN | 11.99 EUR ≈ 52.32 PLN / month | 16 500 / month |
| Artlist Core | 39.99 EUR ≈ 174.49 PLN | 23.99 EUR ≈ 104.68 PLN / month | 40 000 / month |
| Runway Pro | 35 USD ≈ 132.99 PLN | 28 USD ≈ 106.39 PLN / month | 2 250 / month |
| OpenArt Plus | 34 USD ≈ 129.19 PLN | 27 USD ≈ 102.59 PLN / month | 12 000 / month |
| Adobe Firefly Pro | 19.99 USD ≈ 75.96 PLN | — | 4 000 / month |
Currency conversions are approximate. “Unlimited” plans usually come with limits on concurrent jobs, slower queues, model restrictions or fair-use rules. Credits from different platforms cannot be compared directly, so this table does not name a single best-value option.
09 / COSTS
Work out the cost per attempt before costing the film.
PLAN CALCULATOR
What do repeated attempts really cost?
Enter your plan's details. The calculator does not add tax or editing costs.The lowest cost in this comparison, with a broad range of model capabilities. The promotion runs until 30 September 2026; this is not a quality ranking of all generators.
Data updated18.09.2026
| Model / mode | Rate per second | 5 s | 10 s | 30 s |
|---|---|---|---|---|
| H3 Max Turbo 768p — fal / promotionPRICE / CAPABILITIES | 0.02 USD | 0.38 PLN | 0.76 PLN | 2.28 PLN |
| H3 Max Turbo 1080p — fal / promotion | 0.04 USD | 0.76 PLN | 1.52 PLN | 4.56 PLN |
| H3 Max 768p — fal / promotion | 0.04 USD | 0.76 PLN | 1.52 PLN | 4.56 PLN |
| H3 768p — API MiniMax | 0.08 USD | 1.52 PLN | 3.04 PLN | 9.12 PLN |
| H3 Max 768p — API MiniMax | 0.08 USD | 1.52 PLN | 3.04 PLN | 9.12 PLN |
| H3 2K — API MiniMax | 0.13 USD | 2.47 PLN | 4.94 PLN | 14.82 PLN |
| Veo 3.1 Lite 720p + audio | 0.05 USD | 0.95 PLN | 1.90 PLN | 5.70 PLN |
| Veo 3.1 Fast 720p | 0.10 USD | 1.90 PLN | 3.80 PLN | 11.40 PLN |
| Veo 3.1 Fast 1080p | 0.12 USD | 2.28 PLN | 4.56 PLN | 13.68 PLN |
| Veo 3.1 1080p | 0.40 USD | 7.60 PLN | 15.20 PLN | 45.60 PLN |
| Runway Gen-4 Turbo | 0.05 USD | 0.95 PLN | 1.90 PLN | 5.70 PLN |
| Runway Gen-4.5 | 0.12 USD | 2.28 PLN | 4.56 PLN | 13.68 PLN |
| Seedance 2 Mini | 0.16 USD | 3.04 PLN | 6.08 PLN | 18.24 PLN |
| Seedance 2 Fast | 0.29 USD | 5.51 PLN | 11.02 PLN | 33.06 PLN |
| Seedance 2.5 | 0.30 USD | 5.70 PLN | 11.40 PLN | 34.20 PLN |
| Wan 3 | 0.10 USD | 1.90 PLN | 3.80 PLN | 11.40 PLN |
| Kling 3 Turbo / promotion | 0.07 USD | 1.33 PLN | 2.66 PLN | 7.98 PLN |
Exchange rates used: USD 1 = PLN 3.7998 and EUR 1 = PLN 4.3633. Thirty seconds means several generations: H3 and H3 Max generate clips of up to 15 seconds. Reference materials may increase the MiniMax API bill. Once the fal promotion ends, the H3 Max Turbo 768p rate is expected to double to approximately PLN 0.76 per 5 seconds.
Higgsfieldgeneration cost with a 1,000-credit / PLN 186.19 plan
Data updated18.09.2026
| Model | Credits / duration | 5 s | 10 s | 30 s |
|---|---|---|---|---|
| Seedance 2.5 720p | 52 / 8 s | 6.05 PLN | 12.10 PLN | 36.31 PLN |
| Seedance 2.5 480p | 24 / 8 s | 2.79 PLN | 5.59 PLN | 16.76 PLN |
| Seedance 2.0 720p | 36 / 8 s | 4.19 PLN | 8.38 PLN | 25.14 PLN |
Magnificexample costs for Premium and Premium+ plans
Data updated18.09.2026
| Model | Credits / duration | 5 s | 10 s | 30 s |
|---|---|---|---|---|
| H3 Max TurboPRICE / CAPABILITIES | 200 / 5 s | 0.47 PLN | 0.94 PLN | 2.83 PLN |
| Kling 2.5 | 140 / 5 s | 0.33 PLN | 0.66 PLN | 1.98 PLN |
| Hailuo 2.3 Fast | 150 / 6 s | 0.29 PLN | 0.59 PLN | 1.77 PLN |
| Seedance 2 Fast | 940 / 4 s | 2.77 PLN | 5.54 PLN | 16.61 PLN |
| Seedance 2 Pro | 1 120 / 4 s | 3.30 PLN | 6.60 PLN | 19.79 PLN |
| Kling 3 Turbo | 780 / 3 s | 3.06 PLN | 6.13 PLN | 18.38 PLN |
| Veo 3.1 4K | 2 080 / 4 s | 6.13 PLN | 12.25 PLN | 36.76 PLN |
Artlistcost based on the 40,000-credit Core plan
Data updated18.09.2026
| Model | Credits / second | 5 s | 10 s | 30 s |
|---|---|---|---|---|
| Kling 2.6 ProPRICE / CAPABILITIES | 150 | 1.96 PLN | 3.93 PLN | 11.78 PLN |
| Seedance 2 Mini | 165 | 2.16 PLN | 4.32 PLN | 12.95 PLN |
| Seedance 2.0 | 300 | 3.93 PLN | 7.85 PLN | 23.55 PLN |
| Seedance 2.5 | 600 | 7.85 PLN | 15.70 PLN | 47.10 PLN |
| Kling 3 Pro | 700 | 9.16 PLN | 18.32 PLN | 54.95 PLN |
| Veo 3.1 Fast 1080p | 280 | 3.66 PLN | 7.33 PLN | 21.98 PLN |
| Veo 3.1 1080p | 500 | 6.54 PLN | 13.08 PLN | 39.25 PLN |
| Veo 3.1 4K | 1 025 | 13.41 PLN | 26.82 PLN | 80.47 PLN |
Runwaythe app and API use separate billing systems
In Runway, in-app costs depend on the plan and number of credits, while the API bills by generated seconds. When comparing options, consider editing tools and queue times as well as the price of a single clip.
The H3 Max Turbo calculation for Magnific assumes an annually billed Premium+ plan and use of the full 600,000-credit allowance. H3 prices are not included for Higgsfield and Artlist because public rates for these variants could not be confirmed in their interfaces.
BUDGET EXAMPLE
Aiming for a finished 30-second advert?
With four attempts per shot, you generate around 144 seconds of footage before selecting the final sections.LOW-COST OPTIONapprox. PLN 27–55shorter tests and lower-cost models
PREMIUM OPTIONapprox. PLN 160–390higher-cost models, audio and higher resolution
10 / QUALITY CONTROL
Most common problems can be spotted before publishing.
The product changes appearance
Simplify the movement, use a clearer reference photo and restate which elements must remain unchanged in the prompt.
The camera movement is jerky
Choose one movement instead of several commands. Add “slow, stable” and trim the shot to the section that stays smooth.
Text is illegible
Do not generate important text within the image. Add it in the edit, and treat labelled packaging as an element that needs a reference.
The scenes do not match
Match the colour temperature, contrast and camera movement. Colour correction also helps make the advert feel consistent.
Credits run out too quickly
Test movement at lower quality and with shorter clips. Generate at full resolution only once the approach has been approved.
The film has that “AI look”
Use fewer effects, keep the physics realistic and leave room for simple, unhurried product shots.
PUBLISHING CHECKLIST
Final checks
0 of 8 completeSUMMARY
Control first. Results follow.
A great AI advert is not made with a single click. It comes from well-prepared frames, short and precise prompts, a considered choice of model and careful selection in the edit. AI broadens what production can achieve; quality still depends on human decisions.
Sources and documentation
Prices and model capabilities may change. Check the latest information directly with the provider before production.
- Runway — video generation
- Artlist — Seedance 2.5
- Google — video best practices
- Magnific — pricing
- Google — reference-based generation
- Google AI — Veo documentation
- OpenAI — search in ChatGPT
- Runway — Gen-4 prompting guide
- Google Cloud — video prompt guide
- Runway API — pricing
- NBP — exchange-rate table A
- Higgsfield — Seedance 2.5 pricing
- Magnific — plans and credits
- Artlist Max — pricing
- Runway — plans
- Gemini API — pricing
- Higgsfield — Kling 3.0 Turbo
- Artlist — how credits are charged
- MiniMax — video generation documentation
- MiniMax — API pricing
- fal — H3 Max Turbo and pricing
- Hugging Face — MiniMax H3 open weights
- ComfyUI — workflow MiniMax H3
- ComfyUI — H3 memory requirements
- Comfy — MiniMax H3 commercial licence
- OpenAI — Codex CLI
- Adobe Premiere — aspect ratios
- Adobe Premiere — Speed/Duration
- Adobe Premiere — Time Remapping
- Adobe Premiere — time interpolation