AI creatives for affiliate marketing and media buying

AI models for static and video creatives

How to make affiliate creatives: from offer to test

1. 1

   Break the offer into angles

   Write down a few angles: pain, benefit, social proof. Each one needs its own first frame and its own headline, so the test can show which angle the audience responds to.

2. 2

   Build the base in Photo

   Upload a photo of the product, the package or the app screen as a reference and describe the scene. Seedream 4.5 is the default; when many shots have to meet in one frame, take GPT Image 2 or Nano Banana 2.

3. 3

   Make variations

   Each run returns 1 to 4 images. Send a good one back to the references with Use as a reference for photo and change one element at a time: background, hero, color or headline. Otherwise the test will not tell you what made the difference.

4. 4

   Animate the strongest frames

   The To video button on a card opens the Video section with the image already set as the first frame. Describe the motion and pick a model: Kling 3.0, Seedance, Veo 3.1, WAN 3.0 or MiniMax H3.

5. 5

   Fit the aspect ratio and file size

   Feed, Stories and horizontal placements ask for different proportions, and platforms cap the file size. The Compress button on a card shrinks an image or a clip to the size you need.

6. 6

   Run the test and keep the winners

   Upload the batch to your ad account, compare the metrics and keep what works. Build new variants from the winner: same offer and targeting, fresh first frame.

Angles for one offer

Pain shows the problem the product solves: a sore back after a long shift, orders piling up in a messenger, the fear that English will never click. Benefit shows the result: what a day with the product looks like and what the person gets. Social proof leans on other people: a first-person UGC-style pitch, an unboxing, a real review laid out as a card.

The same scheme fits e-commerce, nutra, apps, online schools and services; only the scenes change. For a product that means hands, the package and the place where it is used; for an app, the phone screen at the moment the task is done; for an online school, a student and a first result; for a service, a pro at work.

Make both a static image and a short clip for every angle: then you can tell whether the angle or the format is doing the work.

Which model to use for static creatives

A product banner in a new scene: Seedream 4.5 and Seedream 4 hold up to 8 references and output 2K or 4K. A collage from many photos: GPT Image 2 and 2.5 accept up to 16 references, Nano Banana 2 up to 14.

If the offer headline is the main thing in the image, try GPT Image 1.5 or GPT-4o Image: both render text accurately, but only in 1:1, 2:3 and 3:2. To sketch an idea without photos, use Z-Image: it draws from text only.

Extra resolution helps when one frame will later be cropped into several formats. 4K comes from Nano Banana 2 and Pro, GPT Image 2 and 2.5, Kling O3 Image, Seedream 4.5 and Seedream 4.

A batch of clips in one go

In the Video section, next to the run button, there is Queue: it sends several different prompts at once, while the model, aspect ratio, duration and references stay shared. That lets you test different openings of the same clip in one go: the product in close-up, a person asking the camera a question, the finished result.

To keep the same person in every variant, create them in the Characters section and pick them in Photo and Video. Then the test compares angles, not faces.

Which model to use for video creatives

A product clip from a single frame: Veo 3.1 always renders 8 seconds, Kling 3.0 3 to 15, MiniMax H3 5 to 15. A half-minute pitch: Seedance 2.5 and WAN 3.0, up to 30 seconds per run. To end the clip on the package or an offer screen, set a last frame. Kling 3.0, Kling 3.0 Omni, Seedance and Veo 3.1 Fast and Quality support it.

When the product, the person and the location must match your references, attach them: up to 30 images on Seedance 2.5, up to 10 on WAN 3.0, up to 9 on Seedance 2 and MiniMax H3, up to 7 on Kling 3.0 Omni. Kling 3.0 and 3.0 Omni, Seedance 2 and Veo 3.1 Fast and Quality go up to 4K.

Kling 3.0 and Kling 3.0 Omni generate sound with a switch, and WAN 3.0 has it on from the start. For a UGC clip with speech, make the voice in Voiceover: the To Lipsync button sends it to Lip sync, where the lips follow the speech. A background track from Music goes into the video composer as a sound track with the To video button; Seedance and WAN accept an audio reference.

Aspect ratios by placement

• 9:16 for TikTok, Reels, Shorts and Stories: every video model has it, and so does every image model except GPT Image 1.5 and GPT-4o Image.
• 1:1 for the feed: any model on this page makes a square.
• 4:5 for the Instagram and Facebook feed: in static it comes from Nano Banana, GPT Image 2 and GPT Image 2.5. No video model renders 4:5, so build a feed clip in 1:1 or 3:4; 3:4 is on every video model except Kling.
• 16:9 for YouTube and horizontal placements: every model has it except GPT Image 1.5 and GPT-4o Image.
• A custom banner size in pixels works on GPT Image 2 and 2.5, Seedream 4.5 and Seedream 5.0 Lite.

When a creative burns out

Fatigue shows in the metrics: CTR falls and clicks cost more while the setup stays the same. Before you change the offer or the audience, refresh the creative itself.

Send the winner to the references and change the background, the hero, the palette or the headline while keeping the composition. For a clip, rebuild the first frame and run the generation again: viewers decide in the first seconds whether to keep watching.

You can edit your own finished clip in Kling 3.0 Omni or Seedance: upload it and describe what to change. The length stays the same as the source, and Kling 3.0 Omni can keep the original sound.

Platform rules: what to avoid in a creative

• Do not promise what the product cannot deliver: a cure, weight loss by a set date, results without effort.
• Keep other companies' brands, logos and packaging out of the creative unless the rights holder has given permission.
• Do not put celebrities or real people in the frame without their consent. For UGC, use an actor who agreed to it or a fictional character.
• Do not invent reviews, figures or awards: social proof has to rest on real data.
• Do not dress an ad up as news, a system notification or a message from a well-known company.
• If a platform asks you to label AI-generated content, add the label. Moderation is always the platform's call, so check its rules before launch.

One balance for every model and the whole team

You do not need a separate subscription for each AI: every model spends tokens from one balance, and the price of a generation is on the button before you run it. If a generation fails or the model rejects the request, the tokens come back on their own.

If several people make creatives, create a team under My team. The team shares one balance, you invite people with a link, and each member can get a spending limit.

Formats and tools in detail

• Ad posters and banners
• Marketplace product photos
• AI UGC video
• Product ad video
• Videos for TikTok and Reels
• Image to video
• AI voiceover
• Lip sync
• AI music generator
• Pricing