From: gregisenberg
Creating winning advertisements with AI, particularly ChatGPT 4.0, can significantly increase return on ad spend (ROAS) and business profitability [00:08:00]. This guide outlines effective techniques for prompting AI models like ChatGPT 4.0 to generate compelling visual content for ads [00:17:00].
Leveraging ChatGPT 4.0 for Ad Creation
Jacob Posel, an expert in creating viral ads and visual content with ChatGPT 4.0, shares his insights on producing “banging ads” in about two minutes [01:03:00]. The core benefit of using AI for ad creation is the ability to experiment at scale, testing various demographics and visual elements to identify what converts best [05:17:00].
The Importance of Inspiration and Product Images
Unlike pure image models like Midjourney or Ideogram, ChatGPT 4.0 is a multimodal language model [02:22:00]. This means prompting is less about highly specific text descriptions and more about providing strong inspiration images and product images [02:42:00]. The model has a better understanding of intent and creativity, making visual inputs highly effective [02:37:00].
Case Study: Jones Road Beauty
An example shows how providing an inspiration ad and a product image (a woman displaying the product) allowed ChatGPT 4.0 to create a consistent and reliable ad layout. The prompt was simple: “add this image to the top of the ad with the product image overlapping the bottom of the image and the heading and subheading below” [03:11:00].
This approach enables rapid generation of ads tailored to different audiences (men, female, old, young, diverse colors) by swapping out model images [04:15:00]. With API access, this process could become trivial, creating hundreds of ads in minutes [04:51:00].
When to Use Reference Images (and When Not To)
AI models have a strong understanding of well-known concepts, brands (e.g., Nike, Adidas, Ridge Wallets) [06:17:00].
- No Reference Image Needed: If your product or brand has a significant online presence (millions in ad spend monthly), ChatGPT 4.0 can generate accurate representations without a reference image [06:48:00].
- Reference Image Essential: For newer, smaller, or less-known products, always provide a reference image [06:43:00].
Case Study: Adidas Ad (No Reference)
An Adidas ad demonstrating a female runner in the Egyptian desert with pyramids required no reference image. The prompt utilized detailed prose descriptions for the desired image, emphasizing keywords like “ultra realistic” or “photo realistic” to prevent cartoonish or strangely colored outputs [07:34:00].
A female Runner is captured in mid stride running laterally in the for ground wearing her performance gear...behind stands the um the Egyptian desert with the pyramids...
The prompt also included specific instructions for font type (e.g., “font type exo2 bold italic”) and can include aspect ratios [08:20:00].
Iterative Prompting and Workflow
Sourcing Ad Inspiration
To find inspiration and templates for ads, consider these resources:
- Sora Explore Page: Offers an “explore page” where users can see how others are prompting and view the exact prompts used for images [09:15:00]. This is a powerful tool for ideation and understanding effective prompt structures [09:23:00]. Sora also appears to have less strict policy guidelines and better photo realism than ChatGPT 4.0 for certain outputs [09:54:00].
- Ad Library Tools: Creative OS, ForPlay, and Icon.me (which reads Meta’s ad library) provide tagged, searchable ad inspiration and templates [19:30:00].
Refining Prompts: Start a New Chat
If your initial prompt doesn’t yield the desired results, it’s crucial not to continue adjusting within the same conversation [12:46:00].
[13:01:00] “It’s using the generated image as its reference for the next one, which can have its own inherent flaws.”
Instead, start a new chat and make adjustments to your prompt there [13:10:00]. This ensures the model starts fresh with your refined instructions, preventing degradation of image quality or increasing “hallucinations” [34:42:00].
Case Study: Athletic Greens (AG1)
Using a Rise ad as inspiration and AG1 product imagery, the prompt “recreate this ad for ag1 adjust the background to better fit athletic green Style change the price … to $85 now and then I said generate the image” [13:53:00]. The phrase “generate the image” is important to ensure an image output rather than text describing an ad [14:06:00]. The resulting ad accurately placed products, adjusted pricing, and had fitting copy [14:17:00].
This method of remixing existing successful ads is already prevalent in the marketing industry, making AI an accessible tool for what was previously a manual process involving templates and inspiration scouting [15:26:00].
Case Study: Ridge Wallet (Concept Change)
An existing Black Friday Ridge Wallet ad was used as inspiration to create a Christmas-themed ad for a green Ridge Wallet [16:27:00]. The prompt was simple due to the strong reference image: “Swap it out and change it to a Christmas style ad” [16:46:00]. This demonstrates the ability to quickly adapt existing concepts for different seasons or occasions. This opens up opportunities for highly personalized and timely ads based on specific holidays (even niche ones like National Hot Dog Day) or individual events like birthdays [17:30:00].
Case Study: Rippling (News Cycle Leverage)
A Rippling ad was created to capitalize on a recent “Spy Deal Controversy” [18:26:00]. The prompt aimed for an old Wall Street Journal style newspaper ad with clever copy: “Rippling so good our competitors can’t even resist sneaking a peak” [18:48:00]. This highlights AI’s capability to quickly produce creative, timely ads that leverage current events, which was previously prohibitively expensive and slow [18:32:00].
Live Example: Late Checkout Agency (LCA)
A demonstration of creating an ad for LCA, a design agency for the AI age, illustrates the process:
- Objective: Create a luxury-style ad for LCA, aiming for a “PCH Philip” (Patek Philippe) watch ad aesthetic, not a newspaper style [23:25:00].
- Input: Screenshots of LCA’s logo, typography, client brands (Dropbox, Salesforce), and a vintage Rolex ad as inspiration [21:16:00].
- Initial Prompt:
Note: “generate the image” is key to ensure an image output [29:08:00].Generate a digital ad for LCA (Late Checkout Agency). LCA is a SWAT team design agency only for the most interesting brands on the planet helping them with building AI first tools. It only works with the best companies. I have provided some screenshots from their website for you to use as inspiration in the ad as well as a reference vintage Rolex ad as inspiration. Use this to capture the luxury Vibe of the company. Generate the image. - First Attempt Analysis: The ad generated a man pointing, with a Rolex watch, and copy like “Leaders at Billion Dollar Brands choose LCA to define their future” [31:47:00]. The overall vibe was good, but the Rolex watch was an unexpected “hallucination” from the inspiration image [32:20:00].
- Second Attempt (New Chat): To correct the Rolex, a new chat was started to avoid the model using its own generated image as a flawed reference [34:20:00].
- Adjusted Prompt: Added: “rather than generating a watch show some of the companies LCA has worked with provided in the reference images” [35:53:00].
- Result: The new ad successfully integrated client logos (though it still included Disney, which LCA doesn’t work with, highlighting the need for specific logo instructions [40:09:00]). The man was still present, but now showing his wrist (less jarring than pointing) [39:07:00]. The ad maintained the desired luxury aesthetic with the correct logo and subheading [39:51:00].
This live example illustrates that while AI can produce creative results from vague prompts, further refinement often requires more specific instructions in new chat threads [41:04:00]. For example, to include only specific logos, you would specify them by name [40:51:00].
General Prompting Tips
- Clarity and Directness: Be very clear and direct in what you want [12:36:00].
- Start Fresh: If struggling with the initial prompt, start a new chat rather than adjusting in the same conversation to avoid compounding flaws [12:46:00].
- Expect Iteration: For complex or less-known products, expect 2-3 iterations to get the desired result [31:01:00].
- Specify Dimensions: ChatGPT 4.0 usually defaults to 9x6 aspect ratio, but you can specify others like 1x1 or 9x16 in the prompt [37:54:00].
- Video Content: While direct video generation isn’t the primary focus, generated images can be animated into video content using tools like Runway and Sora [38:41:00].
This approach transforms the process of ad creation, making it faster, more accessible, and highly scalable for marketers.