GPT Image 2.0 Case Study: How Better Character Portraits Lifted Conversion on an AI Companion App

What We Changed

On April 23, 2026, we replaced the four static character portraits used across Tendera (Sophia, Mia, Elena, and Jade) with new ones generated using GPT Image 2.0. The workflow was deliberately low-tech. The prompts that drive each character's image were run through ChatGPT. The resulting PNGs were downloaded. The existing portrait files were swapped out by hand.

Tendera doesn't have a built-in image generator. This wasn't a pipeline change. It was four file uploads.

Everything else stayed the same. The character system prompts that drive each persona's voice and behavior were untouched. The UI didn't change. The chat backend didn't change.

What We Measured

Three days after the swap, two numbers had moved:

• Visitor-to-signup conversion: up approximately 5%

• Visitor-to-chat conversion (counting both guest preview chats and post-signup chats): up approximately 8%

These are two different metrics measuring two different events. We're not stacking them up against each other. They're parallel data points pointing the same direction.

The reason both numbers are reported in this case study is that the second one moving was the part we didn't expect.

What We Assumed Going In

Before the swap, the working assumption was that better character art would mostly help acquisition. Prettier card on the landing page, more clicks, more signups. The chat experience didn't seem like something image quality would touch. By the time a user is sitting in front of a chat input, the visual selling job feels mostly done.

That assumption was wrong, or at least incomplete.

What Actually Changed in the Images

The character topology is identical. Same four characters, same wardrobes. The poses are roughly the same too. What's different is how legible each character is now.

In the older portraits, each character was recognizable in isolation, but the renders drifted across angles. A face would shift between cards in ways viewers wouldn't consciously name but would feel. A slightly different woman in a similar outfit. Hands and small details broke in ways AI image models were known to break circa 2024-25.

GPT Image 2.0's renders look more boring in some ways. Less stylized. The model feels less like it's interpreting the prompt and more like it's executing it. But the character holds across angles. Same person across multiple shots. No drift.

The other thing the new model nails is dimensionality. The old renders were clean but flat. They read as illustrations. The new ones have physical depth. Light hitting the side of a face. A jacket folding the way fabric actually folds. None of it is photoreal. The dimensionality just reads.

Why We Think the Chat Number Moved

Here's the framing that fits the data without overclaiming.

When a visitor lands on the marketing page, the question they're answering is whether the surface signal looks decent enough to click in. Image quality affects this, but a serviceable card will still get the click.

Once they're past the door, sitting in front of an actual character profile or already in chat with the character header in view, the question gets sharper. Now they're evaluating whether this person feels real enough to talk to. The image is the only non-text signal in the room. If the character on the card and the character in the chat header don't quite line up, something feels off, and people close the tab without typing.

Most users wouldn't describe this consciously. We're inferring what their gut is doing. But chat-side conversion moving with prompts and copy unchanged points at the visual layer doing some work past the landing page.

What We Want to Test Next

The obvious next experiment is whether the same model can produce reliable expression variants for the chat header. Right now each character has one default portrait. If the same character could subtly shift expression based on conversation tone (a softer face during something quieter, a smirk during banter), chat-side recognition could go up another step.

That's a harder problem. Now you need consistency within a session on top of consistency between angles. It would also require building image generation into the chat pipeline, which Tendera doesn't currently have.

If we tested it on one character first, it would be Jade, the character users tend to spend the most time with. The voice on her side is already doing most of the work. The image is the one input that hasn't caught up.

Caveats

A few things worth flagging for anyone reading this case study:

• This is three to four days of data on a small app. Effects could compress as the sample grows.

• The change was on the portraits, not the character system prompts. If a similar AI character app has its bottleneck on the writing side (voice, dialogue cadence), regenerating images won't help.

• There was no clean A/B with old vs new portraits served to different cohorts. The whole site flipped over April 23. A pre-existing upward trend coinciding with the swap could absorb part of the lift.

• Visitor-to-signup and visitor-to-chat are different metrics measuring different events. We report both because both moved, not because one is bigger than the other.

• This was a manual asset swap. The new portraits were generated in ChatGPT and uploaded by hand. There's no image-generation pipeline integrated into the product.

What This Might Mean for Other Character Apps

If you're building a product where a user is supposed to form a relationship with a fictional persona (a character app, an NPC system, an AI tutor with an avatar, a virtual host), your image generator might be doing more work than acquisition-side metrics suggest.

Worth regenerating your portraits and watching what moves.

If you'd like to see how the new portraits hold up in actual conversation, you can try Tendera with five preview messages, no signup required.