How Walrus is using AI Image Generation
As AI continues its breakneck evolution, it’s virtually impossible to keep up with the latest releases. There’s a lot of hype, but much of what’s being said about the power of AI seems detached from the reality of what the tools can actually do.We’re asked about this often, so we thought it would be useful to provide an overview of what we’re using, how we’re using it, and where we think the technology is going. We use a wide range of tools at Walrus – too many to include in one post, so we’re breaking it up. In this installment, we’re focusing on static image generation, which is currently the function we utilize most. Apologies in advance for wading into the weeds here, but sometimes you just gotta get waist-deep and swampy.
The Current State of Walrus AI Image Generation
Primarily we use image generators for presentational purposes—comps, storyboards, and decks. This artwork is not making its way into the real world with any regularity. We’re not necessarily against using AI for the finished product (under the right circumstances), but most outputs scream “made with AI” and have enough quirks that we wouldn’t want that art to be public-facing unless we didn’t care if people knew they were looking at an AI image, or if its “AI-ness” was actually part of the concept. Consumers have expressed a strong dislike for AI-generated ads, especially those featuring fake humans. At this point, it’s still easy to discern AI from real, but as AI improves, it will be virtually impossible to differentiate, leaving consumers in the dark about the provenance of the pictures they are seeing. What to do at that point is a quandary for another e-newsletter.Our go-to image generation tool is Midjourney, and most of our comps are now created with it. Midjourney images have a distinctive look—somewhere between an illustration and a photo, with high contrast and a hard-to-put-your-finger-on fantasy vibe, which is fine for our purposes, but is instantly recognizable as AI.
Like all AI models, Midjourney struggles with unusual or highly original ideas. The more novel the concept, the harder it is for the tool to render, because there’s no visual precedent in its image library for it to draw from. But if you want something we’ve all seen before, like a cute puppy looking up at the camera, Midjourney can generate endless versions. It’s a good test: if the AI can do a great job with your prompt, it might not be a very original idea.
Character Consistency
Midjourney has been steadily improving its character consistency—i.e., keeping the same person across multiple frames—which is essential for storyboards. Character consistency and other prompts are application-specific and esoteric, so you need to be fluent in the language of Midjourney to get good results. There are also nuances in the way the application weights parts of the prompt. For instance, it prioritizes terms at the beginning of a prompt more heavily than those at the end. So if you want a red car in the scene, you need to mention “red car” early in the prompt if you want it front and center.
In many cases, we still have to do a good amount of Photoshop work on Midjourney images. Often things are just a little off. There can be odd hallucinations— it will add a sixth finger (which has become a meme)—or an incorrectly oriented object, like a sideways toothbrush.
As for alternatives, DALL·E, ChatGPT’s image generation tool, has improved significantly with recent updates. It tends to generate more photographic results but struggles more than Midjourney with highly creative prompts. It prefers making logical, plausible images. It’s also considerably slower, which can be an issue if you need to make a lot of images.
We haven’t toyed with Stable Diffusion yet. It’s got a much steeper learning curve, but for certain use cases it’s the most versatile. Because it’s open source, you can train it to your own specifications – artist and architect Andrew Kudless has tweaked it to his style (which involves surreal architecture combined with fabrics) so he has much more consistency across outputs:
Other tools:
Photoshop has excellent built-in AI tools but it’s not great for from-scratch image creation (nor is Firefly, Adobe’s web-based generator). It only generates a few options and offers very little in terms of adjustability, but for retouching it’s fantastic—especially for background elements or quick object removal. Adobe has been refining these tools for years via content-aware fills and pattern brushes even if they weren’t labeled “AI” until recently.
A Genuine Innovation: Upscaling
One area of need that’s been greatly helped by AI is image upscaling–adding detail to low-resolution image. Sometimes a low 72dpi JPG is the only product shot available and it needs to go on a billboard. What to do!?!
Topaz Gigapixel is the up-rezer from heaven. It can’t fix everything, but it can take something that’s on the edge of usability, and make it something you can work with. It does leave artifacts so you’ll need to do more than set it and forget it. But it’s a game changer.
So are all these tools making us more efficient? It’s not a simple evaluation. AI is faster and cheaper than going to a comp artist to do the same thing. We can now generate professional looking visuals in-house. It makes our work seem extremely polished, and helps bring ideas to life in a vivid way. But it takes more time than outsourcing or doing it by hand.
There are other drawbacks. As we’ve mentioned, the AIs are not great at original visual concepts. Many of our ideas are intentionally unique – we are trying to do things that are new. To make truly fresh images you have to fight the AI’s tendencies, which usually means making the image in parts, which in turn look less slick because you’ve Photoshopped it all together. Ultimately, this means our freshest ideas will be represented by the least polished comps, making them harder to sell.
Another problem arises when the imagery is so good that it appears finished and clients just want to buy the comp. We call this “demo love” and it can cut both ways. You may love the choice of sweater on the main character, and expect to see that on-set – or you may despise it and wonder how any competent creative business could recommend it. Either way, it’s a decision that was made by an AI as an example of what could be, not a rigid guideline to be executed by a costume stylist eight weeks down the road.
We used to use cartoon drawings for comps, which was really the best way to do it because it left the door open to imagination. Everyone lights things perfectly in their own head. AI makes it much easier to see exactly what it could look like finished, whether that’s what it will actually look like or not.
AI is here to stay, however, and it would be unwise to ignore it, lest ye be left in the past.
