Main menu


Companies Including Stitch Fix Are Already Trying DALL-E 2 – TechCrunch

It has only been a few weeks since OpenAI made images produced by DALL-E 2, an extremely powerful AI text-to-image system, commercially available to customers. But despite the current technical limitations and lack of volume licensing, not to mention APIs, some pioneers say they are already testing their systems with a variety of business use cases. .

Stitch Fix, an online service that uses recommendation algorithms to personalize apparel, says it experimented with DALL-E 2 to visualize products based on specific characteristics such as color, fabric, and style. For example, if a Stitch Fix customer asked for “high rise, red, stretch, skinny his jeans” during a pilot, the DALL-E 2 would be tapped to generate an image of that item, which the stylist would then use. I was able to match it with similar jeans. Products in stock at Stitch Fix.

“DALL-E 2 helps reveal the most beneficial features of a product in a visual way, ultimately helping the stylist find the perfect item that matches what the client requested in written feedback. It helps,” a spokesperson told TechCrunch in an email.

DALL-E 2 generation of stitch fix pilots. The prompt was “soft, olive green, great colors, pockets, patterns, cute textures, long, cardigan.” Image credit: Open AI

Of course, the DALL-E 2 has its quirks — some of which put early enterprise users on hold. Data for his Klaviyo e-commerce startup Eric Silberstein, his vice president of science, outlines his mixed impressions of the system as a potential marketing tool in a blog post.

He said that facial expressions in human models generated by DALL-E 2 tended to be inappropriate, tended to be imbalanced in muscles and joints, and the system did not always fully understand instructions. pointing out that there is no When Silberstein asked his DALL-E 2 to create an image of a candle on a wooden table against a gray background, the DALL-E 2 sometimes extinguished the lid of the candle to blend in with the desk or I added an unnatural border around the candle.

DALL-E 2 Eric Silverstein

Silberstein’s experiment using DALL-E 2 for product visualization. Image credit: Open AI

“We weren’t able to use photos of people or people modeled on products as-is,” writes Silberstein. Still, he said he would consider using his DALL-E 2 for tasks like showing a starting point for editing or communicating ideas to graphic his artists. “For stock photos without humans and illustrations without specific branding his guidelines, the DALL E 2, in my non-expert eyes, can reasonably replace the current ‘old way’. ‘ continued Silberstein.

Cosmopolitan editors came to a similar conclusion when they teamed up with digital artist Karen X. Cheng to create a magazine cover using the DALL-E 2. It shows the limits of DALL-E 2 as an art generator.

But AI weirdness can sometimes act as a feature rather than a bug. In the Draw Ketchup campaign, Heinz uses natural language terms such as ‘ketchup’, ‘ketchup art’, ‘fuzzy ketchup’, ‘ketchup in space’, and ‘ketchup renaissance’ to direct his DALL-E 2 to ketchup bottles. generated a series of images. The company invited fans to send in their own prompts, which Heinz curated and shared across his social channels.

Heinz DALL-E 2

A bottle of Heinz “imagined” by DALL-E 2, part of Heinz’s recent advertising campaign. Image credit: Open AI

With AI imagery dominating news and social feeds, we saw a natural opportunity to expand our “Draw Ketchup” campaign. Heinz is rooted in the insight that the word ketchup is synonymous with the word ketchup, and to test this theory in her AI field,” Jacqueline Chao, her senior brand manager at Heinz, said in a press release. said in her release.

Clearly, a DALL-E 2-led campaign works when AI is the subject matter. However, several of his DALL-E 2 business users say they have taken advantage of the system to generate assets that show no clear signs of AI limitations.

Software Engineer Jacob Martin used a DALL-E 2 to create the logo for OctoSQL, an open source project he is developing. For about $30 (about what Fiverr’s logo design costs for his services), he created a cartoon image of an octopus that looks like a human illustration to the naked eye.

“The end result is not ideal, but we are very happy with it,” Martin wrote in a blog post. “As far as DALL-E 2 is concerned, I think it’s still in the ‘first iteration’ stage for most parts and purposes at this point. The main exception is pencil sketches. They’re amazingly good… I think the real breakthrough will happen when the DALL-E 2 is 10x to 100x cheaper and faster. ”


OctoSQL logo generated after a few tries with DALL-E 2. Image credit: Open AI

One DALL-E 2 user — Don McKenzie, head of design at development startup Deephaven — has taken this idea one step further. He has tested applying this system to generate thumbnails for his company’s blog. This is based on the idea that posts with images can get a lot more engagement than posts without images.

“As a small team, mostly engineers, we don’t have the time or budget to request custom artwork for each blog post,” Mackenzie said in a blog post. We spent a lot of time scrolling through relevant but ultimately inappropriate images from stock photo sites, downloaded the not-so-horrible ones, slammed them into front matter, and published them.”

After spending $45 in credit over the weekend, McKenzie says he was able to replace about 100 blog posts with DALL-E 2-generated images. The prompts had to be fine-tuned to get the best results, but McKenzie says it was worth the effort.

“On average, I think it took a few minutes and about 4-5 prompts per blog post to get something I was happy with,” he wrote. “The more money and time we spent on stock images per month, the worse the results.”

For companies that don’t have time to spend brainstorming prompts, there are already startups looking to commercialize DALL-E 2’s asset generation capabilities. Built on top of DALL-E 2, his promises to “deliver high-quality images and illustrations on demand,” but for now it’s free. The customer enters a prompt (e.g., ‘top view of his three goldfish in a bowl’), selects a preferred style (vector art, photorealistic, pencil drawing) to create the image. This image can be cropped and resized. essentially automates prompt engineering. This is an AI concept that embeds task descriptions in text. The idea is to provide detailed instructions to the AI ​​system to ensure that it accomplishes what it is asked to do. In general, prompts such as “Woman having coffee, walking to work, telephoto” yield much more consistent results than “Woman walking.”

This may bode well for future applications. When asked for comment, OpenAI declined to share figures regarding business users of the DALL-E 2. But anecdotally, the demand seems to be there. Unofficial workarounds for DALL-E 2’s lack of APIs have popped up all over his web, put together by developers eager to incorporate this system into their apps, services, websites, and even video games. I’m here.