Marketers know all about hype. And generative AI is in the middle of a feeding frenzy. While many marketers are working with ChatGPT, which debuted in November 2022, or its competitors, generative AI-focused startups are also building custom products designed for marketers to solve more nuanced or marketing-focused use cases. But with so many new entrants popping up in generative AI, the increasingly jam-packed generative AI startup landscape can make it difficult for marketers to distinguish the fakers from the movers and shakers. AdExchanger consulted with agencies about how they are evaluating new generative AI tools for marketing use cases as well as the startups breaking new ground in generative AI. How to evaluate generative AI companies Assuming that marketers decide to give generative AI startups a chance, they need to sort out the answers to some fundamental questions: What model(s) is the startup’s AI model built on top of? While originality is usually prized in a startup, building their own large language model (LLM) isn’t feasible for most startups. They might lack the infrastructure, financial resources, data and expertise to build a LLM from scratch and compete with Big Tech companies, according to Gartner analyst Andrew Frank. That’s why so many build their products on top of existing foundational GPT models from cloud vendors like Microsoft, Google and Amazon. Some agencies prefer to go direct to the LLM rather than work with a generative AI startup with more custom features. Generative AI tools have become “part and parcel” of the creative processes – such as storyboarding, preproduction work and copywriting – for Code and Theory’s art and copy teams, said Dave DiCamillo, CTO of the Stagwell-owned digital creative agency. But Code and Theory mostly uses tools from larger players, like Google’s Bard/Gemini and OpenAI. “I don’t think I’ve seen a good product that we would put into production use from a startup,” DiCamillo said. Still, startups can be valuable to marketers in how they fine-tune prompt engineering; ensure data complies with privacy regulations and is properly formatted for AI-driven ad creative production, targeting and optimization; or provide a higher level of human feedback, Frank said. Startups can also zero in on niche problems that larger companies might not be interested in addressing. AdExchanger Daily Get our editors’ roundup delivered to your inbox every weekday. What data is the model trained on, and how did the startup secure the rights to the training data? Ideally, a startup will be able to confirm it’s licensed the training data it feeds into its AI model(s). Or perhaps it owns the rights to the content it used to create the model. One agency executive mentioned a startup that partners with stock photo providers like Getty Images and Shutterstock to train solely on licensed, royalty-free assets. The startup, Bria.ai, compensates creators when it generates images grounded in creator photos. Marketers should also be careful that their own confidential data doesn’t end up as training fodder, informing future iterations of a generative AI tool. “If you have enterprise agreements with these companies, they will agree to not use your data in their model training,” said Brian Brown, SVP and executive creative director at Publicis digital agency Razorfish. If those agreements aren’t in place, err on the side of caution. Don’t put any client information or names into generative AI tools. Does the startup offer indemnification protection against copyright issues? Some Big Tech providers of LLMs, including OpenAI, Google, Adobe, Anthropic, Getty Images and Microsoft, are offering indemnity shields to their enterprise clients that protect them from liability exposure if someone sues them for copyright infringement. Even if a generative AI company promises indemnification protection, there’s no guarantee it will hold up. Multiple lawsuits against generative AI providers are wending their way through the courts right now, which could take years to resolve. But having a lawyer review a startup’s terms of service for potential copyright issues doesn’t hurt. How much will it cost? Marketers should understand a startup’s cost structure and commercial model, including how their costs will mount as their usage of generative AI tools increase, Gartner’s Frank said. Since many people’s first experience with generative AI was having limitless conversations with ChatGPT for free, they might assume that such tools will remain free or low cost. They would be wrong. What generative AI companies charge for ad creative can vary from two-digit numbers per month for a license to six or seven digits per month, Frank said. “The range is enormous,” he said. “You can tailor the solution to your level of commitment.” Speaking of commitment, startups “look for ways to hook you into their tools,” he said. Organizations should ensure their data is secure and protected – and that they’re not locked into a given startup’s tool if they want to switch down the line. Does the model fit your use case? Different models have different strengths. Some image generation tools, for example, excel at making “cinematic-looking images,” but struggle to create images like a medieval fresco or tapestry, Brown said. Others, built on stock photo libraries, shine “if you need a happy businessman leaping down the hallway of their office or somebody buying a home for the first time.” By understanding what the model is trained on, and where the community of users has taken these models, marketers can suss out which ones fit their desired use cases. Does it work? Digital marketing agency Mod Op found that a text-to-image AI generator it tested didn’t save people time, according to CTO Tessa Burg. Employees were reallocating the time they would have spent creating strong assets on their own to tweaking the AI-generated images to reach the necessary quality standard. Creating an original asset and then using generative AI to make different versions of the asset proved more effective. Marketers should determine their use cases, goals and how to measure ROI before trying a generative AI tool, Burg said. Get clear on what you want to get out of the tool to avoid getting caught in a “testing spiral.” A few generative AI startups For marketers tasked with choosing between generative AI startups, here’s a very small sampling of the independent names out there (a few others include Copy.ai, Creatopy, ElevenLabs, Freepik, Rembrand and Synthesia as well as older players in the space, like Movable Ink, Persado and Phrasee): Personalized audio ads: AudioStack AudioStack (née Aflorithmic) specializes in generative AI for audio. Marketers use AudioStack to create thousands of variations of personalized audio ads. Audiostack combines copy (either inputted by marketers or generated from a prompt), text-to-speech technology, a variety of synthetic voices (cloned from human voice actors’ voices) and AI mixes of music and sound effects. All ads made in AudioStack are editable, so marketers can update the ad copy, voice, music or file format after a campaign has run, according to co-founder and CEO Timo Kunz. Founded in 2019, AudioStack uses a range of external LLMs to integrate into its audio platform, such as Mistral, OpenAI and Anthropic. The startup works with the “big six” agency holdcos, namely, WPP, Omnicom, Publicis, Havas, IPG and Dentsu. It also works with audio publishers like iHeartMedia and brands like Porsche and Mountain Dew. Audiostack isn’t the only company doing AI-generated voice ads. SiriusXM-owned digital audio ad platform AdsWizz, for example, just released an automated synthetic voice ads tool in AudioGO, its self-serve platform for SMBs. A Swiss army knife for ad creative: Jasper Jasper AI (née Conversation.ai, then Jarvis) is trying to be a one-stop shop for a marketer’s AI needs. Jasper launched in January 2021 as a writing assistant tool, before ChatGPT. In the last year and a half, Jasper has billed itself as a copilot for marketing teams at midsize to large companies, according to VP of Marketing Meghan Keaney Anderson. And last week, Jasper bought AI image creation and editing platform Clipdrop from Stability AI. When a marketer enters a prompt into Jasper like “I want 50 variations of this ad,” it creates those 50 variations through a unique process. First, it automatically chooses a foundational AI model based on what it deems best for that use, like summarization or short-form content. Then the tool pulls in brand-specific information from material supplied by the client, such as creative briefs, product catalogs and style guides. It also brings in campaign performance data from Google Analytics. The startup’s AI engine is built on a cross-section of major LLMs from providers including OpenAI, Google, HuggingFace, Anthropic and Cohere. Jasper counts iHeartMedia, Accenture, Wayfair and Demandwell among its clients. Snackable social clips: Munch Munch turns long-form video content, such as a webinar, interview recording or video podcast, into short-form video clips for social media. Marketers and social media managers upload their long-from video content to Munch’s platform, which its AI analyzes and breaks down into numerous clips that marketers can customize further to suit the brand voice and style. It can then be distributed as shareable clips on social media. Munch aims to streamline the resource-intensive process of video editing and trend research so marketers can focus on creative ideation and strategic planning. That way, marketers can “consistently produce a high volume of content that will engage and grow their audience,” said Oren Kandel, CEO and co-founder of Munch. Creative that feeds on performance data: Omneky Launched in 2020, San Francisco-based omnichannel generative AI ad platform Omneky creates personalized ads for clients including Sony Music, Softbank, Saleen Automotive and Japan Airlines by finding out what elements of a brand’s messaging and visual assets are driving sales. The platform ingests and analyzes past campaign performance and brand data via an API integration with a brand’s advertising platforms. The AI engine can tag different components of ad creative using computer vision technology, according to Founder and CEO Hikari Senju. Omneky then uses that data to generate more effective copy and images for the brand’s target audiences. Once an ad campaign deploys, Omneky uses real-time performance data to keep tweaking the creative for future ads. For example, the tool might determine that a jar of makeup with a brush sticking out is the best-performing creative for a given audience, with the highest number of impressions and clickthrough rates. That’s a more specific target than assuming pictures of makeup result in higher click-through rates. Omneky’s generative AI tools are partly built on OpenAI, Stability AI and open-source models, but it also builds its own tools. Assembler of ad creative in different formats: Pencil Generative AI platform Pencil puts together copy, images and other assets to create multimodal ads that run on various channels, like Meta and TikTok. The Brandtech Group acquired Pencil last year. Pencil is built on top of a variety of LLMs, including OpenAI GPT-4 and Google PaLM2 for text generation and Stability SDXL1.0, Google Imagen3, and Getty for image generation. It’s also “constantly” adding new models and AI governance tools. Since its founding in 2018, Pencil has worked with more than 5,000 SMBs and made more than a million ads, though it did not share any specific client names. “We called the company ‘Pencil’ to emphasize a belief that AI should feel like a tool for creative people and not a replacement,” said CEO and Co-Founder Will Hanschell. Graphic designer-level photo editing for noobs: Photoroom Photoroom, a Paris-based photo-editing startup, offers background removal, background creation and photo modification tools. Last year, for example, the Warner Bros. marketing team used Photoroom’s API technology for its “Barbie” Selfie Generator. The social marketing campaign saw 19 million people used the API to make versions of their own “Barbie” posters, according to Head of Brand Lauren Sudworth. Founded in 2019, Photoroom historically built its platform on its in-house models and used its own data to train the models. It uses a custom algorithm for its OG offering, a background removal feature. But for its generative AI tools that create backgrounds or add shadows to images, it starts from Stable Diffusion and trains and edits a custom model, Sudworth said. Photoroom has a client list spanning verticals including film and entertainment, digital agencies, consumer tech and ecommerce. Warner Bros., Netflix, Lionsgate, Hennessy, Bulgari, Faire, Zomato and Daiso number among its clients. The startup’s direct competitors include Claid.ai and Remove.bg. One way that Photoroom sets itself apart is its API, which allows brands to quickly integrate it into their workflows and edit thousands of images, like for an ecommerce site’s product catalog. The video maker to beat: Runway In speaking with agencies about AI-generated video, many say the space is still nascent. But one name keeps coming up: Runway. Marketers can use Runway’s generative AI tools to explore concepts and styles, create storyboards during preproduction, and turn static assets into videos during post-production. It has a Gen-1 video-to-video tool, Gen-2 text-to-video tool and 3D texture, image and audio generation tools. Established in 2018, Runway later worked with University of Munich researchers in 2021 to create the initial version of Stability AI’s Stable Diffusion text-to-image model. “We build everything in-house, from research to model training to the deployment of products,” said Anastasis Germanidis, co-founder and CTO of Runway. Runway’s customer base includes Microsoft, CBS, Publicis Groupe, New Balance, Wieden+Kennedy and film directors and musical artists. For instance, some scenes in “Everything Everywhere all at Once” were created using Runway, and Paul Trillo created a short film, “Thank You for Not Answering,” using only Gen-2, Germanidis said. Custom ad creative for many channels: Typeface Enterprise generative AI company Typeface creates personalized marketing campaigns across social, web, email and SMS; ecommerce landing pages and product descriptions; and paid media and social advertising. Typeface can also turn long-form content, like whitepapers and e-books, into multiple variants of LinkedIn and Facebook posts tailored for different audiences. To generate this custom content, Typeface uses its Blend AI model, which trains on client-provided assets, in combination with OpenAI, Microsoft, Google Cloud and other open-source platforms. The startup works with “numerous digital-native and Fortune 500 brands” across retail, CPG, fashion and apparel, financial services, and automotive, said marketing director Arushi Jain. For example, LG Electronics uses Typface to customize its landing pages, web banners and product catalogs for different markets. Since its debut a year ago, Typeface has pushed into multimodal content, such as photos, videos and audio. The company just snapped up AI startup TensorTour in January, citing the company’s AI algorithms, computer vision technologies and expertise in multimedia AI content.
This content was originally published here.
