Design
Last reviewed
May 13, 2026
Sources
45 citations
Review status
Source-backed
Revision
v2 ยท 5,255 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 13, 2026
Sources
45 citations
Review status
Source-backed
Revision
v2 ยท 5,255 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: Design ChatGPT Plugins
Design is one of the creative fields most visibly transformed by artificial intelligence. Between 2022 and 2025, generative models moved from research curiosities to features inside Photoshop, Illustrator, Figma, Canva, Fusion 360, and almost every other professional design application. Designers now use AI for graphic layout, photo retouching, vector recoloring, UI generation, logo creation, 3D asset modeling, mechanical part optimization, building massing studies, and stock imagery. The same period brought sharp controversies: artists discovered their work in training sets without consent, design tools produced output that looked uncomfortably similar to existing apps, and platform terms of service became the subject of viral backlash.
AI in design covers several distinct disciplines that have absorbed machine learning at different speeds. Graphic and image-editing tools led the way because diffusion models map cleanly onto raster image work. UI and UX generators followed, helped by large language models that could produce code alongside layouts. Industrial generative design, which actually predates the recent boom in generative AI by almost a decade, uses topology optimization rather than diffusion. Architecture and engineering tools sit between the two, mixing optimization, parametric rules, and more recently natural language prompting.
A short list of the platforms that matter:
Generative methods entered industrial design well before they reached graphic designers. Autodesk shipped early generative design capability inside Fusion 360 around 2017, the same year its in-house studio The Living completed the bionic partition for the Airbus A320. That partition was the largest metal 3D-printed component flown on a commercial airliner, and it was roughly 45 to 50 percent lighter than the conventional aluminum panel it replaced. The underlying technique was topology optimization, an older mathematical method dressed up in cloud compute, not a neural network.
The consumer-facing wave is much newer. Stable Diffusion was released in August 2022. DALL-E 2 had opened broad access in July 2022. Midjourney was already in open beta. By the end of 2022, Shutterstock had announced its partnership with OpenAI to integrate DALL-E into its stock library, with a launch slated for early 2023. In October 2022, Hollie Mengert, a Disney illustrator, found her style replicated by a Stable Diffusion fine-tune trained on 32 of her illustrations without her permission. The same month, the digital painter Greg Rutkowski became the most-searched artist style on Stable Diffusion, ahead of Picasso and Michelangelo.
March 2023 was the inflection point for the major design platforms. Adobe announced Firefly on March 21, 2023, as a family of generative models intended for commercial use. Microsoft launched the Bing Image Creator, also based on DALL-E. May 2023 brought Photoshop Generative Fill in beta. June 2023 brought Illustrator's Generative Recolor. By October 2023, Canva had launched Magic Studio and Vercel had announced v0.
2024 was the year of integration and embarrassment. Figma rolled out an AI feature called Make Designs at its Config conference in June, then pulled it within weeks after a tweet from Andy Allen showed it producing screens that looked like Apple's Weather app. Adobe's revised Terms of Use sparked a separate uproar in the same month over fears the company would train AI on customer files. Recraft V3 arrived in October 2024 and beat Midjourney, FLUX, and DALL-E on community-voted image leaderboards. Google completed its acquisition of Galileo AI in May 2025, folding it into Google Labs as a tool called Stitch.
The graphic design category is the most crowded and the most directly shaped by image diffusion models. Adobe, Canva, and Microsoft dominate distribution because each had an enormous user base before the AI wave. Newer entrants like Recraft compete on raw model quality and vector output.
| Tool | Year | Company | Key features |
|---|---|---|---|
| Adobe Firefly | March 2023 | Adobe | Text-to-image, generative fill, recolor, video; trained on licensed Adobe Stock and public domain content |
| Photoshop Generative Fill | May 2023 (beta) | Adobe | In-context text-prompted inpainting and outpainting inside Photoshop |
| Illustrator Generative Recolor | June 2023 (beta) | Adobe | Text-prompted recoloring of vector artwork |
| Adobe Express | March 2023 | Adobe | Consumer design app with Firefly integration |
| Canva Magic Studio | October 2023 | Canva | Magic Design, Magic Write, Magic Switch, Magic Animate, Magic Eraser, Magic Media |
| Microsoft Designer | October 2022 (preview), April 2023 (public) | Microsoft | DALL-E-powered design app for social posts, invitations, posters |
| Bing Image Creator | March 2023 | Microsoft | DALL-E text-to-image, later upgraded to DALL-E 3 |
| Recraft V3 | October 2024 | Recraft AI | 20B parameter model, vector output, topped LMArena text-to-image leaderboard |
| Leonardo.Ai | December 2022 | Leonardo.Ai | Fine-tuned diffusion models, game asset focus, acquired by Canva in July 2024 |
| Ideogram | August 2023 | Ideogram AI | Strong typography, posters, logos |
| NightCafe | 2019 | NightCafe Studio | One of the earliest community AI art platforms |
| Civitai | November 2022 | Civitai | Open model and LoRA sharing hub for Stable Diffusion |
Firefly is interesting commercially because Adobe sells it on the basis of indemnification. Adobe states the first model was trained only on licensed Adobe Stock images, openly licensed content, and public domain material. That claim became the foundation of Adobe's commercial-safety pitch and a source of contributor controversy, discussed below in the section on controversies.
Canva's Magic Studio launched on October 4, 2023, marking the company's tenth anniversary. The suite bundled Magic Design (auto-generate full designs from a prompt or uploaded media), Magic Write (a writing assistant introduced in late 2022, now extended with brand voice), Magic Switch (one-click format and language conversion of an entire design), Magic Animate, Magic Eraser, and Magic Media for text-to-image and text-to-video. By 2024, Canva said Magic Studio features had been used more than 5 billion times. The company acquired Leonardo.Ai in July 2024 to bring more advanced model R&D in house.
Recraft V3, codenamed red_panda during testing, was released in October 2024. It is a 20 billion parameter model that ships with both raster and SVG vector output. On the Artificial Analysis text-to-image leaderboard hosted on Hugging Face, Recraft V3 reached an Elo rating around 1172 with a 72 percent win rate, beating Midjourney 6.1, FLUX 1.1 Pro, and DALL-E 3 HD at the time. On the LMArena community leaderboard with more than 40,000 votes, it took the number one spot ahead of Ideogram 2.0 and FLUX 1.1 Pro.
The UI design category has been less stable than graphic design. Several startups changed hands quickly between 2023 and 2025.
| Tool | Year | Company | Key features |
|---|---|---|---|
| Vercel v0 | October 2023 | Vercel | Text-to-React UI generation using shadcn/ui and Tailwind |
| Galileo AI | 2022 (launch), May 2025 (acquired) | Galileo AI, acquired by Google | Text prompt to editable UI designs, rebranded inside Google as Stitch |
| Uizard | 2018, 2024 (acquired) | Uizard, acquired by Miro May 2024 | Sketch-to-design, screenshot-to-design, AI wireframing |
| Visily | 2021 | Visily | Sketch-to-wireframe, screenshot-to-design, text-to-design |
| Magician | 2022 | Diagram | Figma plugin acquired by Figma at Config 2023, sunset in 2024 |
| Figma Make | May 2025 | Figma | Prompt-to-prototype, replaces the paused Make Designs feature |
| Builder.io Visual Copilot | February 2024 | Builder.io | Figma-to-code in React, Vue, Svelte, Angular, Qwik, Solid, HTML |
| Stitch (formerly Galileo) | May 2025 | Gemini-powered UI design tool inside Google Labs |
Vercel announced v0 on October 11, 2023, framing it as the first product of a new category called generative UI. The tool produces React code built on top of shadcn/ui and Tailwind CSS. The waitlist passed 100,000 sign-ups within three weeks. Vercel later integrated v0 into a broader product, v0.app, and used it as a showcase for the company's AI SDK.
Galileo AI, founded in 2022 by Arnaud Benard and Helen Zhou, raised under five million dollars before Google acquired it in May 2025. The product turned text descriptions into editable UI designs, and Google relaunched it as Stitch, powered by Gemini, inside Google Labs the same day the acquisition was announced.
Uizard, the Copenhagen startup behind one of the first sketch-to-wireframe tools, was acquired by Miro in May 2024. Diagram, the studio behind the Magician plugin for Figma, was acquired by Figma at Config 2023 in June. The Magician plugin was sunset as those features rolled into Figma AI.
Figma's own AI journey has been bumpy. Figma announced AI features at Config 2024 in June, including Make Designs, which generated initial UI screens from a prompt. Within two weeks, designer Andy Allen, founder of NotBoring Software, posted on X that prompting Make Designs for a weather app produced screens nearly identical to Apple's iOS Weather app. Figma CEO Dylan Field disabled the feature the next day. In a follow-up blog post, Figma said Make Designs had used off-the-shelf models from OpenAI and Amazon's Titan paired with a custom Figma design system, and that some of the design system assets were too close to specific real apps. The feature returned in modified form, and at Config 2025 Figma announced four new products: Figma Sites, Figma Make, Figma Buzz, and Figma Draw. Figma Make turns prompts and existing files into working prototypes and web apps.
Builder.io launched Visual Copilot in February 2024 as a Figma plugin that converts designs into React, Vue, Svelte, Angular, Qwik, Solid, or HTML code. The first version was trained on more than two million data points and uses an open-source compiler called Mitosis along with a fine-tuned LLM to refine the output for the chosen framework and CSS library. Later releases added integrations with Cursor and Windsurf as well as a CLI for bringing generated code into existing repos.
AI logo makers existed before the diffusion era. They were template-and-rules engines that swapped fonts, colors, and icons based on user inputs. The generative wave layered diffusion outputs on top and gave designers a way to suggest visual directions in plain English.
| Tool | Year founded | Founders / Owner | Notes |
|---|---|---|---|
| Looka (formerly Logojoy) | November 2016 | Dawson Whitfield | Rebranded April 2019; over 20 million users |
| Tailor Brands | September 2014 | Yali Saar, Tom Lahat, Nadav Shatz | Series B 2018 raised $15.5 million |
| Hatchful | 2017 | Shopify | Free template-based logo maker |
| Brandmark | 2018 | Jack Qiao | Generates brand kits, icons, mockups |
| LogoAI | 2018 | LogoAI Inc. | Provides matching social and brand collateral |
| DALL-E and Midjourney | 2022 onward | OpenAI, Midjourney | Used widely for one-off logo experiments |
Looka started in November 2016 as Logojoy, founded by Toronto designer Dawson Whitfield with CTO Andrew B. Martin. Whitfield rebranded the company to Looka in April 2019 to reflect the expansion from logos into a broader brand kit. The company has claimed over 20 million users across more than 180 countries and billions of generated logos.
Tailor Brands was founded in 2014 in New York by Yali Saar, Tom Lahat, and Nadav Shatz. It was one of the first machine-learning logo generators, and it raised a $15.5 million Series B in 2018. The product has expanded over time to include business formation tools, websites, and social content scheduling.
Hatchful is Shopify's free logo maker, aimed at e-commerce founders who need something quick. It guides users through industry, style, and color choices and produces a small set of variations.
Brandmark and LogoAI cover a similar middle ground. Both produce logos, brand kits, social mockups, and color palettes, and both rely on a mix of template logic and generative components. Designers in 2023 and 2024 also began using DALL-E, Midjourney, and Recraft directly for logo experiments, although AI image models are still uneven on typography and exact symmetry.
Color tools were one of the first design categories to use small neural networks because palette generation maps cleanly onto a low-dimensional representation that is easy to learn.
| Tool | Approach | Notes |
|---|---|---|
| Khroma | Personal neural network trained on 50 user-selected colors | Created by George Hastings, free, browser-based |
| Adobe Color | Sensei-powered palette extraction from images; trend themes | Integrated across Creative Cloud and Adobe Stock |
| Adobe Illustrator Generative Recolor | Firefly text-to-recolor | Live since June 2023 beta |
| Huemint | Transformer and diffusion modes for context-aware palettes | Free, supports brand, web, gradient, illustration |
| Colormind | Deep learning on photography and design data | Generates web color schemes |
| Coolors | Curated and rule-based | Popular non-AI generator, widely used alongside AI tools |
Khroma is the most cited example. The user picks 50 colors they like, the site trains a small neural network on those preferences, and then generates an effectively unlimited stream of palettes that match the user's taste. Khroma also shows WCAG contrast ratings for each pairing, which makes it usable for production work rather than just exploration. Adobe Color extracts palettes from any uploaded photo, ties into Adobe Stock and Behance for thematic search, and is integrated across Creative Cloud. Huemint adds context: the palette generator outputs colors mapped to specific layout slots like background, primary, and accent, so the result looks coherent in a brand or web setting rather than as a flat set of swatches.
3D generation matured more slowly than 2D because the geometry, texture, and rigging problems are all harder. By 2024, several tools could produce usable game and product assets from a text prompt or single image.
| Tool | Approach | Notes |
|---|---|---|
| Spline | Browser-based 3D editor with AI texture and style transfer | Spline AI Alpha launched December 21, 2023; raised $15M Series A in July 2023 |
| Meshy | Text-to-3D and image-to-3D | Auto-rigging, 500+ animation presets, PBR textures |
| Tripo | Text-to-3D, image-to-3D, AI rigging and texturing | Free tier, fast generation |
| Luma AI | NeRF-based capture, plus Genie text-to-3D and Dream Machine video | Photorealism focus from real-world capture |
| Common Sense Machines (CSM) | Cube platform converts sketches, images, or text to 3D meshes | Founded 2020 by Tejas Kulkarni and Max Kleiman-Weiner; acquired by Google in 2025 |
| Rodin | High-quality text-to-3D | Used in production game art pipelines |
Spline was founded in 2020 and accepted into Y Combinator in 2021. The company raised a $15 million Series A in July 2023 and launched its Spline AI Alpha on December 21, 2023, adding style transfer, AI textures, and prompt-driven content. The company positions itself as a browser-native alternative to Cinema 4D and Blender.
Meshy and Tripo are the workhorses for game and product asset generation. Both turn text or a reference image into a textured mesh in seconds. Meshy includes built-in auto-rigging, animation presets, and PBR texture generation. Luma AI began as a NeRF capture company and expanded into text-to-3D with Genie and AI video with Dream Machine. Common Sense Machines, founded in 2020 by Tejas Kulkarni (formerly DeepMind) and Max Kleiman-Weiner (MIT), built the Cube platform for production-ready 3D assets and was acquired by Alphabet in 2025, with the team joining Google DeepMind.
Industrial generative design is older than the generative AI consumer boom by roughly a decade. The dominant vendors are Autodesk, Siemens, and the lattice specialist nTop (formerly nTopology).
Autodesk shipped generative design as a feature inside Fusion 360 around 2017. Unlike diffusion models, it uses topology optimization and design-space exploration. Engineers define a problem with loads, supports, materials, manufacturing constraints, and obstacles. The software produces dozens or hundreds of candidate geometries that meet the requirements.
The Airbus partition project, built in collaboration with Autodesk's research arm The Living and Airbus subsidiary APWorks, was the public proof point. The bionic partition is the largest metal 3D-printed component flown on a commercial airliner, milled from Scalmalloy, a high-performance aluminum-magnesium-scandium alloy designed for additive manufacturing. The final part is roughly 45 to 50 percent lighter than the conventional partition while meeting all safety requirements. Across an A320 fleet, the weight savings translate to meaningful fuel and emissions reductions.
General Motors began a multi-year generative design partnership with Autodesk in 2018. The first joint case study was a redesigned seat-bracket. The original part was eight pieces of stainless steel welded together. The generative version is a single piece, 40 percent lighter, and 20 percent stronger. The software proposed more than 150 alternatives during the exploration. Black & Decker's Breakthrough Innovation group used Autodesk's topology optimization and Netfabb additive manufacturing tools to redesign a hydraulic crimper used by line workers. The new design is roughly 3 pounds lighter than the original 15.4 pound tool, with the lattice-like geometry typical of generative output.
| Platform | Vendor | Use case |
|---|---|---|
| Fusion 360 Generative Design | Autodesk | Mechanical parts, mass reduction, manufacturing-aware geometry |
| Generative Design for Revit | Autodesk | Architectural space planning |
| nTop | nTop (formerly nTopology) | Lattice structures, lightweighting, support optimization, field-driven design |
| Siemens NX with Topology Optimizer | Siemens | Manufacturing-aware topology optimization on convergent geometry |
| Simcenter HEEDS | Siemens | Multi-objective design space exploration |
nTop was founded in 2015. Its lattice tools allow engineers to define cell types, fill rules, and parameter fields at every point in space. The third-generation lattice tools can generate a lattice with more than 50,000 unit cells in close to real time on a GPU. nTop 3.0 introduced GPU-accelerated visualization with 10x to 100x performance gains and is used heavily in aerospace, medical devices, and motorsport.
Siemens NX includes a topology optimizer integrated with the company's convergent modeling kernel, which lets designers work on facet and B-rep data together. The optimizer is manufacturing-aware: by adding constraints for the chosen process, like cast, machined, or additive, the result is a part that can actually be made. NX's Design Space Explorer pairs the optimizer with Simcenter HEEDS for multi-objective optimization across many design variables.
Architecture AI has two layers. The older layer is parametric design and rules-based optimization, sometimes labeled computational design. The newer layer adds LLMs and diffusion models for natural language prompting and presentation imagery.
| Platform | Founded / acquired | Focus |
|---|---|---|
| Spacemaker | Acquired by Autodesk November 2020 for $240 million | Site analysis, massing, daylight, wind, noise |
| Hypar | Founded 2018 by Anthony Hauck and Ian Keough | Cloud building design automation; Hypar 2.0 added AI in September 2024 |
| TestFit | Real estate feasibility | Site solver for parking, units, FAR; generative design launched June 2024 |
| Maket | 2023 | AI residential floor plan generator |
| Snaptrude | AI-native BIM | Universal Graph Representation, exports to Revit and Rhino |
| Galileo AI / Stitch | Acquired by Google May 2025 | Originally UI design, now in Google Labs |
Autodesk's $240 million acquisition of Oslo-based Spacemaker in November 2020 brought one of the most mature urban design AI products into a major CAD platform. Spacemaker analyzes site conditions like terrain, wind, daylight, noise, and zoning rules to generate and rank massing options. Hypar, founded in 2018 by Anthony Hauck (formerly Autodesk, where he led Generative Design for Revit) and Ian Keough (creator of Dynamo, the visual programming environment for Revit), offers cloud-based building design automation with a focus on healthcare, workplace, and data center floor plates. Hypar 2.0, released in September 2024, integrated AI features including layout suggestions and a text-to-BIM function.
TestFit is the tool real estate developers use to scope a site before they commit to design fees. It produces parking, unit-mix, and feasibility studies in minutes rather than weeks, often inside a 10 to 15 percent margin against the eventual design. TestFit Generative Design, released in June 2024, lets users optimize for floor area ratio, parking ratio, and yield on cost. Maket, launched in 2023, focuses on residential floor plans, and has reported more than a million registered users. Snaptrude is the most ambitious newer entrant: an AI-native BIM platform with ten specialized agents for zoning, massing, floor plans, and code compliance, built on a graph data model that connects geometry to building rules.
The stock photo industry was caught between two threats and one opportunity. Generative models threatened to replace much of the work contributors sold. The same models needed enormous image datasets to train. The opportunity was to license those datasets in exchange for being part of the AI supply chain.
Shutterstock and OpenAI began their relationship in 2021 when Shutterstock licensed its catalog to OpenAI for training DALL-E. In October 2022, Shutterstock announced its public integration of DALL-E into the Shutterstock site, launched a contributor compensation fund to share royalties with artists whose work trained the model, and committed to shipping AI generation tools in the months following. The expanded six-year partnership was announced in July 2023. Shutterstock later launched Shutterstock Generate, branded as a commercially safe AI image generator.
Getty Images took a different path. The company sued Stability AI for copyright infringement in early 2023 and built its own generative tool. Generative AI by Getty Images launched on September 25, 2023, trained on Getty's own creative library using the Edify model architecture from NVIDIA Picasso. The product comes with the same uncapped indemnification Getty offers on stock photos, perpetual worldwide license, and royalty payments to contributors whose work contributed to training. Generative AI by iStock arrived later, aimed at small and mid-sized customers. Omnicom received first-mover access to the Getty system through a separate agreement.
Beyond the headline platforms, a long tail of specialized tools serves specific design needs. PatternedAI, based in London, generates seamless print-ready patterns from text prompts, used by more than 600,000 designers and textile manufacturers. The platform supports recoloring, mockup previews on physical products, and high-resolution PNG or vector export for rotary, screen, and digital printing. Civitai is the open community hub for Stable Diffusion checkpoints, LoRAs, and embeddings, with users sharing custom-trained models for character design, illustration styles, and architectural visualization. NightCafe is one of the earliest community-facing AI art platforms and supports multiple models in a single interface. Leonardo.Ai, founded in 2022 and acquired by Canva in July 2024, specializes in game asset and concept art workflows, with multiple fine-tuned models and a Canvas Editor for in-place edits.
Figma announced Make Designs as part of its AI suite at Config in June 2024. Within days of broader access, Andy Allen, the founder of NotBoring Software, tweeted three side-by-side comparisons showing Make Designs producing weather-app screens that looked nearly identical to Apple's iOS Weather. Allen ran the prompt three times and got the same result each time. CEO Dylan Field disabled Make Designs the following day and committed to a full quality audit before re-enabling the feature.
Figma's blog post a few weeks later explained that Make Designs combined off-the-shelf large language models from OpenAI and Amazon's Titan with a custom Figma design system. Several assets inside that custom design system were too closely modeled on specific real-world apps, which is how the output ended up looking like Apple Weather under certain prompts. The episode became a reference point inside the design community for the difficulty of building generative UI tools that do not regurgitate existing IP. The feature returned in modified form, and Figma later replaced it entirely with Figma Make, announced at Config 2025 in May.
In June 2024, Adobe pushed a Creative Cloud update that included a new acceptance screen for its Terms of Use. One clause in particular caught the attention of users: "We may access your content through both automated and manual methods, such as for content review." A single tweet citing that language collected more than 5 million views and 50,000 likes. Designers across Photoshop, Substance, Premiere, and Lightroom feared the terms allowed Adobe to train its generative models on their files.
Adobe responded within days. The company stated it had never trained generative AI on customer content, that it did not take ownership of customer work, and that the scanning clause referred to checking cloud-stored material for illegal content such as CSAM. On June 18, 2024, Adobe published revised Terms of Use that explicitly stated Adobe Firefly is trained only on licensed Adobe Stock, public domain, and openly licensed content, distinguished local files from cloud files, and gave users the ability to opt out of the product improvement program. The episode hurt trust even after the clarification, and many creators publicly cancelled subscriptions.
Firefly is marketed as commercially safe in part because Adobe trained the model on Adobe Stock. That decision sparked separate frustration among Adobe Stock contributors, who said they had not been clearly notified that their submissions would be used for training. Contributors also pointed out that AI-generated imagery has flooded the Stock library, often using prompts that include the names of human contributors, which can produce works that mimic an artist's recognizable style. Adobe has said it is committed to compensating contributors and has expanded payout programs tied to Firefly use, but the policy remains contested.
The artist controversies that defined the early Stable Diffusion era still shape policy discussions in design AI. In October 2022, a Reddit user fine-tuned a Stable Diffusion model on 32 illustrations by Disney artist Hollie Mengert and shared the result. Mengert told Andy Baio's Waxy blog that she had not been asked and described the experience as invasive. That same month, MIT Technology Review reported on Greg Rutkowski, a Polish digital painter whose name had become one of the most-used prompts on Stable Diffusion, ahead of Picasso and Michelangelo. Stability AI later removed his name as a recognized prompt in Stable Diffusion 2.0. The community responded by training LoRAs that recreated his style and sharing them on Civitai, putting the question of consent back into the open. Both cases continue to be cited in litigation and in industry conversations about training data and opt-out mechanisms.