The AI Race You Could Actually See

There is a version of the AI story that most people know. ChatGPT arrives, the big labs scramble, models multiply, benchmarks fly. I covered that story in a previous post, tracing the language model race from November 2022 to last week. There have been more releases since, because of course there have.

But there is another race that was running in parallel the whole time, and in some ways this one was more visceral, more immediate, and landed closer to home for anyone who works in marketing, design, or any kind of creative field. You did not need to understand tokens or reasoning benchmarks to feel this one. You could actually see it.

This is the story of AI image generation, how it started, how it went cinematic, and what it means for the rest of us.

The AI image generation story started earlier than you think

Most people place the start of the AI image story at 2022, when Stable Diffusion and DALL-E 2 went public. But the groundwork was laid earlier than that.

In January 2021, OpenAI released the original DALL-E. It was research-grade, not publicly available, and the outputs were strange and pixelated by today’s standards. The famous “avocado chair” image from that era looked nothing like a real photograph, but it proved something important. A machine could take words and turn them into a picture. Not perfectly. Not even close to perfectly. But it could do it.

Midjourney launched its first private demo in September 2021, built by a small team of engineers with a philosophy that stood apart from the rest. Rather than chasing photorealism and technical benchmarks, founder David Holz wanted to build something closer to an artistic tool, something that expanded what people could imagine rather than simply replicated what already existed.

By early 2022, Midjourney V1 was in the hands of a small invited group. The outputs were raw, dreamlike, and strange. Hands were wrong. Faces melted. But something about the aesthetic stopped people in their tracks. This was not photography. It was not illustration. It was something new.

2022: The year it went public

Three things on AI image generation happened in 2022 that changed everything.

In April, DALL-E 2 launched to a waitlist. OpenAI had dramatically improved on the original, sharper, more coherent, more controllable. For the first time, people outside research labs could generate images from text descriptions and get results that looked like something you might actually use.

In July, Midjourney opened its public beta, accessible entirely through Discord. You joined a server, typed a command, and four image options appeared. The community exploded almost immediately. By the time Midjourney V3 launched in the same month, the Discord server had grown to one million users, overtaking the Fortnite and Minecraft servers as the largest on the platform.

In August, Stability AI released Stable Diffusion as fully open-source. This was the moment that changed the structural nature of the race. Not because it was the most beautiful output, but because anyone could download it, run it on their own computer, modify it, fine-tune it on their own images, and build on top of it for free. The genie was out of the bottle in a way that could not be put back.

By November, Midjourney V4 had arrived and the quality leap was significant enough to go viral. AI-generated images were showing up in news articles, social media feeds, and creative briefs. The conversation had started.

2023: The quality crosses a threshold

If 2022 was about proving the concept, 2023 was about AI image generation crossing a line that most people did not realise existed until it had already been crossed.

In April 2023, Midjourney V5 arrived. Not “impressive for AI.” Just impressive. Portrait photography that required a second look to determine whether a camera had been involved. Concept art that working professionals could use without embarrassment. Designers and marketers started incorporating it into real workflows, not as a novelty but as a genuine production tool.

Adobe entered in March 2023 with Firefly, and the positioning was deliberate. Trained only on licensed Adobe Stock images and public domain content, Firefly was the copyright-safe option for enterprises that could not afford legal risk. It reached one billion images generated in its first three months. The creative industry had a legitimate tool it could actually defend in front of a client or a legal team.

In October 2023, DALL-E 3 launched inside ChatGPT. The text rendering improved dramatically. For the first time, an AI image model could put legible words inside a generated image with reasonable reliability. That single capability unlocked an entire category of use cases that had previously been off the table.

Also in 2023, Getty Images sued Stability AI for copyright infringement. A group of artists filed a class action lawsuit against Midjourney, Stable Diffusion, and DeviantArt. The creative industry had gone from watching to litigating.

The aesthetic tool had become an industry threat.

2024: Video walks into the room

Everything above had been still images. Remarkable, disruptive, legally contested still images. Text-to-video had been around since 2022, when Meta’s Make-A-Video and Google’s Imagen Video showed it was theoretically possible. The results were blurry, three seconds long, and nowhere near usable. Runway Gen-2 in early 2023 was the first commercially available tool that produced something you could actually work with. But in February 2024, OpenAI showed the world something that changed the scale of the conversation entirely.

Enter Sora.

The demo videos were not “good for AI.” They were cinematic. A woman walking through neon-lit Tokyo rain, her reflection catching in wet pavement. A woolly mammoth charging across snowy grasslands with a camera pull that felt like a David Attenborough production. A slow drone shot moving through a fantasy cityscape. People who had spent careers on film sets watched the demos in awe.

Sora was not publicly available yet. OpenAI gave access to a small group of filmmakers and researchers. But the signal was clear. AI had learned to make video that looked real.

Meanwhile, a company called Runway had been building away in the background, producing the tool that professionals actually started using. Runway Gen-2 had shipped in 2023. Gen-3 Alpha arrived in 2024 and was the first public model to produce multi-second video with stable motion, coherent camera movement, and usable output. While Sora generated headlines, Runway became the working production standard.

From China, Kuaishou launched Kling. The motion quality on human subjects was striking, and the cost was significantly lower than Western equivalents. The pattern from the language model race was repeating itself. Chinese challengers arriving faster than the West had anticipated, matching capability at lower cost.

Pika launched its 2.0 version with a feature it called Pikaffects, stylised creative transformations that made it the tool of choice for social media creators who wanted something quick, expressive, and feed-native. It wasn’t trying to be production grade.

2024 into 2025: Hollywood fights back

The IP war that had started with still images escalated sharply when video entered the picture.

CAA, the talent agency representing some of the biggest names in Hollywood, called Sora a serious and harmful risk to their clients’ intellectual property. The Motion Picture Association demanded OpenAI take immediate and decisive action over copyright infringement. Writers, illustrators, voice actors, photographers, and musicians filed lawsuits across multiple jurisdictions. The EU AI Act kicked in with transparency and disclosure requirements for AI-generated content.

The creative industry had stopped being scattered and anxious. By this point it was organised, funded, and litigating on multiple fronts simultaneously.

OpenAI released Sora publicly in December 2024. The public version was more constrained than the demos suggested. Users found it difficult to access, expensive to run, and inconsistent in ways the curated demos had not revealed. The gap between the highlight reel and the actual product was significant.

Adobe continued adding AI capabilities to Photoshop, Illustrator, and Premiere, threading Firefly through the tools creative professionals already used rather than asking them to learn something new. For agencies and brand teams with existing Adobe workflows, this was the path of least resistance.

Google released Imagen 3 and began integrating video generation through what would become Veo. The Google approach was integration over standalone product, with AI capabilities threaded through Workspace, Gemini, and YouTube rather than a separate app to learn.

2025: The race goes wide

If 2024 had four or five serious players, 2025 had a dozen. The race went from a sprint to a field event.

Runway Gen-4 arrived in spring 2025 and became the professional production standard. Character consistency across scenes, precise camera controls, a reference image system that let you maintain visual continuity shot to shot. Ad agencies and film studios started building it into actual workflows, not just experimenting. If you worked in production in 2025, you knew Runway.

Google Veo 3 landed and changed the video conversation fundamentally. For the first time, a video generation model could produce native audio in the same pass as the footage. Ambient sound, dialogue, and music generated simultaneously rather than stitched together afterwards. The implications for content production were immediate. You could generate a scene with a person speaking and have their voice, the room tone, and background sound all arrive together.

Midjourney continued iterating through V7, with a strong emphasis on character consistency and a new feature called Consistent Characters that let you generate the same person across multiple scenes and angles. A capability that had been the exclusive domain of expensive production pipelines suddenly became available to anyone with a subscription.

Black Forest Labs released Flux, an open-source image model that stunned the community with its photorealism and prompt accuracy. Flux became the reference benchmark for what open-source image generation could achieve, and the community built an ecosystem of fine-tuned models around it rapidly.

Stable Diffusion 3.5 arrived as Stability AI’s most capable model, though the company’s financial troubles had begun to slow its competitive pace. The open-source ecosystem it had spawned remained vibrant even as the company itself struggled.

From China, ByteDance released image and video generation capabilities through its creative tools. Alibaba’s Wan 2.1 and 2.2 released open-source weights for a multimodal image and video model that held up against closed Western peers with few obvious weaknesses. Tencent’s HunyuanVideo had already released strong open-source video weights in December 2024. The Chinese open-source video models were arriving faster and more capable than anyone had forecasted.

Ideogram and Recraft carved out specific niches. Ideogram became known for graphic design and text-heavy image generation, posters, logos, and social graphics where accurate text rendering mattered. Recraft focused on clean vector-style illustration for designers who wanted something closer to professional graphic design output than photographic realism.

By this point, the AI image generation landscape had grown crowded enough that a new category emerged to make sense of it all. Aggregator platforms like Higgsfield, Fal.ai, Replicate, and Leonardo.ai gave users access to multiple models in one place, paying per generation rather than juggling five separate subscriptions. Higgsfield pulled together Runway, Kling, Luma, and image models like GPT Image and Recraft under one dashboard. Replicate let developers run open-source models on demand, paying per second of compute. Leonardo.ai built its own fine-tuned models on top of Stable Diffusion, aimed squarely at game artists and concept designers. When the market needs platforms just to manage access to all the platforms, that tells you something about how far things have come.

HeyGen and Synthesia dominated the talking-head video category. Not cinematic video, but AI avatars that could present in multiple languages with accurate lip-sync. The enterprise training video and corporate communications market found its tool.

2026: The market settles, mostly

Sora was shut down in March 2026. OpenAI announced the web and app experiences would be discontinued in April, the API in September. The economics had not worked. The IP battles had not helped. The most hyped AI product of 2024 became one of the more expensive product exits of 2026. Teams that had built workflows around it were told to migrate to Veo, Kling, Runway, or Seedance.

Midjourney reached V8.1. The company remained one of the most unusual success stories in the entire AI race, profitable since its second month, no venture capital funding, roughly 107 employees, $500 million in annual revenue, and valued at around $10 billion.

GPT Image 2 from OpenAI became the go-to for complex prompt-following image generation. If you needed an image that precisely matched a detailed description, including legible text, GPT Image 2 was the benchmark.

The market as of mid-2026 looks something like this. For image generation: Midjourney V8.1 for artistic and editorial output, GPT Image 2 for precise prompt execution, Adobe Firefly Image 3 for commercially safe enterprise use, Flux for open-source and developer workflows, Ideogram for text-heavy graphic design work, Recraft for clean illustration. For video: Google Veo 3.1 as the all-around quality benchmark, Runway Gen-4 as the professional production tool, Kling 3.0 for high-motion scenes and human subjects, Pika for social media content, HeyGen and Synthesia for talking-head and corporate video. Most professional teams in 2026 use two models, not one, a workhorse for volume work and a premium model for hero shots.

What this actually means

Here is the thing about the image and video race that makes it different from the language model race. The LLM story is largely about productivity, who can answer questions faster, write code more accurately, synthesise information more reliably. It affects how we work.

The image and video story cuts deeper than that. It goes after craft. It goes after the skills that creative professionals spent years developing. The photographer who spent a decade learning light. The illustrator who built a visual style that clients commissioned specifically. The motion graphics designer who understood how to make something feel cinematic. All of these skills are now being partially replicated by tools that cost a few cents per generation.

That does not mean the skills have no value. The best creative work still requires taste, judgement, direction, and a point of view that no prompt can substitute for. But the floor of what is acceptable has moved dramatically, and the economics of creative production have shifted in ways that are still working themselves out.

The lawsuits are not over. The IP questions are not resolved. The question of what happens to the next generation of designers and photographers and filmmakers who are learning their craft in an environment where AI can produce a decent version of almost anything is genuinely open.

I do not have a clean answer to that. But I think it is the more important question to sit with than “which model should I use this week.”

Here is the sum of it all. In under five years, AI image generation went from generating a blurry avocado on a chair to producing cinematic video with native audio. The tools went from research labs to Discord servers to professional production workflows. The lawsuits went from a few concerned artists to Hollywood studios and government regulators. And the cost of generating a five-second video clip dropped from $2.50 to less than 30 cents in under eighteen months. Whatever you think about where this is heading, the pace of what has already happened is worth sitting with for a moment.

How are you navigating AI in your creative or marketing work? I would love to hear your perspective in the comments below.