AI Breakthroughs Everyone Missed: Drama 3, OmniHuman 2.0 and World Models

Key takeaways

• Drama 3 lets you write voice direction in square brackets, like notes to an actor, and it scored top for pronunciation on one independent leaderboard
• Drama 3’s API is about $10 per million characters at launch, rising to $50 after 12 November, according to the video
• OmniHuman 2.0 costs about 7 cents a second, so a five-minute video is roughly $21, at 720p only
• World models are AI-generated spaces you can move around in; their big weakness is forgetting what was behind you
• Matrix Game 4.0 and NVIDIA StreamRender-H3 fix that by letting ordinary game code run the world while AI only paints the picture
• Several projects come with their own benchmarks or demo reels only, so treat the results as claims until tested

Most weekly AI news is about the same few companies. That is useful, but it means smaller research projects and tools slip past almost everyone.

This article walks through the AI breakthroughs everyone missed in Matt Wolfe’s latest round-up, in plain words. For each one we explain what it does, whether you can use it now, what it costs where that is known, and which claims still need checking.

What are the AI breakthroughs everyone missed?

Ten releases across voice, video, world models and 3D, most of them from research teams rather than the big chat apps. Here is the short version:

What is Fish Audio Drama 3?

Drama 3 is a text-to-speech model where you direct the performance in plain English instead of picking “happy” or “sad” from a menu. You write the direction in square brackets inside the script, the way you would brief an actor.

• Whole-line direction, such as “barely holding it together, forcing a smile”
• Word-level direction, so only one phrase is whispered
• Sound cues placed exactly where you want them, such as a small laugh or a shaky breath
• Several speakers with different voices in one scene

On the Artificial Analysis leaderboard Matt shows, Drama 3 scored highest for pronunciation, ahead of ElevenLabs. In blind listening tests where people pick the voice they prefer, it did less well. In short, it reads hard words very accurately, but many listeners still prefer other voices.

If you write ads, explainer videos or audiobooks, the direction-in-brackets idea is the useful part. Our guide How to Direct AI Voice Acting With Plain-English Script Notes shows how to write those notes so they work with most modern voice tools.

What does OmniHuman 2.0 do with one photo?

OmniHuman 2.0 takes a single image and an audio track and creates a video of that person or character talking, singing or acting. The big change from version one is length: up to five minutes in one continuous shot, up from 60 seconds.

• Text prompts can add camera moves or objects that were not in the photo
• It works with cartoon characters and animals, not only people
• Output is 720p, so no 1080p or 4K
• Price dropped from 12 cents to 7 cents a second, about $21 for a full five minutes

Matt’s honest view is that AI lip sync still has a long way to go, even though it is far better than a couple of years ago. Everything shown comes from ByteDance’s own demo reel, and early users noticed backgrounds shifting.

What are AI world models, and why do they forget?

A world model is an AI that generates a video world you can move around in, frame by frame. Their biggest weakness is memory: look away and turn back, and things have often vanished or changed, because the model only remembers the last few seconds.

WorldCrafter (Tencent ARC Lab)

WorldCrafter starts from one image or one line of text. Its fix is a memory you look up by camera position: when you point the camera somewhere, it pulls earlier views of that spot and feeds them back in. To show it works, the team rebuilt a 3D point cloud from the generated video and got clean geometry rather than mush. It also kept track of a moving horse that left the frame and came back.

The team says it beat the other models on its own test of 145 scenes. That is their benchmark, so treat it with care. The code is on GitHub and there is a Hugging Face demo, which Matt found slow but working.

ByteDance’s unnamed world model

At an event in Beijing, ByteDance showed a world model that created a new camera path through a scene from Stephen Chow’s film King of Comedy, in effect reshooting an old shot from new angles. It also spawned a giant spoon you could climb and a portal into space. There is no name, no release date and no public access, and the only footage is from people filming at the event.

How do Matrix Game 4.0 and StreamRender-H3 fix the memory problem?

They stop asking the AI to remember the world at all. Ordinary game code keeps track of where everything is and what the rules are, and the AI’s only job is to make each frame look good.

Matrix Game has catches: each game has its own model of about 5 billion parameters, the model card does not guarantee a frame rate or that it fits in a home graphics card’s memory, and the training code is not included. One fun detail on NVIDIA’s side: the team says Claude Code built a faster video decoder for the project in about three hours.

What is the fruit fly brain in Odyssey 3?

Odyssey says it connected a simulated fruit fly brain to its world model, so the fly now flies around AI-generated worlds by itself. The brain simulation comes from Eon Systems and is based on a complete map of a fly’s brain, about 140,000 neurons and 50 million connections.

It is a fun clip, but there is no paper, no code and no explanation of how the two are joined, and viewers noticed the fly never tries to land. You can try Odyssey 3 itself in the browser, steering a character through a world you describe.

Which 3D tools are worth watching?

• T3lescope (Preferred Networks, Japan) turns a set of photos into a detailed 3D mesh, from one object up to city blocks. It builds a rough version first, then zooms in where detail is needed, and handles shiny and glass surfaces better than the usual Gaussian splatting method. It needs many photos (the demos used 56 to 256) and only rebuilds what your camera covered. The code is “coming soon”.
• Gaussianimate animates characters that are soft outside but rigid inside, such as loose clothes, dresses and animals, by adding floating points that capture the jiggle and flow. It needs a captured motion sequence, not a single image. Code is on GitHub for research use.
• WorldGacha is a free browser tool from a Japanese developer that generates random fictional regions with mountains, rivers, towns, railways and rice paddies, in line-art, tone or clay styles. It is not AI, just maths and random numbers, and exports PNG, SVG or 3D files.

Matt’s idea for WorldGacha is to build a rough city there, then use an AI image or video model to reskin it into something realistic. That is the same “structure first, AI paint job second” pattern as Matrix Game.

Which of these can you use today?

Only a few are ready for everyday work. Here is how we would sort them:

What is confirmed and what is not?

None of that makes these projects less interesting. It just means you should test on your own work before you rely on any of them. Our guide How to Verify AI Model Claims Before You Believe the Hype gives you a simple way to do that.

How could a small business use these?

Start with voice, be careful with faces, and keep an eye on world models.

• Better voiceovers: directed voices make ads, course videos and product demos sound less robotic. Begin with our guide to directing AI voice acting.
• Talking avatars for your own content: a five-minute talking photo of yourself costs about $21, but check quality and consent first with our talking photo checklist.
• Product and property visuals: tools like T3lescope point towards cheap 3D models from ordinary photos, useful later for shops, venues and estate agents.
• Content you own: whichever tool you use, keep the scripts, voices and source files in your own library so you can switch tools without starting again.

Frequently asked questions

The bottom line

The AI breakthroughs everyone missed are worth a look because they show where creative AI is heading. Directed voices like Drama 3 and long talking photos like OmniHuman 2.0 are usable today, with care over cost, quality and consent. World models are getting better at remembering, mostly by letting ordinary code run the world while AI paints it. Try the voice tools now, and keep the rest on your watch list.

Related posts

• GPT-6.5 Leak? MiniMax M3.1 Flash, Gemini 4 Carbon and Kimi K3.1 Explained
• How to Write Natural AI Voiceover Scripts
• How to Write a UGC Ad Script for an AI Spokesperson Video

Garry Knight

I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

See Full Bio