The longest tail of this week's releases had nothing to do with chatbots. It was generative media walking away from the flat image and the five-second clip and starting to deal in geometry, time, and physics — 3D meshes, 4D reconstructions that move, motion you can transfer between videos, forces you can apply to a scene, and the consistency problem that has always broken long video. Nine releases worth knowing about landed in the span of a few days.
That's too many to absorb as a flat list, and most lists like this quietly pretend everything is equally usable. It isn't. So the cut that matters is readiness: what you can drop into a pipeline this week, and what's still a research signal pointing at where the next year goes. I've sorted them that way.
Available now
Scale 2 — open-source motion transfer that holds up against closed Kling 3. This is probably the best open motion-transfer model released so far. You take the movement from one video and map it onto a different character, and the impressive part is how it behaves in the cases that usually fall apart. Multi-character scenes hold together: two figures fighting in a forest keep their separate details instead of bleeding into each other. It works on non-human subjects, mapping a person's motion onto a flamingo and keeping the movement roughly intact. It copes with odd proportions, animating a chubby creature with abnormal dimensions by inferring a sensible skeleton anyway. Realistic footage, anime, other art styles — all in range. It even picks up the camera movement of the driving video, which the closed Kling could not. On quality it's reported on par with Kling 3 and clearly past Wan Animate and the original Scale. The cost of entry is size: roughly 81 GB of weights, so broad consumer use waits on quantized or GGUF builds. One detail worth noting is the affiliation — Scale 2 comes out of Z.ai, the lab behind GLM, which we cover in the open-weight surge.
Mesh Flow (Meta) — fast text, image, or point cloud into a real 3D mesh. What you get out is an actual mesh, with vertices and edges, not a point cloud or a vague blob you have to clean up afterward. The selling point is speed. Most mesh generators build the thing up step by step, almost a token at a time. Mesh Flow instead compresses meshes into a latent space using a MeshVAE and samples from that space directly, which it claims runs around 18 times faster than the step-by-step approach. Code is released.
World Tracing — one image into layered 3D. Most depth estimation gives you the front surface and stops. World Tracing gives every pixel a small stack of 3D points: the visible surface first, then the geometry hiding behind it — the back of an object, the wall behind a piece of furniture. That layered cloud lifts into a textured mesh without retraining, and because the hidden geometry is there, you can do real edits like adding or removing objects rather than just spinning a nice preview. It ships in three variants for static objects, static scenes, and moving objects, and it's tiny at about 6.2 GB. Code is released.
Video MDM — 3D human motion with no motion-capture data. It generates coherent 3D human motion straight from a text prompt, and the trick is in how it was trained: on body poses extracted from ordinary 2D video rather than on expensive mocap rigs. It covers a wide action vocabulary — waving, walking backward in a straight line, deadlifts, push-ups, barbell rows — which suggests the pose-extraction approach generalizes rather than memorizing a handful of canned moves. MIT licensed, code released.
Research signals, code not yet out
These you can't deploy this week. But each one names a capability worth designing toward, because the gap between "demo on a project page" and "available on GitHub" has been closing fast for this whole category.
- Flex 4D Human turns ordinary human videos into full-body 4D reconstructions — 3D plus time — that you can view from any angle and drop into another scene. Feed it more reference videos, from one to two to four, and the accuracy and consistency climb. It does this without explicit skeletons, depth maps, or surface normals, reconstructing straight from raw footage. Code release is in preparation.
- Surflow fuses multiple images of a scene into one clean global 3D model, and it doesn't need the images aligned or the camera positions known up front. The shared global state gets cleaner as you add views instead of ballooning in size, which is the opposite of how naive multi-view methods behave. Code is promised "as soon as possible."
- Moverse turns a single image into a navigable 360-degree world that runs at 8 frames per second in real time on a single RTX 4090 — a consumer card, not a data-center GPU. It pulls that off by splitting the job in two: first expand the image into a full 360-degree panorama and bake it into a 3D Gaussian model that acts as reusable spatial memory, then let a video renderer follow a user-controlled camera path through it. The output is low-resolution and blurry today, but real-time interactive worlds on a gaming GPU is the kind of efficiency that tends to improve quickly once the structure is right. Code is under corporate review, with the team estimating about a month.
- MilliVid goes after the hardest problem in AI video, which is staying consistent over a long runtime. Current models make great 5-to-20-second clips and then lose the plot when you ask for more — forgetting what parts of the scene looked like, drifting, accumulating errors. MilliVid represents each frame at multiple levels of detail, coarse layout and semantics at one level and fine texture at another, and generates coarse-to-fine. That structure lets it hold onto scene layout far longer. Paper and code were due right around the time you're reading this.
- StreamForce gives a video model a physics joystick. Instead of describing motion in a prompt, you apply forces directly to the scene — a global force that acts like wind across the whole frame, or a local one that pushes a single chair or apple — and you can change that force over time. It's causal and streaming, so it reacts in real time as you adjust the input, at up to 16.6 frames per second on a single CPU. Code is "coming soon."
The thread connecting all nine
It's tempting to treat these as nine unrelated drops that happened to share a week, but there's a common move underneath them. For the last few years, generative media meant producing an output you look at — an image, a clip — and then you were done with it. The model's job ended at the surface. What this batch has in common is that the output is no longer the end state. It's a thing you keep working with: a mesh you can edit, a layered scene you can rearrange, a 4D human you can re-pose and drop into new footage, a world you can navigate, a video you can push on with forces. The artifact persists and stays manipulable, instead of being a finished render you either accept or regenerate from scratch.
That shift matters more than any single model's quality numbers, because it changes where these tools fit in a workflow. A one-shot image generator sits at the end of a pipeline, producing final assets. A model that outputs editable geometry sits in the middle of one, producing things that downstream tools and human artists continue to shape. The second position is far more valuable and far stickier, because it integrates with how creative work actually gets done — iteratively, collaboratively, with revisions — rather than demanding the prompt be perfect on the first try.
It also explains why so many of these releases foreground the artist staying in control. Scale 2 preserves the original's camera movement instead of inventing its own. World Tracing exposes hidden geometry so you can edit what's behind an object. StreamForce hands you a force you apply by hand. None of these is trying to take the human out of the loop. They're trying to give the human better material to work with, which is a more durable proposition than full automation and, not coincidentally, a much easier one to sell to a studio that has spent decades building craft it doesn't want to throw away.
What it means if you build with media
Across all nine, the direction is the same: generative AI is moving from pixels to world state. Geometry that persists when the camera turns away. Motion that transfers between subjects. Scenes you can walk through and push on rather than just look at. For a studio or a product team, that breaks down into two practical reads.
The first is that the deployable set is real today, not aspirational. Scale 2, Mesh Flow, World Tracing, and Video MDM are production-grade for any team with the GPU budget to run them. Motion transfer, fast mesh generation, layered 3D from a single photo, and mocap-free animation have all crossed from research curiosity to tools you can put in a pipeline this quarter. The thing gating adoption is hardware footprint, not capability, and footprint is a problem that quantization and the next generation of cards tend to solve on their own.
There's a hardware caveat worth stating plainly, because it's the thing that quietly decides which of these you can use. Several of the available models are heavy — Scale 2's 81 GB of weights is the clearest example — and "production-grade" assumes you have the GPUs to run them at the resolution and speed your work demands. For a studio that already owns a render farm, that's a non-issue. For a small team, it's the whole question, and the honest answer is to wait for the quantized and GGUF builds that the community reliably produces a few weeks after a release like this. That lag is predictable enough to plan around. The capability arrives first for the well-resourced and then, on a short delay, for everyone else, which is roughly the same pattern every previous wave of generative media followed before the tools became something you ran on a laptop. Plan your pipeline around the capability landing, not around today's footprint, because today's footprint is the part that always shrinks.
The second is that the research signals tell you where to place architectural bets before they ship. Long-form video consistency from MilliVid, real-time navigable worlds on consumer hardware from Moverse, force-controllable physics from StreamForce — once those land, they change what an interactive or generated experience can even be. Design with the assumption they're months away rather than years, because for this category that's been the pattern. And the discipline carries over from the language-model side: the same vendor-independence and routing instinct you'd apply to your LLM stack applies here too. This field is moving fast enough that committing your whole pipeline to a single generator is a bet against your own flexibility, and flexibility is the thing you'll want most when the next nine releases land next month.
Part of our coverage of an unusually dense week in AI. See also: Fable 5's six-day life, the open-weight surge, world models and robotics, agent benchmarks and harnesses, and the week's open AI primitives.