Project Casper

1.How an image model paints

Nearly every frontier image and video model today is a diffusion model or a close relative[1][2]. The idea is simple to state. Take a real image, add a little Gaussian noise, then a little more, until after enough steps nothing remains but static. Train a network to undo one step. Then start from pure static and run the network backwards.

Formally, the forward process mixes a clean image $x_0$ with noise according to a schedule $\beta_t$. Because each step is Gaussian, any step can be sampled directly:

A network $\epsilon_\theta$ learns to predict the noise that was added, given the noisy image, the step, and a condition $c$ such as a text prompt. The training objective is a plain squared error:

Sampling runs the chain in reverse, subtracting a scaled estimate of the noise at each step and adding a small amount back to keep the process stochastic. Structure appears coarse to fine, and nowhere in the loop is there a scene, only increasingly confident pixels:

Modern systems run this loop in a compressed latent space rather than on raw pixels[3], and have moved the denoiser from convolutional U-Nets to transformers, which scale more predictably with compute[4]. How quickly signal is destroyed is a design choice. The cosine schedule keeps more of the image alive in the middle steps than the original linear one, which is where global layout gets decided[5].

Text steering comes from classifier-free guidance, which extrapolates from an unconditioned prediction toward a conditioned one[6]:

Read that equation closely and the problem appears. Everything the model knows about the scene has to arrive through $c$ or live implicitly in its weights. A sentence is a remarkably poor description of a room.

2.Where the world goes missing

A diffusion model learns a distribution over images, $p(x \mid c)$. It never learns a distribution over scenes. Geometry, occlusion, contact, and scale are whatever was statistically typical in its training pixels. When a prompt pins them down, the model obliges. When it doesn't, they are free variables, and sampling fills them with something plausible. Plausible and correct are different things.

The benchmarks are blunt about this. GenEval checks simple, verifiable properties of generated images with an object detector. On its position task, which only asks whether one object sits left of, right of, above, or below another, SDXL scored 0.15 in the original paper, against 0.98 for simply rendering a single named object[7]. T2I-CompBench finds the same pattern across spatial relations, attribute binding, and object counts[8]. Models know what things look like far better than where things are.

Video makes the gap visible. A moving camera is a continuous demand for geometric consistency. OpenAI's Sora report noted that 3D consistency and object permanence emerged with scale, without explicit 3D supervision, and also that the model still fails to simulate basic physical interactions like glass shattering[9]. That is the honest state of the field: spatial structure can emerge from data, partially, and it degrades exactly where you most need it.

There is a simple reason partial is not enough. A scene is not one constraint, it is dozens. The strap meets the shoulder. The shadow falls away from the window. The bag is the same size in both hands. If each is held independently with probability $p$, the chance an image holds all $k$ of them is:

This is the core claim of the paper. Hallucination in generated scenes is, to a first approximation, an under-constraint problem. You do not fix it by making the painter more talented. You fix it by giving the painter a room.

3.What a spatial construct is

Computer vision has a precise language for rooms, and it predates deep learning by decades[10]. A point $X$ in the world lands on a pixel through a camera's intrinsics $K$ and its pose, a rotation $R$ and translation $t$:

The division by depth is where a single image loses information. Every point along a ray through the camera center lands on the same pixel. A second view resolves it: the candidates spread out along a line in the other image, and only one of them matches.

A spatial construct, as we use the term, is any representation that makes these relationships explicit and checkable: where surfaces are, how the cameras relate, what is in front of what. The field has built several, each with a different trade between fidelity, speed, and editability.

Neural radiance fields store a scene as a function from position and viewing direction to density and color, and render by integrating along rays[11]. 3D Gaussian splatting replaces the function with millions of explicit, anisotropic primitives that render in real time at 1080p[12]. Both produce a pixel by alpha compositing whatever lies along its ray, front to back:

That one line is the difference between a picture and a world. Occlusion, depth order, and view consistency are not learned tendencies here. They are arithmetic.

4.What the field has learned

The last four years have produced four distinct strategies for marrying image priors to geometry. Each one taught us something.

Lift 2D into 3D. DreamFusion showed that a frozen 2D diffusion model can supervise a 3D scene with no 3D training data, by rendering the scene, noising the render, and pushing the scene toward whatever the denoiser prefers[13]. The gradient, score distillation sampling, is:

It works, slowly, and inherits every inconsistency of the 2D prior, including the well-known tendency to put a face on the back of a head. Zero-1-to-3 fine-tuned a diffusion model on the relative camera transform between views[14], and MVDream generated several views jointly with attention shared across them[15]. Consistency improved when views were generated together rather than one at a time.

Condition on geometry. ControlNet showed that a pretrained generator can be made to obey a depth map, edge map, segmentation, or pose skeleton by training a parallel copy of its encoder while the original stays frozen[16]. The generator keeps its taste and gains a spine. Monocular depth estimators such as Depth Anything, trained on tens of millions of unlabeled images, now make those depth maps available from any photograph[17].

Recover geometry feed-forward. DUSt3R regressed dense 3D point maps from pairs of ordinary photos with no known camera parameters[18]. VGGT extended this to predict cameras, depth, and point maps from one to hundreds of views in a single forward pass, in under a second[19]. Geometry stopped being an optimization problem and became an inference call.

Simulate the world. Video world models carry an implicit state forward in time. Genie 2 generated playable 3D environments from a single image[20], and Genie 3 extended this to real-time interaction at 720p and 24 frames per second, staying consistent for a few minutes[21]. The ambition is right. The state is still implicit, which means it is still forgettable.

5.Project Casper

Casper is our attempt to take that pattern to its conclusion for one domain: images of a specific person, in a specific place, with a specific product. We believe the next generation of generative models will not paint into a blank canvas. They will generate from a scene: an explicit, persistent, multi-dimensional state that the model renders, conditions on, and is held accountable to.

The practical consequence is that consistency stops being a hope and becomes a property you can measure. When a scene is written down, two views of it either agree about the surfaces they share or they don't. A face either matches the person it belongs to or it doesn't. A product either keeps its shape and surface or it doesn't. A hallucination stops being a vibe and becomes a disagreement with a threshold.

The literature already shows that sampling can be steered by signals outside the prompt[22], and that generators can be made to obey structure they were never trained to infer[16]. Casper builds on both ideas. What is new is less any single mechanism than the decision to make one persistent scene the shared source of truth for everything that happens before, during, and after generation.

6.Why commerce is the right first world

General world models have to be right about everything. Commerce only has to be right about three things, and it has to be right about them every time: the person looks like themselves, the product looks exactly like the product, and the scene is somewhere they could plausibly be. Subject-driven methods such as DreamBooth showed that a model can learn a specific person or object from a handful of photos[23]. Keeping that subject correct across poses, rooms, and lighting is a spatial problem, which is why we think it is the frontier.

It is also a domain with a ground truth. Every product has a reference image. Every person has a selfie. The scene state can be checked against real inputs, which makes commerce an unusually honest testbed for spatial generation: errors are visible, measurable, and expensive.

Our first product, Reflection, already turns a brand's existing product imagery into personalized creative. In preliminary internal estimates, 70 percent of outputs meet our product fidelity criteria and 60 percent are usable without manual edits. Casper is aimed squarely at the remainder. The failures we see are almost never about style. They are about space: a strap that floats, a scale that drifts, a product that changes shape when the camera moves.

7.Open problems

We would rather name the hard parts than pretend they are solved. These are the questions we are working on and do not yet have answers to.

Where this leaves us

The first generation of generative models learned what the world looks like. The next generation will have to learn where the world is. We think that shift, from pixels to scenes, from appearance to state, is the most important open problem in visual generation, and the one most directly tied to whether people can trust what they see.

If you work on 3D vision, world models, or generative evaluation and see a flaw in this reasoning, we want to hear it. That's the most useful email we can get: founders@clockingminds.com.