A painting can make a road seem to recede without containing a single measured distance. Perspective, overlap and familiar object sizes provide enough clues for us to read depth. A model estimating depth from one photograph has to make related inferences, but it must express them as numbers.
Marigold repurposes a Stable Diffusion prior for monocular depth estimation. The December paper describes fine-tuning with synthetic data to produce affine-invariant depth. The attraction is the visual knowledge already present in the generative model: patterns about objects and arrangements that may help interpret images beyond a narrow training set.
A map of relationships
A depth image can be useful without specifying meters. It can indicate that the chair is closer than the wall and that the floor recedes behind it. Those relationships can support compositing, visual effects and approximate scene interpretation.
Affine-invariant depth leaves scale and shift unresolved in the depth representation. Two maps related by an allowed global transformation can describe the same normalized result. This is why a plausible grayscale image should not be read as a calibrated distance sensor output.
A miniature room and a full-sized room can produce similar photographs under corresponding camera arrangements. The pixels alone do not always resolve physical scale. A model can make an informed guess based on familiar objects, but that guess is different from an independent measurement.
The same ambiguity appears with texture and shading. A dark shape may be a recess, a shadow or a painted mark. Image understanding can help select a plausible interpretation. It cannot make the missing evidence appear.
What would make the result useful?
For a visual effect, the first question might be whether foreground boundaries are coherent. A depth-based blur that leaks background values into a person's hair can be distracting even when most of the scene is correct. Thin objects and transparent surfaces deserve special attention because they are easy to simplify or misinterpret.
For a small camera move, the question changes. A single depth map can support a limited reprojection, but moving the view exposes regions that were hidden in the source image. Depth tells us where observed pixels might go. It does not provide the missing background behind them.
That makes a useful experiment quite modest: choose an image with clear layers, generate a depth estimate and inspect a restrained effect. If the aim is parallax, move only far enough to reveal the relationship between layers before missing information becomes dominant. A small amount of parallax can be expressive even when the image only supports a limited viewpoint change.
I would also compare images that challenge the learned expectations: unusual scale, reflections, illustrations or unfamiliar interiors. A method's ability to generalize matters most when the scene differs from the examples that make the task easy.
The generative origin is the part I find especially interesting. A model trained to produce images has learned patterns that can be adapted to interpreting them. Marigold turns that prior toward a geometric output, while still inheriting the uncertainty of observing a three-dimensional world through one picture.
For graphics work, that can be a very useful kind of estimate. The key is to choose an operation that benefits from relative structure and to preserve the distinction between an inferred arrangement and a measured scene.