Why scale matters
A roof and the road beyond it tell us different things about a household’s surroundings. A tight image shows the plot; a wide one includes the village, roads, and fields. Fit either into the same number of pixels, and something must give: the wider the view, the less detail each pixel records.
In Zhu, Jerzak and Daoud (2025), we asked how this trade-off affects estimates of who benefits from an anti-poverty program. Our method, Multi-Scale Representation Concatenation, converts each image scale into a numerical description using the same image model, CLIP-RSICD, then joins those descriptions. A causal forest uses them to estimate average program effects for units with different surroundings.
Combining scales helped. The paper evaluates performance with the RATE ratio, which measures the strength of evidence that a model’s ranking identifies differences in treatment effects. In Uganda’s youth business-grant trial, using two scales raised the ratio over the best single scale by 0.95 (standard error 0.10) after compressing image features into principal components, and by 0.41 (standard error 0.09) with raw features. The best pairs combined an image about 64 Landsat pixels across with one about 350 pixels across; each pixel represented 14.5 m. Across the Peru and Uganda trials, the average ratio rose from about 2.0 with one scale to about 2.5 with five.
A generative world model offers a possible shortcut. Given images and their camera positions, it can draw the scene from another viewpoint. If those views preserve the information a missing scale would have supplied, one image might do the work of two.
We tested this on one Ugandan village, comparing generated views with drone imagery.
The two input images
Both images show the same place. One records a single compound at 5 cm per pixel. The other covers ten times the width at 50 cm per pixel.
Input given to Atlas · Neighborhood image
Household image
Input given to Atlas · Household image
| Household | Neighborhood | |
|---|---|---|
| Pixels | 1280 × 720 | 1280 × 720 |
| Ground per pixel | 5 cm | 50 cm |
| Area shown | 64 × 36 m | 640 × 360 m |
| Centered on | the study compound | the study compound |
The images are crops of one orthomosaic: a map stitched from overhead drone photographs around Azua, Terego District, West Nile, northern Uganda. The original map has a resolution of 3.9 cm per pixel and a CC BY 4.0 license. We chose this site for its openly licensed imagery.
The cameras and the prompt
Atlas requires a camera position for each input. We treated the crops as straight-down photographs over flat ground, with a 60° field of view: the household camera at 55 m and the neighborhood camera at 550 m.
Under this flat-ground setup, each requested overhead view has a matching crop in the drone map. We requested 13 views: 10 straight down and 3 at oblique angles for which we had no reference photographs.
We ran each of the three input conditions twice for 78 generated views in all (13 views per run). Seeds let us see how sensitive the model is to randomness, how much the output varies between runs.
The text prompt
“Aerial drone photograph of rural farmland and homesteads in northern Uganda, seen from the air in daylight.”
Atlas also accepts a text description. We sent this same prompt with every request, so requests differed only in which input images Atlas received, the camera positions of the views it was asked to generate, and the seed.
Atlas also generated two 20-frame descents, from 1,100 m above the study compound to a low oblique view 28 m from it: one with both images, one with the household image alone (section 7).
These runs and a 6-frame pilot produced 124 frames across 10 requests. Requests returned in 5 to 31 seconds. We used the beta World Labs Atlas (Marble 2) developer API on October 20, 2026. See also Methods and sources.
What the inputs show at each requested view
The household image supplies detail at one compound. The neighborhood image covers most requested views, but at lower resolution. Beyond those limits, Atlas must infer the scene.
Household image showsNeighborhood image shows
| Requested Atlas view | Household image shows | Neighborhood image shows |
|---|---|---|
| Study compound64 × 36 m at 5 cm per pixel | ||
| Compounds 1 to 5the same size, 130 to 250 m away (boxes in section 2) | ||
| Around the study compound192 × 108 m at 15 cm per pixel | ||
| Between compounds 1 and 2192 × 108 m at 15 cm per pixel | ||
| Neighborhood640 × 360 m at 50 cm per pixel | ||
| Zoomed out1,280 × 720 m at 1 m per pixel | ||
| Three oblique views45° from 70 m; 14 m up, 42 m southwest; 22 m up, 47 m northeast | Overhead imagery of the study compound | Overhead imagery of the neighborhood |
Compare the generated views with the map
Choose a view and an input condition. The panels show the inputs and outputs to Atlas. Drag the handle, or use the slider below the image, to compare Atlas’s view on the left with the real drone map on the right.
What Atlas received
The same inputs for all 13 views in this run


- Image
- Prompt
What we asked it to generate
This view only
- Camera
- Available evidence
Where this view is
Scores for this view
How well the views match
The neighborhood image helped Atlas place fields, paths, and yards. For fine detail at the five neighboring compounds, every generated view scored below a plain enlargement of that image.
Each dot represents one overhead view: five compounds × two seeds, or ten per row. The gray band shows the enlargement baseline. For the correlation scores, values near 1 indicate close correspondence; values near 0 indicate little linear correspondence.
Results across the five neighboring compounds
Ranges cover both seeds: ten generated views per condition.
At the study compound, where the household image supplied the detail directly, fine-detail correlations reached 0.65 to 0.93 with that image alone or with both inputs. With the neighborhood image alone, they were 0.13 to 0.15, below the enlargement’s 0.27.
At compounds 3 and 4, both-input views had sharpness scores of 0.75 to 1.09, close to the real images’ strength, but fine-detail correlations of about 0.01. Household-only views were similarly sharp, with a mean of 1.09. Their texture also failed to match the ground.
Table of every score for all ten straight-down views
Atlas imagining unseen neighbors
Looking at the images is enough to show that the household-only views of the neighboring compounds differ from the real ones. Across the five real views, we counted 16 thatched round huts and 4 rectangular buildings with metal roofs. In the ten household-only views, we counted no thatched huts.
Five generated views contained rectangular buildings, mostly with metal or tiled roofs. One included a parked car. Others showed palms, forest, or open fields. A model extracting features from these images would encounter different roofs, vegetation, and buildings from those in the drone map.
Descending toward the compound
The two 20-frame descents follow the same camera path and prompt, from 1,100 m above the study compound to a low oblique view 28 m from it. One run receives both images; the other receives only the household image.
At the starting height, the neighborhood image covers the central quarter of the requested ground footprint. Atlas must supply the landscape beyond it. With only the household image, it places the compound among a forest belt, a forked track, and fields that differ from the map.
Turn on coverage to see which input each pixel projects back into, using Atlas’s returned depth.
Both images

Household image only

Coverage of each frame, both images
Coverage of each frame, household image only
“Coarse only” means one input pixel spans more than 3 × 3 generated pixels. Gray marks pixels without returned depth, which cannot be traced to an input. In the final frame, 94% of pixels lack depth. Click a bar to jump to a frame.
Table of coverage shares for every frame
The huts look solid; the returned geometry is nearly flat
In the low oblique views, Atlas draws walls and conical roofs. Yet within 15 m of the study compound, 99.9% of the reconstructed points lie less than 0.9 m above the ground. We modeled the input images as lying on a flat plane (as the orthomosaic did not provide heights). The returned geometry remained nearly flat.
In the point cloud, the huts lie like pictures on that flat surface rather than standing as raised structures. Because the coverage tint records where each point lands on the overhead inputs in Atlas’s geometry, even a generated wall can receive a tint.
The point clouds combine frames 9 to 19, thinned to one point per 35 cm near the compound.

Frame 16 of the both-images descent. The 3-D view requires WebGL. Drag to rotate; scroll or pinch to zoom.
What this means for causal research
For a study of poverty, a thatched roof, a metal roof, and a parked car can carry different information. Replacing one with another could change the features an image model extracts and, in turn, the households it ranks as likely to benefit. We measured whether the generated views preserved the place.
The household image supplied roofs and yard detail at the study compound. The neighborhood image supplied the arrangement of fields, paths, and compounds farther away. This illustrates the information trade-off that motivates using several scales in the paper.
Both inputs did improve Atlas’s geometry: mean scale error fell from 8.2% to 1.5%, and the maximum fell from 22.5% to 6.3%. But a follow-up found similarly low errors when the household image was replaced by a blurred view or an empty field. That points to the second camera at 55 m as the source of the scale improvement. The household image’s actual detail helped where it showed the ground directly.
This investigation is not without limitations. These findings come from a single test of the Atlas beta model, run on October 20, 2026, in one Ugandan village with two seeds. We scored the generated views against held-out drone imagery at 5 cm and 50 cm per pixel, much finer than the paper’s 14.5 m Landsat pixels.
Across those runs, no input recovered fine detail at the five neighboring compounds, and household-only views drew none of the 16 real thatched huts. Using generated views as study observations would require checking a sample against held-out imagery of the same sites, and testing whether encoding them changes the causal forest’s RATE ratio.
Looking ahead, future tests can build on what worked here: the wide image preserved layout, and a second camera tightened the scale. Supplying real elevation data could let the geometry capture heights instead of staying flat, while expanding to more villages and to Landsat resolution would match the paper’s actual inputs. As world models move past beta, these same checks can track whether one image comes closer to doing the work of two scales.
Methods and sources
Requests. World Labs Atlas (Marble 2) developer API, beta, October 20, 2026. Each request supplied input images, cameras over flat ground, the fixed prompt, and target cameras. Depth for the coverage tints came from the descents and the both-images run with seed 2.
Scores. Layout match is the Pearson correlation of 32 × 18 block averages, averaged over red, green, and blue channels. We report it before and after the best rescaling (0.8 to 1.25) and shift (up to 160 pixels). Fine-detail match correlates brightness residuals after subtracting a Gaussian blur with a 25 cm standard deviation. Sharpness compares the strength of those residuals with the real view. The baseline is a bicubic enlargement of the neighborhood image.
Coverage. Pixels with returned depth, in meters, are projected into the input images. “Coarse only” means one input pixel spans more than 3 × 3 generated pixels; “neither image” means the projection falls outside the inputs; “no depth returned” means the projection cannot be made.
Hand counts. We counted huts and rectangular roofs by eye in the generated and real views.
Follow-up: camera height or image content?
We kept both input cameras and replaced the household image with either the neighborhood image’s center enlarged to the household frame or a real 5 cm-per-pixel image of an empty field outside all target views. Both replacements yielded scale errors as low as those from the two real images.
The real household image improved detail at the study compound: fine-detail correlations reached 0.83 to 0.88, compared with 0.27 for the blurred replacement. Neither replacement produced the crisp, incorrect texture at compounds 3 and 4, supporting the interpretation that Atlas transferred it from the household image.
| Input given to Atlas | Mean / maximum scale error | Views drawn too large / too small | Layout match, as drawn | Sharpness at compounds 3 and 4 |
|---|---|---|---|---|
| Neighborhood image only | 8.2% / 22.5% | 9 / 0 of 10 | 0.71 | 0.23 to 0.31 |
| Both real images | 1.5% / 6.3% | 2 / 1 of 10 | 0.87 | 0.75 to 1.09 |
| Neighborhood image plus blurred household frame | 0.8% / 2.0% | 4 / 0 of 10 | 0.93 | 0.17 to 0.27 |
| Neighborhood image plus field frame (one seed) | 1.2% / 4.1% | 2 / 0 of 5 | 0.94 | 0.16 to 0.28 |
Imagery. “Terego Azua” by OpenStreetMap Uganda, uploaded to OpenAerialMap by Ronald Oyom, acquired August 7, 2021. Licensed CC BY 4.0.
Paper. F. W. Zhu, C. T. Jerzak and A. Daoud, “Optimizing Multi-Scale Representations to Detect Effect Heterogeneity Using Earth Observation and Computer Vision: Applications to Two Anti-Poverty RCTs,” Proceedings of the 4th Conference on Causal Learning and Reasoning (CLeaR 2025), PMLR 275:894–919. PDF · Publication.