FlowSat

FlowSat generates 512 px satellite imagery from a caption and the acquisition metadata that defines how the image was captured — where on Earth, at what ground resolution, on what date, under how much cloud. Most text-to-image models fold such metadata, when they use it at all, into a single undifferentiated vector added to the timestep embedding. FlowSat encodes each field on its own geometry and injects the result through a zero-initialised graft on the pretrained modulation path, so metadata influence is learned rather than imposed. On FMoW-RGB it reaches FID 28.74 in 20 sampling steps, against 35.27 in 100 for the closest prior work.

Interactive

Turn the dial yourself

Every image below was generated from the same caption and the same noise seed. Only one metadata field changes. Drag a slider to see what that field controls.

Comparison

Same prompt, different acquisition

Drag the divider. Both sides share a caption and a seed; only the metadata differs.

static/compare/a.jpg  +  static/compare/b.jpg

⇆

⟳ January July

Method

How it works

Satellite imagery is defined as much by how it was captured as by what it contains. Captions rarely state the season, the sensor resolution or the atmosphere — yet those are exactly what a user wants to control.

The seven fields

Conditioning is seven numbers, normalised to a common range. A caption almost never states any of them.

Encoding each field on the geometry it actually lives on is the point. Coordinates are a position on a sphere, not two independent numbers, so places either side of the dateline stay neighbours. Dates are a circle, so December neighbours January. Ground resolution is a scalar spanning an order of magnitude. Folding all seven into one flat vector throws that structure away before the model sees it.

The resulting embedding does not get its own new pathway into the network. It joins the modulation signal the pretrained backbone already uses, through a projection initialised to exactly zero — so at step 0 the model is byte-identical to the text-to-image checkpoint it started from, and every unit of metadata influence has to be earned by gradient descent rather than imposed by construction.

Flow matching, not denoising

Training does not add noise in steps and learn to undo it. It draws a straight path between a real latent and a sample of noise, xt = (1−t)·x₀ + t·ε, and learns the constant velocity along that path, v = ε − x₀. Generation integrates the velocity back from noise with an Euler solver. Because the path is straight, few steps suffice — 20 rather than the 100 a denoising sampler needs. The backbone is natively flow-matched, so time is used directly in [0, 1] with no timestep rescaling.

Results

FMoW-RGB at 512 px

Identical FID reference statistics and sample set across all models.

Three-seed variance, as submitted: FID 31.53 ± 0.32, CLIP 0.3018 ± 0.0007. The step-count advantage derives from flow matching and the DC-AE/Sana backbone rather than from metadata conditioning; the encoder's own contribution is isolated in the ablation.

Reproducing these numbers

One script measures all four reported metrics under the published protocol — 10,000 FMoW test samples, 20 Euler steps, guidance 2.5, seed 42. It writes the protocol it ran under alongside the results, so two runs can be checked against each other before their numbers are compared.

copypip install -e ".[eval,data]"

python -m flowsat.evaluation.evaluate_fmow \
    --checkpoint     Djisgod/flowsat-fmow-512 \
    --fmow_test_root /path/to/fmow-full/test \
    --caption_root   /path/to/fmow_captions_test \
    --output_dir     evaluations/flowsat

On one A100 the run above takes about 64 minutes and scores 9,999 of 10,000 — one FMoW test GeoTIFF is unreadable and is dropped rather than substituted.

FID against sample count

FID is biased upward at small N, so a number measured at one sample count cannot be compared with a number measured at another. Every row below comes from the same generation pass, read at increasing counts:

CLIP is flat across the sweep; FID is not. A run that stops early reports a worse FID than it earned, which is the single most common reason a reproduction disagrees.

The same run reports the paired reconstruction metrics, SSIM 0.1564 and LPIPS 0.6574. Those score each image against the one real acquisition its caption and metadata came from, so they measure conditioning fidelity rather than image quality: two real FMoW images of the same place reach only SSIM 0.214 / LPIPS 0.425 through the same pipeline.

Samples

Controllability

What makes a model metadata-controllable

Metadata conditioning either works or quietly does nothing, and the difference is rarely the architecture.

A generative model has to explain the variation in its training images, and every conditioning signal competes to do that explaining. Text is cheap: it arrives through a large pretrained encoder, cross-attended at every block. Metadata is expensive: a small learned embedding through a narrow injection path. So when a caption already states what a metadata field encodes, that field is redundant during training — gradient descent routes around it, the pathway starves, and at inference the field looks inert. The model is behaving correctly; there was nothing left for it to explain.

Controllability is inversely proportional to how much your captions cover that field's own domain. The mechanism is overlap, not richness: a 300-word caption about rooftop geometry and pavement texture leaves latitude's variance untouched, while a 30-word "an arid Egyptian site in winter" destroys it.

The caption mixture we train on

The empty bucket is the one people omit. Our earlier 70:20:10 split left only 10% caption-free samples and produced visibly weaker metadata response than 40:30:30.

Why GSD is the most controllable field

Ground sample distance needed no special encoder design, no tuning and no extra loss term. It works because it is nearly orthogonal to everything else in the conditioning: no VLM ever writes "0.5 m per pixel", resolution is uncorrelated with latitude, month or cloud, and its visual effect is monotonic. Nothing else could account for that variation, so the model was forced to use the channel. Latitude, month and cloud cover all struggle for the opposite reason — captions describe country, season, haze and contrast constantly.

Limitations

What the numbers above do not show

Two of the seven fields are demonstrated rather than scored, the paired metrics have a floor well above zero, and FID belongs to a protocol rather than to a model. Worth knowing before comparing anything against this page.

Get started

Generate an image

copygit clone https://github.com/dsp81/flowsat-satellite-image.git
cd flowsat-satellite-image && pip install -e .

python generate.py --prompt "An airport with a long grey runway."

Metadata is optional — sensible defaults are built in. When you want control, every field is a flag:

copypython generate.py \
    --prompt "A farmland with irrigated green fields." \
    --lat 45.0 --lon 5.0 --month 7 --gsd 0.5

Runs on GPU if one is available, CPU otherwise.

Pretrained weights

The released model is v1 checkpoint-125000 — 612,119,200 parameters, fp32, 2.28 GiB. It is the checkpoint every number on this page was measured from.

Download the checkpoint

Or let the evaluation and generation scripts fetch it for you — --checkpoint accepts a Hub repo id as readily as a local path:

copypython -m flowsat.evaluation.evaluate_fmow \
    --checkpoint Djisgod/flowsat-fmow-512 \
    --fmow_test_root /path/to/fmow-full/test \
    --caption_root   /path/to/fmow_captions_test \
    --output_dir     evaluations/flowsat

Use it on your own dataset

FlowSat is not tied to FMoW. Any corpus supplying (image, caption, metadata) triples plugs in through a single adapter class returning three keys:

copy{
    "pixel_values": (3, H, W)   # float, normalised to [-1, 1]
    "input_ids":    (L,)        # tokenised caption
    "metadata":     (7,)        # lon, lat, gsd, cloud, year, month, day
}

Citation

BibTeX

copy@inproceedings{parihar2026flowsat,
  title     = {FlowSat: Flow-Matching Diffusion Transformers with Metadata
               Conditioning for Satellite Image Generation},
  author    = {Parihar, Digvijay Singh and Mondal, Rishabh and Batra, Nipun},
  booktitle = {British Machine Vision Conference (BMVC)},
  year      = {2026}
}