RSDiff: Remote Sensing Image Generation from Text¶

RSDiff: Remote Sensing Image Generation from Text Using Diffusion Model Ahmad Sebaq, Mohamed ElHelw Center for Informatics Science, Nile University Neural Computing and Applications, 2024 [Paper] · [Code] · [Model]

A T5-conditioned cascaded diffusion model for text-to-satellite-image generation at 256×256, trained on RSICD. The released checkpoint reaches FID 65.70 and CLIP-score 0.278 on the full RSICD test split (N=1,093, cond_scale=5).

The cascade is text-conditioned via a frozen T5-base encoder, classifier-free-guided at inference, and split across a 27 M-param 128² base UNet and a 92 M-param 256² super-resolution UNet (119.9 M parameters across the two UNets, 1000-step DDPM). Pre-trained weights are released on the HuggingFace Hub at asebaq/rsdiff-sr-cascade-ep650.

pip install rsdiff
rsdiff sample --prompt "many buildings and green trees are around a playground ." -n 4

All generated imagery is synthetic: it is not an observation or a prediction of any real place.

Headline¶

Headline FID + CLIP

Metric	Value
FID (torchmetrics Inception, feature=2048, 256, N=1093)	65.70
CLIP-score (OpenAI ViT-B/32)	0.278 ± 0.030
CLIP-score (shuffled-caption null)	0.232
CLIP-score delta vs null	+0.046

See Results for the full FID-vs-epoch sweep, CFG-scale ablation, and discussion.

Samples¶

Nine synthetic 256×256 samples from the RSICD test split, generated with the released cascade at cond_scale=5. Captions shown verbatim — no truncation.

sample 0000 — "there's a green pool like an airplane closing the house with a grey roof ."

sample 0096 — "there are many large trees on both sides of the wide road ."

sample 0199 — "the lake cover the most area of the lake ."

sample 0349 — "the waves are crushing on the wet sand ."

sample 0499 — "roads and rivers can be seen in the valleys ."

sample 0649 — "the airport covers a large area and has many planes ."

sample 0799 — "the rectangular pond close to the trapezoidal pond is next to a building with a parking lot surrounded by trees ."

sample 0949 — "some streets divide advertising into several pieces ."

sample 1049 — "white advertising with surrounding trees is next to a main road and some apartments ."

Full 1,093-image generation bundle on the HF Hub release under samples/.

Acknowledgments¶

Built on lucidrains/imagen-pytorch for the cascade scaffolding and the HuggingFace datasets mirror of RSICD. Developed at the Center for Informatics Science, Nile University.