BiliSakura/GeoCore-9B-diffusers

Self-contained Diffusers checkpoint for GeoCore-9B: 256×256 optical satellite generation from text, with optional GSD, latitude, and longitude.

Converted from JeonghyeokDo/GeoCore-9B. DiffusionPipeline.from_pretrained loads the DiT, Flux.2 VAE, CLIP-L, T5-XXL, and FlowMatchEulerDiscreteScheduler from this repo (trust_remote_code=True).

Project · Code · Original weights

Prompt

A parking lot full of cars is located next to some trees.

Prompt

Many buildings are located in a commercial area.

Prompt

A roundabout is surrounded by grass.

Prompt

Two tennis courts are semi-surrounded by several green trees.

Demo

256×256, 50 Euler steps, guidance_scale=4.0, seed 42, lon=126.97, lat=37.56, res=0.0 (Seoul).

Parking lot	Commercial	Roundabout	Tennis courts

Model details

Field	Value
Pipeline	GeoCorePipeline (pipeline.py)
Transformer	Flow-matching DiT, 8 double-stream + 24 single-stream blocks, hidden 4096, 32 heads, 9.24B
Text encoders	CLIP ViT-L/14 + T5-XXL (bundled)
VAE	Flux.2 AutoencoderKLFlux2 (Apache-2.0, bundled)
Scheduler	FlowMatchEulerDiscreteScheduler, shift=1.0
Resolution	256×256
Training data	Git-10M
License	Apache-2.0

Load from Hugging Face

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "BiliSakura/GeoCore-9B-diffusers",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
).to("cuda")

image = pipe(
    prompt="A parking lot full of cars is located next to some trees.",
    lon=126.97,
    lat=37.56,
    res=0.0,
    height=256,
    width=256,
    num_inference_steps=50,
    guidance_scale=4.0,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]
image.save("demo.png")

The DiT is ~18.5 GB in bf16. A 24 GB GPU is tight; on two GPUs, split only the transformer with Accelerate and keep CLIP, T5, and the VAE on one device.

Conditioning

Captions follow the Git-10M style (short scene descriptions). Geospatial fields are optional.

Argument	Meaning
prompt	Natural-language description
res	Resolution index 17 - z for Google XYZ zoom z. 0 ≈ 1.2 m/px at the equator; each +1 doubles GSD
lon, lat	Degrees
-999.0	Learned null embedding — any of res / lon / lat may be omitted

License

DiT weights © 2026 Jeonghyeok Do and Munchurl Kim, Apache-2.0. VAE is an unmodified Apache-2.0 Flux.2 autoencoder from black-forest-labs/FLUX.2-klein-base-4B; see LICENSE-FLUX2-VAE.md. Git-10M is CC BY-NC-ND 4.0 (non-commercial retraining).

Citation

@article{do2026geocore,
  title   = {GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation},
  author  = {Do, Jeonghyeok and Kim, Munchurl},
  year    = {2026}
}