GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation".
Jane: Here's the gist of the paper. GeoCore-9B introduces a 9-billion-parameter generative foundation model trained from scratch exclusively on Earth Observation data, which is the first of its scale to be trained from scratch on EO data The main contributions of our work are summarized as follows Model Architecture and Conditioning GeoCore-9B is built upon a Flow Matching-based Diffusion Transformer (DiT) and natively conditions generation on text descriptions and continuous geospatial metadata, including ground sample distances, latitudes, and longitudes The model operates in the latent space of a pre-trained VAE To condition generation on a text description d, we use a global text embedding Tg(d) from CLIP-ViT-L/14
31: and a token-level text embedding Tl(d) from T5-XXL
32: as cg = Tg(d) ∈ R 1×cg, zl = Tl(d) ∈ R M×cl Given a GSD r, latitude ϕ, and longitude λ, we map each scalar with a sinusoidal projection Φ(·) and combine them to obtain the geospatial context vector cgeo as cgeo = MLPr (Φ(r)) + MLPϕ (Φ(ϕ)) + MLPλ (Φ(λ)) Then, we obtain a global conditioning vector c by adding cgeo, the timestep embedding ct, and the global text embedding cg as c = cgeo + ct + cg This design avoids reliance on natural image priors and enables geo-aware EO synthesis Training Stabilization Mechanism To overcome the convergence and spatial disorientation challenges of training at this scale, we introduce a Geospatial Semantic Alignment loss, a training-only alignment objective This objective distills structural Earth surface priors from a frozen specialist teacher network Specifically, we use a frozen DINOv3-Sat
39: as a satellite-specialist teacher and align intermediate DiT features with its dense structural representations The Geospatial Semantic Alignment loss is defined as LGSA = Ez0,x,t,C Wproj h(k)θ(zt, t, C) - F(x) 2 This training-only alignment objective distills structural Earth surface priors from a frozen DINOv3-Sat
39: as a teacher network The final training objective combines the Flow Matching loss (LFM) and the GSA loss as Ltotal = Ez0,z1,t,C h∥vθ(zt, t, C) − (z1 − z0)∥2 2 i LFM +µLGSA Empirical Validation and Performance We demonstrate practical downstream transferability by adapting GeoCore-9B to highly challenging tasks such as cloud removal and SAR-to-optical translation Beyond standard proxy generative tasks, we show that GeoCore-9B can be effectively adapted for practical EO applications, including highly challenging tasks such as cloud removal and SAR-to-optical cross-modal translation For cloud removal on Sen2-MTC
10: , GeoCore-9B outperforms task-specific cloud removal methods in PSNR and SSIM, while achieving LPIPS comparable to the strongest specialist baseline
26, 58: For SAR-to-optical image translation on QXS-SAROPT
11: , GeoCore-9B achieves the best FID and HF-SCC, and remains comparable to the task-specific C-DiffSET baseline
2: Ablation and Transferability Analysis We evaluate the effect of Geospatial Semantic Alignment (GSA) by training a variant without the GSA objective, i.e., µ = 0 As shown in Fig.
Tom: 5, removing GSA leads to fragmented textures and distorted boundaries The inclusion of GSA loss (bottom row) improves structural fidelity and geometric consistency compared to the baseline without alignment (top row), which suffers from distorted boundaries and fragmented patterns In Table 4, the task w/ GSA (µ = 0.5) shows better performance in RSICD text-to-image compared to w/o GSA (µ = 0) The inclusion of GSA loss improves structural fidelity and geometric consistency compared to the baseline without alignment (top row), which suffers from distorted boundaries and fragmented patterns Limitations and Future Work GeoCore-9B leaves several directions for further extension First, although we validate its transferability on practical tasks such as cloud removal and SAR-to-optical translation, broader evaluations remain as future work for additional real-world applications such as pan-sharpening, super-resolution, and segmentation-conditioned generation Second, GeoCore-9B adopts a pre-trained VAE for efficient latent-space training for which future work will explore specialized latent learning for satellite imagery by jointly optimizing the VAE encoder with the DiT backbone Finally, GeoCore-9B can be extended to handle multispectral satellite imagery based upon the large datasets with geo metadata and rich text prompts The VAE is also an information bottleneck whose reconstruction quality upper-bounds input-faithful detail We presented GeoCore-9B, a 9-billion-parameter generative foundation model trained from the scratch for Earth Observation Built upon a Flow Matching-based Diffusion Transformer, GeoCore9B conditions generation on text and geospatial metadata, reducing reliance on natural image priors To stabilize training at this large scale, we introduced a Geospatial Semantic Alignment loss, which distills structural Earth surface priors from a frozen DINOv3-Sat teacher network during training without adding inference overhead Experiments show that GeoCore-9B achieves strong geo-aware generation capability and can be efficiently adapted to practical downstream tasks, including cloud removal and SAR-to-optical translation These results suggest that large-scale generative pre-training on satellite RGB data provides a promising foundation for geo-aware remote sensing generation A Broader Impacts On the positive side, our generative foundation model can substantially advance Earth Observation (EO) applications by providing high-quality data augmentation for scarce regions, and by enhancing downstream tasks such as environmental monitoring, disaster response, and urban planning Conversely, the ability to generate highly realistic, geo-aware synthetic satellite imagery introduces the risk of creating geographical deepfakes If misused, such technology could be exploited to generate disinformation regarding geopolitical events, natural disasters, or environmental conditions To mitigate these negative societal impacts, we emphasize the necessity of developing robust synthetic image detection frameworks tailored specifically for satellite imagery and advocate for the responsible deployment and disclosure of generative EO models Validation of Pre-trained VAE on EO Data Under the submitted frozen-VAE encode–decode protocol, a random set of 100,000 Git-10M
23: images gives 31.48 dB PSNR, 0.9310 SSIM, and 0.710 high-pass spatial correlation coefficient (HF-SCC) These averages indicate strong reconstruction fidelity and substantial preservation of high-frequency structure at 256 × 256 resolution The positive HF-SCC values, ranging from 0.416 to 0.
Title and authors: Tom: So, we’re looking at this paper called "GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation <ref:2608.01896#pg1,GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation>." It’s about building a nine billion parameter generative model from scratch using only Earth Observation data.
Jane: That sounds huge, Tom. What is the main problem they are trying to solve with this model? What makes standard models struggle with satellite images?
Tom: Well, the paper says that existing generative models for Earth Observation mostly rely on fine-tuning natural image priors. And that introduces these perspective biases that just don't match the constraints of geospatial data.
Lu: Exactly. The core idea here is to train this model from scratch exclusively on EO data to avoid those biases inherent in using natural image priors, which the authors call a major problem for scalability >
Meng: But training from scratch at that scale, nine billion parameters? That sounds like it’s going to take an enormous amount of time and compute power. What are they doing technically to make that training feasible?
Tom: They built it on a Flow Matching-based Diffusion Transformer, which is the architecture they used for the generation part. And instead of just looking at the image, it conditions generation directly on text descriptions and continuous geospatial metadata like ground sample distances, latitudes, and longitudes >
Jane: So they are tying the visual generation process directly to location information. That’s interesting because it means if you tell it to generate something in a specific area with certain ground sample distances, it can actually incorporate that into what it creates >
Lu: And they handle the conditioning by using global text embeddings from CLIP-ViT and token-level text embeddings from T5 XXL. Then they combine those with a geospatial context vector made from mapping scalar values like the ground sample distance, latitude, and longitude using sinusoidal projections >
Tom: That combination creates a global conditioning vector called cgeo. They then add that to the timestep embedding and the global text embedding to get their final conditioning vector c >
Title and authors: Meng: It sounds complicated just putting all those vectors together. What’s the specific mechanism they use for alignment, since training at this scale causes spatial disorientation issues?
Lu: That's where they introduce something called a Geospatial Semantic Alignment loss. It's a training-only alignment objective designed to distill structural Earth surface priors from a frozen specialist teacher network >
Jane: A frozen teacher network? Like what is that, Lu? Is it another model they already trained, or something specific to satellites?
Lu: They use a frozen DINOv3-Sat as the satellite specialist teacher. The loss is defined as L GSA equals Ez0,x,t,C Wproj h(k)θ(zt, t, C) minus F(x) squared >
Tom: So they are forcing the intermediate features of their Diffusion Transformer to align with the dense structural representations learned by that frozen satellite model. That’s a clever way to get structural fidelity without adding any inference overhead during generation >
Meng: If you remove that loss, what happens? Does the model just generate blurry stuff, or something else breaks at scale? We need to know what happens when you take away this alignment mechanism >
Tom: The paper shows that without the Geospatial Semantic Alignment loss, you get fragmented textures and distorted boundaries in the images generated >
Jane: So so they show that this loss is actually necessary for structural consistency. It seems like it’s not just a nice addition but something crucial for making sure the output looks geographically sound >
Lu: They also showed that when they include this alignment loss, like with a weight of zero point five, the performance in the RSICD text-to-image task actually gets better compared to training without it >
Meng: That’s good evidence, but I still worry about deployment. What about the practical application? Can this model just be fine-tuned for a specific task like cloud removal or SAR translation easily?
Tom: Yes, they demonstrate strong transferability on tasks like cloud removal and SAR-to-optical translation. For instance, on Sen2-MTC, GeoCore-9B outperforms task specific methods in both PSNR and SSIM > <ref:2608.01896#pg1>
Jane: That’s a big deal for real applications. If it can beat specialized methods there, that means we could actually use this foundation model for things like disaster response or environmental monitoring >
Title and authors: Lu: And for SAR-to-optical translation on QXS-SAROPT, they achieve the best FID and HF-SCC scores and remain comparable to the task specific C DiffSET baseline >
Meng: But I’m still concerned about running this. The paper mentions they are using a pre-trained VAE for efficient latent space training, which is good for speed, but it also introduces an information bottleneck because the reconstruction quality upper bounds input faithful detail >
Tom: They do flag that limitation clearly. The VAE is an information bottleneck whose reconstruction quality upper bounds input faithful detail >
Jane: So, while the model generation part is powerful, the way they train it through that VAE means you can't get perfect fidelity back into the generated image if your starting point isn't good enough >
Lu: They also suggest future work on specialized latent learning for satellite imagery by jointly optimizing the VAE encoder with the DiT backbone, which is a direction to explore >
Meng: And they mention extending it to handle multispectral satellite imagery using large datasets and rich text prompts, but that’s still future work because of the VAE bottleneck issue >
Tom: So, to wrap up on GeoCore-9B: it’s a nine billion parameter model trained from scratch on EO data, conditioned on text and location metadata, stabilized by a Geospatial Semantic Alignment loss against a frozen teacher network >
Jane: It shows that large scale generative pre-training specifically for satellite RGB data provides a solid foundation for geo-aware remote sensing generation >
Lu: That is the main contribution: combining the Flow Matching-based DiT with this novel alignment objective to impose structural Earth surface priors directly into the latent trajectory during training >
Meng: From a deployment standpoint, it’s promising because of the performance gains on cloud removal and SAR translation, but we need more testing on real traffic scenarios before we can say it’s ready for production >
Tom: That's what I think. It’s a very strong foundation for anyone working in remote sensing generation right now >
The paper's summary: Tom: So, GeoCore-9B is this massive nine billion parameter generative model trained from scratch using only Earth Observation data, and the main takeaway is that it can generate highly realistic and structurally accurate satellite imagery when you condition it on both text descriptions and specific geospatial details like latitude, longitude, and ground sample distance <ref:2608.01896#pg1>.
Jane: It’s a big deal because it means we don't have to rely on those natural image priors that standard models use. Instead, this model learns what actually exists in Earth Observation data, which avoids those perspective biases when it synthesizes new scenes.
Lu: The authors tackle the training challenge by introducing this Geospatial Semantic Alignment loss. They distill structural information from a frozen satellite specialist network directly into the diffusion process during training without adding any extra work for inference later on.
Meng: So they’re trying to get structural fidelity built in from the start, which is smart because those generated images often look messy or have weird boundaries when you push these models to their limits.
Tom: The experimental results show this works practically. They tested it on tough tasks like cloud removal and SAR-to-optical translation, and GeoCore-9B actually beats task specific methods in things like PSNR and SSIM for cloud removal on the Sen2-MTC data set <ref:2608.01896#pg1>.
Jane: That’s where the real impact is for people working in remote sensing. If you're trying to monitor environmental changes or plan disaster responses, having a model that understands geometry and location constraints is a huge step up from using models trained on general photos.
Lu: The paper also shows how important that alignment loss is by comparing it to a version without it, which just creates fragmented textures and distorted boundaries in the generated outputs.
Meng: I see the trade-off then. You get this strong structural consistency, but you have to spend time training that alignment loss correctly, and you still have those limitations we talked about regarding what fidelity you can expect from the VAE bottleneck they used for latent space training.
Tom: Exactly, so it’s a powerful tool for generation right now because of how well it adapts to these real world tasks, but we still need to figure out how to push that fidelity boundary further without running into those training headaches or the inherent limitations of their current setup.
The paper's improvements: Tom: The authors point out that while GeoCore-9B is solid, there are still big areas where they think they can improve things to make it even better for real world use, especially when we talk about other imaging tasks <ref:2608.01896#pg1>.
Jane: So what are these improvements? Are we talking about more training data or maybe a different way to condition the model on that location information?
Lu: They suggest extending the application beyond just cloud removal and SAR translation into things like pan-sharpening, super resolution, and even segmentation conditioned generation.
Meng: That sounds ambitious because those are pretty complex tasks. What’s the technical hurdle there? How do you train a model to do segmentation when it's already doing image synthesis?
Tom: The authors suggest exploring joint optimization of the VAE encoder with the diffusion transformer backbone specifically for latent learning, aiming to make that latent space more specialized for satellite imagery.
Jane: That addresses some of those earlier concerns about the information bottleneck they mentioned regarding the VAE. So it’s about making sure that compressed representation is really useful for EO data.
Lu: And they also think there’s room to extend the model to handle multispectral satellite imagery, using those large datasets along with rich text prompts, which would open up a whole new class of generation capabilities.
Meng: I still have my reservations about the VAE bottleneck though. If you're optimizing two things at once, you risk one thing getting worse instead of both improving. What is the concrete performance gain they are expecting from that joint optimization?
Tom: They anticipate that by jointly optimizing those components, you can achieve better reconstruction quality while maintaining the structural integrity we saw with the alignment loss.
Jane: So they’re looking for a balance there: better detail in the image without losing that crucial geo-aware context. It seems like they are trying to solve a multi-layered problem all at once.
Lu: The implication is that this model isn't just a static tool for one task; it’s meant to be the foundation for a whole suite of specialized remote sensing AI tools.
Conclusion: Tom: So, to wrap up, GeoCore-9B is this nine billion parameter generative foundation model trained from scratch on Earth Observation data that uses text and location metadata to make super realistic and structurally accurate satellite images without relying on natural image priors <ref:2608.01896#pg1,billion parameter generative foundation model>.
Jane: It’s a massive step because it shows how powerful you can get a generative model if you train it specifically on the data you care about, instead of just general photos.
Lu: The main thing they did was the Geospatial Semantic Alignment loss, which forces the diffusion process to respect real-world Earth surface structure during training.
Meng: I think that structural fidelity is what makes it deployable for certain tasks; if the generated image looks geometrically sound, it’s easier to integrate into a production pipeline.
Lalam: I'm confident in this answer; the evidence shows that this approach effectively stabilizes training at this massive scale and provides a strong foundation for geo-aware synthesis.
Tom: Right, so we have a model trained from scratch on EO data, stabilized by an alignment loss against a satellite specialist teacher, which can handle hard tasks like cloud removal and SAR translation better than existing methods.
Jane: It really shows that large scale generative pre-training on satellite RGB data provides a solid foundation for geo-aware remote sensing generation.
Lu: The core contribution is combining the flow matching based diffusion transformer with this novel alignment objective to impose structural Earth surface priors directly into the latent trajectory during training.
Meng: Still, we have to remember that it’s still being used as a starting point; we need more rigorous testing on real traffic scenarios before we can say it's ready for full production deployment.
Tom: That makes sense; there's always a gap between research performance and real world operational reliability.
Jane: It’s exciting stuff, but the road to practical application is still long as we figure out how to manage those latent space constraints effectively.
Lu: We'll keep looking at those joint optimization ideas for the VAE encoder next because that could unlock even higher fidelity in the latent representation.
Meng: And I’m looking forward to seeing what happens when they extend this model to handle multispectral imagery; that opens up a whole new set of data possibilities.
Lalam: We are very sure that GeoCore-9B sets a high bar for foundational models in this domain, and it’s definitely something worth watching closely <ref:2608.01896#pg1>.
Information & Electronics Research Institute · KAIST
cs.CV
Submitted: 2026-08-03
Updated: 2026-11-05
Comments: NeurIPS 2026. Please visit our project page at https://kaist-viclab.github.io/GeoCore-9B_site/
Project page: https://kaist-viclab.github.io/GeoCore-9B_site/Abstract
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 96/100
The gist: GeoCore-9B introduces a 9-billion-parameter generative foundation model trained from scratch exclusively on Earth Observation data…
Key concepts
- Flow Matching-based Diffusion Transformer (DiT)
- This is the core architecture of the model, using a diffusion process combined with a transformer structure. It allows the model to generate complex images step-by-step by gradually denoising random noise into a coherent image. This approach is used to synthesize highly realistic satellite imagery from scratch.
- Geospatial Semantic Alignment loss (LGSA)
- This is a training objective designed to improve structural accuracy. It forces the model's intermediate features to align with dense structural representations learned from a frozen satellite specialist network during training. This process distills essential Earth surface knowledge, ensuring the generated images have correct terrain and boundary structures.
- Geospatial Context Vector (cgeo)
- This vector combines various geospatial inputs—ground sample distance (GSD), latitude, and longitude—using sinusoidal projections. It serves as a crucial conditioning signal that tells the model exactly where in space the generated image should be located, enabling true geo-aware synthesis.
- Satellite-specialist teacher (DINOv3-Sat)
- This is a frozen, pre-trained network used only during training to provide structural guidance. It acts as an expert that has already learned the dense structural priors of Earth surfaces from satellite data. By aligning the main model with this teacher, GeoCore-9B learns robust geometric consistency.
Terminology
Summary
GeoCore-9B introduces a 9-billion-parameter generative foundation model trained from scratch exclusively on Earth Observation data, which is the first of its scale to be trained from scratch on EO data The main contributions of our work are summarized as follows
Model Architecture and Conditioning
GeoCore-9B is built upon a Flow Matching-based Diffusion Transformer (DiT) and natively conditions generation on text descriptions and continuous geospatial metadata, including ground sample distances, latitudes, and longitudes The model operates in the latent space of a pre-trained VAE To condition generation on a text description d, we use a global text embedding Tg(d) from CLIP-ViT-L/14 [31] and a token-level text embedding Tl(d) from T5-XXL [32] as cg = Tg(d) ∈ R 1×cg, zl = Tl(d) ∈ R M×cl Given a GSD r, latitude ϕ, and longitude λ, we map each scalar with a sinusoidal projection Φ(·) and combine them to obtain the geospatial context vector cgeo as cgeo = MLPr (Φ(r)) + MLPϕ (Φ(ϕ)) + MLPλ (Φ(λ)) Then, we obtain a global conditioning vector c by adding cgeo, the timestep embedding ct, and the global text embedding cg as c = cgeo + ct + cg This design avoids reliance on natural image priors and enables geo-aware EO synthesis
Training Stabilization Mechanism
To overcome the convergence and spatial disorientation challenges of training at this scale, we introduce a Geospatial Semantic Alignment loss, a training-only alignment objective This objective distills structural Earth surface priors from a frozen specialist teacher network Specifically, we use a frozen DINOv3-Sat [39] as a satellite-specialist teacher and align intermediate DiT features with its dense structural representations The Geospatial Semantic Alignment loss is defined as LGSA = Ez0,x,t,C Wproj h(k)θ(zt, t, C) - F(x) 2 This training-only alignment objective distills structural Earth surface priors from a frozen DINOv3-Sat [39] as a teacher network The final training objective combines the Flow Matching loss (LFM) and the GSA loss as Ltotal = Ez0,z1,t,C h∥vθ(zt, t, C) − (z1 − z0)∥2 2 i LFM +µLGSA
Empirical Validation and Performance
We demonstrate practical downstream transferability by adapting GeoCore-9B to highly challenging tasks such as cloud removal and SAR-to-optical translation Beyond standard proxy generative tasks, we show that GeoCore-9B can be effectively adapted for practical EO applications, including highly challenging tasks such as cloud removal and SAR-to-optical cross-modal translation For cloud removal on Sen2-MTC [10], GeoCore-9B outperforms task-specific cloud removal methods in PSNR and SSIM, while achieving LPIPS comparable to the strongest specialist baseline [26, 58] For SAR-to-optical image translation on QXS-SAROPT [11], GeoCore-9B achieves the best FID and HF-SCC, and remains comparable to the task-specific C-DiffSET baseline [2]
Ablation and Transferability Analysis
We evaluate the effect of Geospatial Semantic Alignment (GSA) by training a variant without the GSA objective, i.e., µ = 0 As shown in Fig. p. 5 To evaluate semantic controllability, we vary the text prompts while fixing the geospatial
metadata. As shown in Fig. ”
Figure 5: Ablation study on the Geospatial Semantic Alignment (GSA) loss. The inclusion of GSA
loss (bottom row) improves structural fidelity and geometric consistency compared to the baseline
without alignment (top row), which suffers from distorted boundaries and fragmented patterns.Where the paper says it
Where the paper says it
Limitations and Future Work
GeoCore-9B leaves several directions for further extension First, although we validate its transferability on practical tasks such as cloud removal and SAR-to-optical translation, broader evaluations remain as future work for additional real-world applications such as pan-sharpening, super-resolution, and segmentation-conditioned generation Second, GeoCore-9B adopts a pre-trained VAE for efficient latent-space training for which future work will explore specialized latent learning for satellite imagery by jointly optimizing the VAE encoder with the DiT backbone Finally, GeoCore-9B can be extended to handle multispectral satellite imagery based upon the large datasets with geo metadata and rich text prompts The VAE is also an information bottleneck whose reconstruction quality upper-bounds input-faithful detail
Conclusion
We presented GeoCore-9B, a 9-billion-parameter generative foundation model trained from the scratch for Earth Observation Built upon a Flow Matching-based Diffusion Transformer, GeoCore9B conditions generation on text and geospatial metadata, reducing reliance on natural image priors To stabilize training at this large scale, we introduced a Geospatial Semantic Alignment loss, which distills structural Earth surface priors from a frozen DINOv3-Sat teacher network during training without adding inference overhead Experiments show that GeoCore-9B achieves strong geo-aware generation capability and can be efficiently adapted to practical downstream tasks, including cloud removal and SAR-to-optical translation These results suggest that large-scale generative pre-training on satellite RGB data provides a promising foundation for geo-aware remote sensing generation
A Broader Impacts
On the positive side, our generative foundation model can substantially advance Earth Observation (EO) applications by providing high-quality data augmentation for scarce regions, and by enhancing downstream tasks such as environmental monitoring, disaster response, and urban planning Conversely, the ability to generate highly realistic, geo-aware synthetic satellite imagery introduces the risk of creating geographical deepfakes If misused, such technology could be exploited to generate disinformation regarding geopolitical events, natural disasters, or environmental conditions To mitigate these negative societal impacts, we emphasize the necessity of developing robust synthetic image detection frameworks tailored specifically for satellite imagery and advocate for the responsible deployment and disclosure of generative EO models
Validation of Pre-trained VAE on EO Data
Under the submitted frozen-VAE encode–decode protocol, a random set of 100,000 Git-10M [23] images gives 31.48 dB PSNR, 0. p. 7 Bold and
underline indicate the best and second-best results, respectively. (a) Cloud Removal
Methods PSNR↑ SSIM↑ LPIPS↓
Task-specific specialist methods
McGAN [5] 17. 48 dB PSNR, 0. 9310 SSIM, and 0. 710 high-pass spatial correlation coefficient
(HF-SCC). These averages indicate strong reconstruction fidelity and substantial preservation of
high-frequency structure at 256 × 256 resolution.Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved AI system can synthesize highly realistic and structurally accurate satellite imagery conditioned on text descriptions and continuous geospatial metadata including ground sample distances, latitudes, and longitudes. This is achieved by training GeoCore-9B from scratch exclusively on EO data to avoid perspective biases inherent in natural image priors.
-
The system incorporates a Geospatial Semantic Alignment loss which
distills structural Earth surface priors (e.g., terrain and urban areas) from a frozen specialist teacher network, constraining the diffusion latent trajectory during training without adding inference overhead.
This ensures improved structural fidelity in generated EO images. -
The system can adapt to highly challenging tasks such as cloud removal and SAR-to-optical cross-modal translation by performing parameter-efficient adaptation using LoRA adapters. These adaptations
outperform or remain competitive compared to task-specific specialist methods
on these practical applications.
Abstract
Existing generative models for earth observation (EO) predominantly rely on fine-tuning natural image priors, which limits their scalability and introduces perspective biases that conflict with geospatial constraints. To address this, we introduce GeoCore-9B, a 9-billion-parameter generative foundation model, which is the first of its scale to be trained from scratch exclusively on EO data. Unlike previous EO foundation models, GeoCore-9B is built upon a Flow Matching-based Diffusion Transformer (DiT) and natively conditions generation on text descriptions and continuous geospatial metadata, including ground sample distances, latitudes, and longitudes. To overcome the convergence and spatial disorientation challenges of training at this scale, we propose a Geospatial Semantic Alignment loss. This objective distills structural Earth surface priors (e.g., terrain and urban areas) from a frozen specialist teacher network, constraining the diffusion latent trajectory during training without adding inference overhead. Pre-trained on the global-scale Git-10M dataset, GeoCore-9B demonstrates strong downstream versatility. Beyond standard proxy generative tasks, we show that GeoCore-9B can be effectively adapted for practical EO applications, including highly challenging tasks such as cloud removal and SAR-to-optical cross-modal translation. Extensive evaluations confirm that GeoCore-9B establishes new state-of-the-art performance in both visual fidelity and geographic structural accuracy.
Sources
- C-DiffSET: Leveraging Latent Diffusion for SAR-to-EO Image Translation with Confidence-Guided Reliable Object Generation
- Classifier-Free Diffusion Guidance
- The QXS-SAROPT Dataset for Deep Learning in SAR-Optical Data Fusion
- FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space
- Flow Matching for Generative Modeling
- Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow
- DINOv3
- Deterministic Guidance Diffusion Model for Probabilistic Weather Forecasting
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think
- PMAA: A Progressive Multi-scale Attention Autoencoder Model for High-performance Cloud Removal from Multi-temporal Satellite Imagery
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs