EO-TTS: Earth Observation Test-Time Scaling

George Plessias (@Redmonkeycloud) · supervised by Ioannis Kakogeorgiou (@gkakogeorgiou)

MSc thesis, MSc in Artificial Intelligence, University of Piraeus and NCSR "Demokritos"

TerraDiT-Ω-XL, 100 Euler steps, same tiles and noise: fp32 (top) vs bf16 autocast (bottom). Mean PSNR 54.2 dB; bf16 is 1.7× faster on an RTX 3060 (6.3 vs 10.8 s per image).

EO-TTS studies test-time scaling for Earth observation diffusion transformers: inference-only search (Best-of-N, then particle-based search with Rollover Budget Forcing) over pretrained satellite image generators, guided by remote-sensing verifiers.

Base model

All experiments build on TerraDiT-Ω-XL by Wei, Sastry, Cher, Xing and Jacobs (Multimodal Vision Research Lab, Washington University in St. Louis). Their code is pinned here as a git submodule at commit 8d90f0d and used unmodified.

	Link
Paper (TerraDiT-Ω, ECCV 2026)	arXiv:2606.31029 · project page
Paper (TerraDiT-α/Σ)	arXiv:2603.02172 · project page
Code	github.com/mvrl/TerraDiT
Weights	huggingface.co/MVRL/TerraDiT
Derived data and test splits	huggingface.co/datasets/MVRL/TerraDiT-data
Imagery (Git-10M)	huggingface.co/datasets/lcybuaa/Git-10M

If you use this repository, please cite their work:

@inproceedings{wei2026terraditomega,
  title     = {TerraDiT-{\Omega}: Unified Spatial Control for Satellite Image Synthesis with Any Geospatial Primitive},
  author    = {Wei, Brian and Sastry, Srikumar and Cher, Daniel and Xing, Eric and Jacobs, Nathan},
  booktitle = {European Conference on Computer Vision},
  year      = {2026}
}

@article{sastry2026terradit,
  title   = {TerraDiT: Point-Conditioned Diffusion Transformer for Satellite Image Synthesis},
  author  = {Sastry, Srikumar and Cher, Daniel and Wei, Brian and Dhakal, Aayush and Khanal, Subash and Gupta, Dev and Jacobs, Nathan},
  journal = {arXiv:2603.02172},
  year    = {2026}
}

Status

Stage	State
Paper-protocol baseline, TerraDiT-Ω-XL, Git-Rand	done: FID 9.81 vs 9.25 in the paper (details)
Experiment protocol for Best-of-N, fixed before any run	done (protocol)
Best-of-N on 2,000 Git-Rand tiles, N ∈ {1, 2, 4, 8, 16}, 2 verifiers	generating (started 2026-10-30)
Best-of-N on 500 Git-Spatial tiles (held-out region), with fp32 anchor	queued after R2000
Full Git-Spatial baseline	not run; Spatial is used as a held-out test set for search (why)
Best verifier on the full Git-Rand split, N = 4	after the subset results
Rollover Budget Forcing with custom allocation rules	later

Baseline results

TerraDiT-Ω-XL re-run with the authors' code and weights on the full Git-Rand split (14,932 tiles, 100 Euler steps, no guidance, batch 32, seed 42, fp32):

Metric	Ours	Paper	Ours − paper
FID ↓	9.81	9.25	+0.56
LPIPS ↓	0.3401	0.3438	−0.0037
SSIM ↑	0.2669	0.2665	+0.0004
CLIP score (cosine) ↑	0.2882	0.2812	+0.0070
KID ×10³ ↓	3.98 ± 0.41	not reported	–

The per-image metrics match the paper; FID is 6 % higher. What each metric measures, what the paper leaves unstated, and the likely reasons for the gap (fp16 release weights, seed, software and hardware) are in docs/baseline_reproduction.md. Raw output: results/baseline/omega_xl_random_omega.json. All later experiments are compared against this re-run.

Reproducing the paper baseline

The baseline calls TerraDiT's own generate.py and eval/metrics.py in-process. The only substitution is where imagery comes from: instead of downloading all of Git-10M (394 GB, 787 shards), Git10MSubset keeps exactly the rows a test split uses, checked against the split metadata by img_name, so pixels and captions are byte-identical.

Split	Tiles	Git-10M shards needed	Download
random	14,932	151 of 787	29.3 GB
spatial	14,387	2 of 787	0.3 GB

Shards are fetched one at a time and deleted after extraction; the stored subset is ~0.2 GB per split.

Paper targets (TerraDiT-Ω-XL, condition Ω, from Table 1 and Table S3 of arXiv:2606.31029) are in configs/paper_targets.json; every result file stores them next to our numbers.

Split	FID ↓	LPIPS ↓	SSIM ↑	CLIP (cos) ↑
random	9.25	0.3438	0.2665	0.2812
spatial	17.54	0.3614	0.3421	0.2761

Differences between the paper's stated protocol and the released code, recorded in each result:

• Batch size: the paper says 32, the repo defaults to 16. Noise is drawn per batch, so this changes individual samples, not the distribution. Use --batch-size 32 if it fits.
• Seed: not reported in the paper; we use the repo default, 42.
• CLIP: the repo's metrics.py reports the CLIP logit (cosine × 100); the paper reports cosine. Results carry both clip_score and clip_cos.
• sFID: in the paper, not computed by the repo; not reported here.
• KID: not in the paper; added because it is unbiased on small subsets (Best-of-N).

Setup (Windows 11, NVIDIA GPU)

git clone --recursive https://github.com/Redmonkeycloud/eo-tts.git
cd eo-tts
conda create -n eotts python=3.11 -y
conda activate eotts
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124
pip install -e "third_party/TerraDiT[eval]"
pip install -e ".[dev]"
huggingface-cli login            # optional, avoids Hub rate limits on the 29 GB download
python scripts/00_check_env.py

If you cloned without --recursive: git submodule update --init.

Run order

python scripts/01_download_assets.py                        # omega_xl weights + test split files (~2.5 GB)
python scripts/02_build_git10m_subset.py --split spatial    # ~0.3 GB, about a minute
python scripts/03_benchmark.py                              # speed, VRAM, fp32 vs bf16 (~10 min)
python scripts/04_run_baseline.py --split spatial --limit 64   # smoke test
python scripts/02_build_git10m_subset.py --split random     # ~29 GB download, resumable
python scripts/04_run_baseline.py --split random            # the paper number (long run, resumable)
python scripts/04_run_baseline.py --split spatial

Generation resumes from the images already on disk, so an interrupted run is restarted with the same command. Results land in runs/baseline/<run>/results.json.

Best-of-N run order

Fixed in advance by docs/experiment_protocol.md. GPU hours are on an RTX 3060.

python scripts/05_generate_bon.py --set S500 --limit-tiles 2        # smoke test, a few minutes
python scripts/05_generate_bon.py --set R2000                       # 32,000 samples, bf16, ~56 h
python scripts/05_generate_bon.py --set S500                        # 8,000 samples, bf16, ~14 h
python scripts/05_generate_bon.py --set S500 --anchor-fp32          # sample 0 in fp32, same noise, ~1.6 h
python scripts/06_score_bon.py --run R2000
python scripts/06_score_bon.py --run S500
python scripts/06_score_bon.py --run S500_fp32_k0
python scripts/06_score_bon.py --run R2000_baseline_fp32            # the fp32 baseline images on R2000 (anchors 1-2)
python scripts/07_analyze_bon.py                                    # -> results/bon/report.md

Every step resumes after an interruption. Each sample has its own seed, derived from (split, tile, sample index), so a sample is identical however and whenever it is generated.

Layout

configs/paper_targets.json      paper numbers we compare against
docs/baseline_reproduction.md   baseline results, metric definitions, gap analysis
docs/experiment_protocol.md     Best-of-N questions, tiles, verifiers, endpoints, statistics (pre-registered)
configs/subsets/                fixed evaluation tile lists and their representativeness statistics
eotts/data/subsets.py           how the subsets are drawn and checked
eotts/bon/                      Best-of-N: sets and seeds, generation, selection rules, analysis, report
eotts/scoring/                  verifiers (V1, V2) and evaluators (E1-E5)
eotts/stats/                    KID from features, paired bootstrap, Holm
results/                        result files of finished runs (images stay local)
eotts/data/git10m_subset.py     exact local subset of Git-10M (shard math, verify, read)
eotts/baselines/                paper-protocol baseline runner
eotts/eval/kid.py               KID (extra metric for small subsets)
eotts/provenance.py             code versions, environment, GPU, written into every result
scripts/00..07                  the steps above, in order
third_party/TerraDiT            upstream code, pinned, unmodified
tests/                          subset builder, statistics and selection tests (no GPU, no network)

Licences

Code licence for this repo: to be decided. TerraDiT code is Apache-2.0; its weights are CC-BY-NC-4.0 and Git-10M imagery is CC-BY-NC-ND-4.0, so this work is non-commercial research only. Git-10M data is never redistributed by this repo; it is downloaded from the original repository.