Satellites capture far more imagery than their radio links can send to Earth. A recent fix is generative compression: the satellite sends a tiny code, and a large diffusion model on the ground expands it back into a full image. A diffusion model is a generator that produces an image by repeatedly removing noise, starting from pure static. At very low bitrates, these decoders invent detail that was never there, such as waves on a calm sea or altered building outlines. The ground station, however, already holds something the decoder ignores: a full-resolution archive of earlier, cloud-free images of the same place. The satellite only has a heavily shrunk copy of that archive, because the uplink to the satellite is so slow.
We propose to start the ground decoder from a partly noised copy of the archived image rather than from pure noise. The starting noise level is chosen region by region: unchanged areas start shallow and stay pinned to verified pixels, while regions where the received code disagrees with the archive start deep and are rebuilt from the downlinked code. The decoder then also outputs a provenance map that labels each pixel as "carried from the archive image of date X" or "synthesized from this pass's downlink." This idea is "temporal warm-starting": seeding an iterative solver with the previous solution. It comes from robotics, where its known weakness is a pull toward old solutions. On the ground, that same pull becomes the property we want, provided we can show it does not erase genuine changes.
Core Claim
Earth observation compression has an asymmetry that current methods do not exploit. Rich side information, the full-resolution archive, exists only at the decoder. Earth+ shares references across a constellation through a ground loop. It does differencing onboard against references shrunk 2601× to fit a 250 kbps uplink, and it sends changed 64×64 tiles as JPEG-2000. Its onboard reference is therefore far poorer than what the ground holds. COSMIC pairs a light onboard encoder with a ground latent diffusion model that is conditioned on sensor metadata. It decodes every tile from pure noise and acknowledges hallucination below 0.1 bits per pixel.
We claim three things. First, a region-wise warm start from the ground archive lowers hallucination in unchanged regions at a fixed bitrate. Second, it cuts the number of denoising steps. Third, when the onboard encoder is fine-tuned knowing the decoder will warm-start, the encoder learns to spend bits mainly on change. That last point is a learned form of coding with decoder-only side information, which neither COSMIC nor Earth+ attempts.
The technique comes from RTI-DP. That method warm-starts a diffusion policy from its previous shifted prediction and denoises in 1–3 steps, cutting Push-T latency from 816 ms to 25 ms. It also gives local contraction conditions, meaning that errors shrink when the starting guess is close to the answer. Two papers constrain the design. RTI-DP itself breaks on discontinuous jumps, its gripper open/close case. CoDiG reports that warm starts bias samplers toward historical paths. In imagery, a new building or a flood is exactly such a jump. A uniform warm start would paint the past over it, which is why the start depth must vary by region.
Mechanism
The ground station has three inputs: the received latent code $\hat z$, the archived reference image $r$, and the frozen onboard encoder $E$. It computes a per-region disagreement $d(p) = \lVert \hat z(p) - E(r)(p) \rVert$ in the encoder's own latent space. Because the ground can encode $r$ itself, this costs no extra downlink bits. A monotone map turns disagreement into a start depth $s(p) \in [s_{\min}, K]$, where $K$ is the full number of diffusion steps. A standard diffusion model expects one noise level per image, so we run the schedule from $\max_p s(p)$ for the whole tile. At each step $\tau$, every region whose target depth has not yet been reached is overwritten with the reference noised to level $\tau$:
$$x_\tau(p) \leftarrow \sqrt{\bar\alpha_\tau}\, E(r)(p) + \sqrt{1-\bar\alpha_\tau}\,\epsilon \quad \text{for } \tau > s(p),$$
Here $\bar\alpha_\tau$ is the noise schedule's signal fraction and $\epsilon$ is fresh Gaussian noise. This is inpainting-style blending, so every intermediate state stays at a single noise level the model was trained on. Below each region's depth, the decoder denoises freely, conditioned on $\hat z$ and metadata as in COSMIC. The final $s(p)$ map is the provenance record.
Research Program
The work has three stages. Stage one is training-free. We reimplement a COSMIC-style encoder and metadata-conditioned latent diffusion decoder, then add region-wise warm-starting at inference. The aim is to measure fidelity, step count, and change preservation as functions of the disagreement-to-depth map.
Stage two fine-tunes the encoder's rate term while the decoder warm-starts. If decoder-side side information is useful, bits per pixel should fall in unchanged areas without loss inside changed areas. The encoder never sees the full-resolution reference, so onboard compute does not grow.
Stage three stress-tests nuisance change, the failure Earth+ documents: snowmelt, crop phenology, sun-angle shadows. We test whether shallow warm starts plus metadata conditioning absorb appearance shifts that tile differencing would flag. As a cheap appearance correction before warm-starting, we will also adapt the content/style split of Learned Representation of Satellite Image Series. That paper re-renders a historical scene under the current pass's illumination with a feed-forward generator, saving 44.22% bitrate over HEVC on Jilin-1.
Baselines
The naive reference point is the cold-start COSMIC-style decoder at matched bitrate and full step count. This is exactly the system we modify, so any gain is attributable to the warm start. COSMIC reports beating JPEG-2000, CDC and COLIC on LPIPS and FID below 0.1 bpp, and 3–4% higher fMoW land-use accuracy than JPEG-2000. The file gives no numeric LPIPS or FID value, so we will report our reimplementation's numbers next to its JPEG-2000 baseline. Beating it means lower LPIPS and FID and higher change-region PSNR at equal bpp, with no loss in fMoW accuracy.
The strongest temporal competitor is Earth+, on the Washington State Sentinel-2 setting it was evaluated on. It reports 1.3–3.3× downlink reduction over Kodan and SatRoI at equal PSNR, rising to 10× in a 16-satellite constellation. We will run it with the same reference ages and cloud masks. Beating it means lower total downlink bytes at equal PSNR, and fewer bytes spent on nuisance-flagged tiles. For satellite video, the anchor is the Learned Representation method's 44.22% saving over the HEVC main profile on Jilin-1. We will compare at equal bitrate.
The ablation that isolates our key ingredient is a uniform warm start at a single depth, plus an oracle variant whose depth map comes from ground-truth change masks. If uniform matches region-wise, the disagreement map adds nothing. All neural variants share the same decoder weights, training data, and a fixed GPU-hour budget, and every method gets the same hyperparameter-search budget.
Why not just…?
Why not just condition the diffusion decoder on the reference image? Concatenating $r$ as an extra input and retraining is the most serious threat. It may match or beat us on LPIPS. We will train it at equal compute and compare. Our expected advantages are narrower and testable. A conditioned network gives no guarantee about where past content enters, so we can measure its "past leakage" into changed regions directly. It also yields no provenance map, and it requires retraining for every reference-handling change. If conditioning beats warm-starting on fidelity and does not leak more past content into changed regions, we will report that. In that case the contribution shrinks to the provenance map and the encoder-rate result, which can sit on top of a conditioned decoder.
Why not just send Earth+ changed tiles and paste the archive elsewhere on the ground? This reuses existing pieces. Its weaknesses are documented in Earth+ itself: rigid tile thresholds trip on phenology and shadows, and pasted tiles leave seams. Our nuisance stress test and the COSMIC Tile Test Set seam evaluation compare the two directly.
Why not just distill the decoder to one step for speed? Speed is a secondary benefit here. Distillation does nothing about hallucination, which is the main claim. RTI-DP reports matching distilled policies without retraining. We will include a few-step cold-start decoder as a further check, to separate "fewer steps" from "better starting point."
Proposed Experiments
We will evaluate on multi-temporal fMoW sequences, which supply metadata, references and classification labels. We add Earth+'s Sentinel-2 Washington data for nuisance change, and Jilin-1 clips for short-interval frames. Because labeled change is scarce (the embedding-uplink benchmark has only 8 change pairs), we also inject controlled changes by pasting buildings or flood extents into later passes. This gives exact change masks.
The decisive check is change preservation. Inside injected or real changed regions, warm-started reconstructions must match cold-start PSNR and keep the change detectable. We will verify detectability with a fixed change detector and with the time-preference test from the embedding-uplink paper. Outside changed regions, warm-starting must lower hallucination, measured as LPIPS and PSNR against the true current image. We will sweep reference age from days to the 51-day single-satellite regime that Earth+ reports. The aim is to locate the age at which warm-starting stops helping, which bounds when a constellation reference loop is required.
References
- Earth+: On-Board Satellite Imagery Compression Leveraging Historical Earth Observations
- COSMIC: Compress Satellite Image Efficiently via Diffusion Compensation
- Real-time Iteration Scheme for Diffusion Policy
- Constraint-Aware Diffusion Guidance for Robotics: Real-Time Obstacle Avoidance for Autonomous Racing
- Learned Representation of Satellite Image Series for Data Compression
- EMBEDDING-ONLY UPLINK FOR ONBOARD RETRIEVAL UNDER SHIFT IN REMOTE SENSING