Skip to content

Command line

synapse-sr INPUT.tif OUTPUT.tif [options]

In a terminal you get a live progress bar (stages, tile count, elapsed time and ETA), then a summary table:

                    synapse-sr result
model          synapse-pro-v2
output         640 × 640 px · 2 m · bands B04 B03 B02 B08
backend        pytorch · bfloat16 · tile 32 halo 16
consistency    B04 0.91τ  B03 0.80τ  B02 0.68τ  B08 0.99τ
support        HIGH 22%  MEDIUM 48%  LOW 29%
invalid input  0.0%
time           80.3 s
done wrote OUTPUT.tif

Common tasks

Task Command
Default (Flash, best available device) synapse-sr scene.tif scene_2m.tif
Highest detail (GPU) synapse-sr scene.tif scene_2m.tif --model pro
Offline, local weights synapse-sr scene.tif scene_2m.tif --weights synapse-pro-v2.safetensors
Explicit cloud mask synapse-sr scene.tif scene_2m.tif --scl scene_SCL_20m.tif
No cloud masking synapse-sr scene.tif scene_2m.tif --scl none
Raw DN from baseline ≥ 04.00 without a tag synapse-sr scene.tif scene_2m.tif --offset -1000
Four bands only (for a GIS) synapse-sr scene.tif scene_2m.tif --no-confidence
Low GPU memory synapse-sr scene.tif scene_2m.tif --batch 2
Part of a large tile synapse-sr T43PGQ.tif part.tif --window 2000,3000,1000,1000
A whole folder, four bands, previews synapse-sr scenes/ out/ --no-confidence --preview previews/
Scripts and pipelines synapse-sr scene.tif scene_2m.tif --json (JSON on stdout, no bars; per-file lines of folder runs go to stderr)
Diagnose the machine synapse-sr --env
What models exist synapse-sr --models

All options

Option Default Meaning
--model flash flash, pro, or a registered name such as pro-v1
--weights local .safetensors checkpoint; no network access needed
--device CUDA, then MPS (Flash), then CPU cuda, cuda:1, cpu, mps
--backend, --backend-device Flash accelerators, see Hardware
--window ROW,COL,H,W process only this pixel window
--bounds L,B,R,T process only this area, in the input CRS
--scl auto scene classification raster; auto uses <input>_scl.tif when it exists; none disables
--offset tag or 0 DN offset added before dividing by 10 000
--tile largest that fits memory tile size in 10 m pixels
--halo 16 context around each tile, in 10 m pixels
--batch as many as fit, up to 8, on CUDA / 1 CPU tiles per forward pass
--no-confidence write B04 B03 B02 B08 only (also in folder mode)
--preview PNG save a 10 m / 2 m / support quicklook; a folder of <name>.png in folder mode
--cog Cloud-Optimised GeoTIFF
--fetch LAT,LON, --dates START:END, --size-m, --rank download a scene first, see Data
--json machine-readable summary on stdout
--quiet no progress bar, summary or per-file lines
--env, --models, --version print and exit (--env --json for JSON)

Output file

Bands Content
1-4 B04 B03 B02 B08 surface reflectance, float32, NaN (the declared NoData) where the input was invalid
5-8 ERRSCALE_*: learned per-pixel error scale for each band
9 SUPPORT: 2 observation-determined, 1 medium, 0 prior-dominated or invalid

Same CRS and bounds as the input, with a 2 m pixel.

Exit codes

0 success, 1 an error (unreadable file, unsupported grid, missing weights, a bad --window / --fetch argument, or any failed file in folder mode), 2 a usage error, 130 interrupted. Every error prints a message that says what to change; see Troubleshooting.