How asebaq/rsdiff works
RSDiff is a text-conditioned remote-sensing image generator. Its documented reference workflow trains a 128² low-resolution diffusion model on RSICD, trains a 256² super-resolution stage using the base model, samples caption-conditioned images, then evaluates generated images. The legacy Imagen implementation is the documented working path; the newer src/rsdiff trainer and sampler are explicitly unfinished. The graph also preserves documented evaluation capabilities. Some training and evaluation wiring is documented in the README but their scripts were not sampled, so uncertain call links are omitted.
Main entry points: cli.py Evaluation commands · Imagen_text_sr_pytorch.py SR training · sample_grid.py Cascade sampler · fid.py FID metric
Components 18
Reference cascade
- LR training
ddpm/models/Imagen_text_pytorch.py - SR training
ddpm/models/Imagen_text_sr_pytorch.py - Joint fine-tune
ddpm/models/Imagen_text_joint_pytorch.py - UNet builder
src/rsdiff/models/builders.py - Run config
src/rsdiff/training/config.py - Native trainer
src/rsdiff/training/trainer.py - CLI commands
src/rsdiff/cli.py
Dataset and inputs
- RSICD data
Generation outputs
- Cascade sampler
ddpm/sample_grid.py - Model checkpoint
- Generated images
Evaluation
- Evaluation commands
src/rsdiff/cli.py - FID metric
src/rsdiff/eval/fid.py - Zero-shot OA
src/rsdiff/eval/zeroshot_oa.py - CLIP evaluation
scripts/eval - Evaluation results
results
Other components
- Engineer
- Hugging Face Hub
Connections 29
- Engineer → LR training: starts training
- RSICD data → LR training: training data
- LR training → Model checkpoint: saves weights
- Engineer → SR training: starts training
- RSICD data → SR training: training data
- Model checkpoint → SR training: seeds base
- SR training → Model checkpoint: saves weights
- Engineer → Joint fine-tune: starts fine-tune
- RSICD data → Joint fine-tune: training data
- Model checkpoint → Joint fine-tune: loads weights
- Engineer → Cascade sampler: requests samples
- RSICD data → Cascade sampler: provides captions
- Model checkpoint → Cascade sampler: loads weights
- Cascade sampler → Generated images: writes images
- Engineer → CLI commands: runs command
- CLI commands → Run config: loads settings
- CLI commands → Native trainer: dispatches train or sample
- Native trainer → UNet builder: builds model
- Engineer → Evaluation commands: requests metrics
- Evaluation commands → FID metric: runs FID
- Evaluation commands → Zero-shot OA: runs classification
- Evaluation commands → RSICD data: reads split metadata
- Generated images → FID metric: generated set
- RSICD data → FID metric: real image set
- Engineer → CLIP evaluation: runs evaluation
- CLIP evaluation → Generated images: scores images
- FID metric → Evaluation results: records metrics
- CLIP evaluation → Evaluation results: records metrics
- Hugging Face Hub → Model checkpoint: provides weights