Latent diffusion with fiddle configuration#16

Open

leifdenby wants to merge 38 commits into

mlcast-community:mainfrom

leifdenby:wip/ldcast-fiddle-refactor

Commits on May 20, 2026

Commits on Jun 3, 2026

feat: add latent diffusion architecture
leifdenby
committed
refactor: split forecasting task modules
leifdenby
committed
refactor: clarify forecasting task module roles
leifdenby
committed
refactor: remove obsolete module aliases
leifdenby
committed
feat: add LDCast two-stage training experiment
leifdenby
committed
docs: add LDCast architecture docs, config diagram, and implementation comparison
leifdenby
committed
refactor: add explicit ensemble dimension throughout forecasting pipeline
leifdenby
committed
refactor: use einops rearrange for ensemble unstack in LatentDiffusionTaskModule
leifdenby
committed
refactor: make fiddlers apply recursively to nested Experiment configs
leifdenby
committed
refactor: move archetype experiments to mlcast.config.archetype subpackage
leifdenby
committed
Merge branch 'main' of https://github.com/mlcast-community/mlcast into wip/ldcast-fiddle-refactor
leifdenby
committed
fix: resolve test failures and align metric naming for LDCast merge
leifdenby
committed
docs: add ldcast-notes.md documenting DMI/Martinbo implementation differences
leifdenby
committed
refactor: align LDCast config with DMI/Martinbo reference
leifdenby
committed
docs: explain why EMA is used for latent diffusion in docstring
leifdenby
committed
docs: simplify LDCast README snippet — single use_random_sampler(cfg) covers both stages via @applies_to_experiments
leifdenby
committed
docs: fix HalfUNet example jaxtyping, einops, and ensemble dim
leifdenby
committed

Commits on Jun 24, 2026

remove DiffusionLoss as separate class
leifdenby
committed