5 Commits
Author SHA1 Message Date
TheSmallHanCat 7773c5415f fix(frontend): stabilize dynamic nodes and catalog refresh 2026-09-30 21:47:26 +08:00
TheSmallHanCat a703206cc3 fix(audio): correct model routing and save edge cases 2026-09-30 21:47:26 +08:00
TheSmallHanCat 14f842eaad fix: harden model selection, stem routing and audio handling
- require exact model matches and preserve links for unavailable custom models
  - restore complete stems from model configs and validate downloaded assets
  - align ensemble channels and batches without dropping audio data
  - stream uploads safely and restore localized model names
  - add Python and frontend regression coverage
2026-09-30 21:47:26 +08:00
TheSmallHanCat 1760a18a6f fix(audio): resample separation inputs to model sample rate (#4)
- resample ComfyUI audio before MSS and VR inference
  - return separated stems with the model-configured sample rate
  - add 48 kHz to 44.1 kHz regression coverage
  - document sample-rate behavior for downstream nodes
2026-09-30 21:45:44 +08:00
TheSmallHanCatandSucial 3a562861e2 Delegate custom model architecture detection to pymss (#3)
* feat: delegate custom model architecture detection to pymss

Default custom separation nodes to model_type=auto while preserving saved manual selections. Delegate catalog detection and runtime selection to the shared upstream implementation. Keep detection errors in ComfyUI's standard execution error flow and cover config preservation, explicit overrides and real inference.

Requires publication of pymss 2.1.6 and pymss-core 0.1.9 before standard dependency installation is available. Validation: 8 node tests passed, plus full PolarFormer weights through CPU and CUDA node execution.

* fix: expose conformer architectures in custom model nodes

* Update README

Removed BS PolarFormer section and clarified existing workflows.

---------

Co-authored-by: Sucial <1584846096@qq.com>
2026-09-22 22:35:41 +08:00