ChitraMaya vs. Jasna - same machine, same models, head to head
Same 2h41m 1080p60 file as the earlier benchmark table (578,704 frames). Everything matched as closely as the two tools allow: same detection model (lada_mosaic_detection_model_v4_fast @ threshold 0.25), same restoration model (lada_mosaic_restoration_model_generic_v1.2 - byte-identical file), Max Clip 90, FP16, TensorRT engines compiled on the machine by each tool, RTX Super-Res 4x secondary on both, HEVC at quality 18. Hardware is deliberately the budget case: the ~$100 ex-mining RTX 3060 Mobile 6GB, on a Xeon W-1250 with 8GB RAM.
Notes:
Both tools were run in good faith at their best known configuration on this machine.
More benchmarks here
Codeberg discussion with screenshots here
Same 2h41m 1080p60 file as the earlier benchmark table (578,704 frames). Everything matched as closely as the two tools allow: same detection model (lada_mosaic_detection_model_v4_fast @ threshold 0.25), same restoration model (lada_mosaic_restoration_model_generic_v1.2 - byte-identical file), Max Clip 90, FP16, TensorRT engines compiled on the machine by each tool, RTX Super-Res 4x secondary on both, HEVC at quality 18. Hardware is deliberately the budget case: the ~$100 ex-mining RTX 3060 Mobile 6GB, on a Xeon W-1250 with 8GB RAM.
| Tool | Stack | Wall time |
| Jasna 0.10.0 | TRT engines + RTX Super-Res 4x | 2h 36m |
| ChitraMaya 1.61.01 | TRT engines + RTX Super-Res 4x (matched stack) | 2h 08m |
| ChitraMaya 1.61.01 | same + Temporal Fix strength 3 (optional stabilizer) | 3h 07m |
Notes:
- On the matched stack, ChitraMaya finished 28 minutes (~18%) ahead on this hardware.
- Temporal Fix is a ChitraMaya-only stage (temporal stabilization of restored regions, which Jasna does not offer); it costs ~59 minutes on this title. A quality/speed choice, now measured - switch it off for maximum speed.
- The two ChitraMaya runs produced identical statistics (331,330 detected frames, 333,624 restored, in both) - only the clock changed, so the Temporal Fix cost is measured cleanly. The pipeline is deterministic.
- Rate control differs slightly by necessity: ChitraMaya constant QP 18 vs Jasna CQ 18.
- CUDA Sysmem Fallback Policy set per each tool's own recommendation (Jasna: Prefer No Sysmem Fallback, applied per-app).
- Detector footnote: v4_fast and the older v2 model found nearly identical coverage on this movie at the same threshold (331,330 vs 328,657 detected frames, a 0.8% difference).
- Movie ID and full methodology on request (DM).
Both tools were run in good faith at their best known configuration on this machine.
More benchmarks here
Codeberg discussion with screenshots here
