Today, 09:50
I performed a test by colorizing a full film (960x720, fps: 23.976) on my RTX5070 Ti.
I obtained the following results:
Using my previous pipeline:
1) Extraction of 3062 reference images (about 1 ref. every 41 frames): 7 min.
2) Colorization DiT: (3062*4)/60 = 204 min.
3) Encoding with CMNET2 at a speed of 12.8 fps: (125236/12.8)/60 = 163 min.
Total Time: 374 min. (5.58 fps)
So with direct encoding, the colorization speed improves by about 1.5% (mainly due to the skipped frame extraction phase and filesystem image access).
The only problem is that in this way is not possible to see the colorized frames and eventually remove the ones with inconsistent colors.
CMNET2 is already doing a good job in keeping the color consistency, but to obtain a perfect results it is necessary to manually analyze the colored reference frames and eventually remove the ones with inconsistent colors. But in my opinion the film colorized with the direct encoding is already very good, much, much better than using HAVC (DeOldify + DDColor).
Dan
I obtained the following results:
encoded 125236 frames in 22145.44s (5.66 fps), 2572.90 kb/s, Avg QP:24.47
finished after 06:12:32.007Using my previous pipeline:
1) Extraction of 3062 reference images (about 1 ref. every 41 frames): 7 min.
2) Colorization DiT: (3062*4)/60 = 204 min.
3) Encoding with CMNET2 at a speed of 12.8 fps: (125236/12.8)/60 = 163 min.
Total Time: 374 min. (5.58 fps)
So with direct encoding, the colorization speed improves by about 1.5% (mainly due to the skipped frame extraction phase and filesystem image access).
The only problem is that in this way is not possible to see the colorized frames and eventually remove the ones with inconsistent colors.
CMNET2 is already doing a good job in keeping the color consistency, but to obtain a perfect results it is necessary to manually analyze the colored reference frames and eventually remove the ones with inconsistent colors. But in my opinion the film colorized with the direct encoding is already very good, much, much better than using HAVC (DeOldify + DDColor).
Dan

