bghira commited on
Commit
b9abc36
·
verified ·
1 Parent(s): 95cbc28

Add Caption v2 training tournament comparisons

Browse files
Files changed (1) hide show
  1. README.md +20 -80
README.md CHANGED
@@ -47,8 +47,7 @@ All renders use a 40-second maximum duration, 30 flow steps, flow CFG 1.7, and s
47
  - **NextLat**: [MiniMaxMusicTraining/soad-mm3-nextlat-20260822](https://huggingface.co/MiniMaxMusicTraining/soad-mm3-nextlat-20260822)
48
  - **NextLat + XM**: [MiniMaxMusicTraining/soad-mm3-nextlat-xm-20260822](https://huggingface.co/MiniMaxMusicTraining/soad-mm3-nextlat-xm-20260822)
49
 
50
- <details>
51
- <summary><strong>No trigger word / off-genre descriptors</strong></summary>
52
 
53
  Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean electric piano, airy male vocals, smooth verse-chorus songwriting, restrained dynamics, polished intimate mix.`
54
 
@@ -100,10 +99,7 @@ Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean elec
100
 
101
  </details>
102
 
103
- </details>
104
-
105
- <details>
106
- <summary><strong>Trigger word only</strong></summary>
107
 
108
  Caption: `system of a down, soad style music`
109
 
@@ -155,10 +151,7 @@ Caption: `system of a down, soad style music`
155
 
156
  </details>
157
 
158
- </details>
159
-
160
- <details>
161
- <summary><strong>Matching descriptors, no trigger word</strong></summary>
162
 
163
  Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
164
 
@@ -210,10 +203,7 @@ Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with lay
210
 
211
  </details>
212
 
213
- </details>
214
-
215
- <details>
216
- <summary><strong>Trigger word + matching descriptors</strong></summary>
217
 
218
  Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
219
 
@@ -267,8 +257,6 @@ Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gent
267
 
268
  </details>
269
 
270
- </details>
271
-
272
  <details>
273
  <summary><strong>Full song without regularisation</strong> (576 renders)</summary>
274
 
@@ -285,8 +273,7 @@ All renders use a 40-second maximum duration, 30 flow steps, flow CFG 1.7, and s
285
 
286
  > **Run note:** These four runs resumed their learned LoRA weights from checkpoint 500 with fresh optimizer, sampler, and RNG state after the original local trainer state was lost.
287
 
288
- <details>
289
- <summary><strong>No trigger word / off-genre descriptors</strong></summary>
290
 
291
  Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean electric piano, airy male vocals, smooth verse-chorus songwriting, restrained dynamics, polished intimate mix.`
292
 
@@ -338,10 +325,7 @@ Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean elec
338
 
339
  </details>
340
 
341
- </details>
342
-
343
- <details>
344
- <summary><strong>Trigger word only</strong></summary>
345
 
346
  Caption: `system of a down, soad style music`
347
 
@@ -393,10 +377,7 @@ Caption: `system of a down, soad style music`
393
 
394
  </details>
395
 
396
- </details>
397
-
398
- <details>
399
- <summary><strong>Matching descriptors, no trigger word</strong></summary>
400
 
401
  Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
402
 
@@ -448,10 +429,7 @@ Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with lay
448
 
449
  </details>
450
 
451
- </details>
452
-
453
- <details>
454
- <summary><strong>Trigger word + matching descriptors</strong></summary>
455
 
456
  Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
457
 
@@ -505,8 +483,6 @@ Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gent
505
 
506
  </details>
507
 
508
- </details>
509
-
510
  <details>
511
  <summary><strong>Random windowing + regularisation</strong> (432 renders)</summary>
512
 
@@ -520,8 +496,7 @@ All renders use a 40-second maximum duration, 30 flow steps, flow CFG 1.7, and s
520
  - **NextLat**: [MiniMaxMusicTraining/soad-mm3-nextlat-20260823-10k-adamw8e-5-randwin128-reg128-nocapdrop](https://huggingface.co/MiniMaxMusicTraining/soad-mm3-nextlat-20260823-10k-adamw8e-5-randwin128-reg128-nocapdrop)
521
  - **NextLat + XM**: [MiniMaxMusicTraining/soad-mm3-nextlat-xm-20260823-10k-adamw8e-5-randwin128-reg128-nocapdrop](https://huggingface.co/MiniMaxMusicTraining/soad-mm3-nextlat-xm-20260823-10k-adamw8e-5-randwin128-reg128-nocapdrop)
522
 
523
- <details>
524
- <summary><strong>No trigger word / off-genre descriptors</strong></summary>
525
 
526
  Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean electric piano, airy male vocals, smooth verse-chorus songwriting, restrained dynamics, polished intimate mix.`
527
 
@@ -569,10 +544,7 @@ Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean elec
569
 
570
  </details>
571
 
572
- </details>
573
-
574
- <details>
575
- <summary><strong>Trigger word only</strong></summary>
576
 
577
  Caption: `system of a down, soad style music`
578
 
@@ -620,10 +592,7 @@ Caption: `system of a down, soad style music`
620
 
621
  </details>
622
 
623
- </details>
624
-
625
- <details>
626
- <summary><strong>Matching descriptors, no trigger word</strong></summary>
627
 
628
  Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
629
 
@@ -671,10 +640,7 @@ Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with lay
671
 
672
  </details>
673
 
674
- </details>
675
-
676
- <details>
677
- <summary><strong>Trigger word + matching descriptors</strong></summary>
678
 
679
  Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
680
 
@@ -724,8 +690,6 @@ Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gent
724
 
725
  </details>
726
 
727
- </details>
728
-
729
  <details>
730
  <summary><strong>Caption v2: training-window comparison</strong> (400 renders)</summary>
731
 
@@ -742,8 +706,7 @@ All renders use a 40-second maximum duration, 30 flow steps, flow CFG 1.7, and s
742
 
743
  > **Run note:** Checkpoint coverage differs by leg; empty table cells were not rendered.
744
 
745
- <details>
746
- <summary><strong>No trigger word / off-genre descriptors</strong></summary>
747
 
748
  Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean electric piano, airy male vocals, smooth verse-chorus songwriting, restrained dynamics, polished intimate mix.`
749
 
@@ -795,10 +758,7 @@ Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean elec
795
 
796
  </details>
797
 
798
- </details>
799
-
800
- <details>
801
- <summary><strong>Trigger word only</strong></summary>
802
 
803
  Caption: `system of a down, soad style music`
804
 
@@ -850,10 +810,7 @@ Caption: `system of a down, soad style music`
850
 
851
  </details>
852
 
853
- </details>
854
-
855
- <details>
856
- <summary><strong>Matching descriptors, no trigger word</strong></summary>
857
 
858
  Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
859
 
@@ -905,10 +862,7 @@ Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with lay
905
 
906
  </details>
907
 
908
- </details>
909
-
910
- <details>
911
- <summary><strong>Trigger word + matching descriptors</strong></summary>
912
 
913
  Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
914
 
@@ -962,8 +916,6 @@ Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gent
962
 
963
  </details>
964
 
965
- </details>
966
-
967
  <details>
968
  <summary><strong>Caption v2: singer-tag comparison</strong> (480 renders)</summary>
969
 
@@ -980,8 +932,7 @@ All renders use a 40-second maximum duration, 30 flow steps, flow CFG 1.7, and s
980
 
981
  > **Run note:** Checkpoint coverage differs by leg; empty table cells were not rendered.
982
 
983
- <details>
984
- <summary><strong>No trigger word / off-genre descriptors</strong></summary>
985
 
986
  Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean electric piano, airy male vocals, smooth verse-chorus songwriting, restrained dynamics, polished intimate mix.`
987
 
@@ -1033,10 +984,7 @@ Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean elec
1033
 
1034
  </details>
1035
 
1036
- </details>
1037
-
1038
- <details>
1039
- <summary><strong>Trigger word only</strong></summary>
1040
 
1041
  Caption: `system of a down, soad style music`
1042
 
@@ -1088,10 +1036,7 @@ Caption: `system of a down, soad style music`
1088
 
1089
  </details>
1090
 
1091
- </details>
1092
-
1093
- <details>
1094
- <summary><strong>Matching descriptors, no trigger word</strong></summary>
1095
 
1096
  Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
1097
 
@@ -1143,10 +1088,7 @@ Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with lay
1143
 
1144
  </details>
1145
 
1146
- </details>
1147
-
1148
- <details>
1149
- <summary><strong>Trigger word + matching descriptors</strong></summary>
1150
 
1151
  Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
1152
 
@@ -1200,8 +1142,6 @@ Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gent
1200
 
1201
  </details>
1202
 
1203
- </details>
1204
-
1205
  ## Evaluation lyrics
1206
 
1207
  The original comparison lyrics are shared by the first four tournaments. The latest singer-tag
 
47
  - **NextLat**: [MiniMaxMusicTraining/soad-mm3-nextlat-20260822](https://huggingface.co/MiniMaxMusicTraining/soad-mm3-nextlat-20260822)
48
  - **NextLat + XM**: [MiniMaxMusicTraining/soad-mm3-nextlat-xm-20260822](https://huggingface.co/MiniMaxMusicTraining/soad-mm3-nextlat-xm-20260822)
49
 
50
+ ### No trigger word / off-genre descriptors
 
51
 
52
  Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean electric piano, airy male vocals, smooth verse-chorus songwriting, restrained dynamics, polished intimate mix.`
53
 
 
99
 
100
  </details>
101
 
102
+ ### Trigger word only
 
 
 
103
 
104
  Caption: `system of a down, soad style music`
105
 
 
151
 
152
  </details>
153
 
154
+ ### Matching descriptors, no trigger word
 
 
 
155
 
156
  Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
157
 
 
203
 
204
  </details>
205
 
206
+ ### Trigger word + matching descriptors
 
 
 
207
 
208
  Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
209
 
 
257
 
258
  </details>
259
 
 
 
260
  <details>
261
  <summary><strong>Full song without regularisation</strong> (576 renders)</summary>
262
 
 
273
 
274
  > **Run note:** These four runs resumed their learned LoRA weights from checkpoint 500 with fresh optimizer, sampler, and RNG state after the original local trainer state was lost.
275
 
276
+ ### No trigger word / off-genre descriptors
 
277
 
278
  Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean electric piano, airy male vocals, smooth verse-chorus songwriting, restrained dynamics, polished intimate mix.`
279
 
 
325
 
326
  </details>
327
 
328
+ ### Trigger word only
 
 
 
329
 
330
  Caption: `system of a down, soad style music`
331
 
 
377
 
378
  </details>
379
 
380
+ ### Matching descriptors, no trigger word
 
 
 
381
 
382
  Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
383
 
 
429
 
430
  </details>
431
 
432
+ ### Trigger word + matching descriptors
 
 
 
433
 
434
  Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
435
 
 
483
 
484
  </details>
485
 
 
 
486
  <details>
487
  <summary><strong>Random windowing + regularisation</strong> (432 renders)</summary>
488
 
 
496
  - **NextLat**: [MiniMaxMusicTraining/soad-mm3-nextlat-20260823-10k-adamw8e-5-randwin128-reg128-nocapdrop](https://huggingface.co/MiniMaxMusicTraining/soad-mm3-nextlat-20260823-10k-adamw8e-5-randwin128-reg128-nocapdrop)
497
  - **NextLat + XM**: [MiniMaxMusicTraining/soad-mm3-nextlat-xm-20260823-10k-adamw8e-5-randwin128-reg128-nocapdrop](https://huggingface.co/MiniMaxMusicTraining/soad-mm3-nextlat-xm-20260823-10k-adamw8e-5-randwin128-reg128-nocapdrop)
498
 
499
+ ### No trigger word / off-genre descriptors
 
500
 
501
  Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean electric piano, airy male vocals, smooth verse-chorus songwriting, restrained dynamics, polished intimate mix.`
502
 
 
544
 
545
  </details>
546
 
547
+ ### Trigger word only
 
 
 
548
 
549
  Caption: `system of a down, soad style music`
550
 
 
592
 
593
  </details>
594
 
595
+ ### Matching descriptors, no trigger word
 
 
 
596
 
597
  Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
598
 
 
640
 
641
  </details>
642
 
643
+ ### Trigger word + matching descriptors
 
 
 
644
 
645
  Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
646
 
 
690
 
691
  </details>
692
 
 
 
693
  <details>
694
  <summary><strong>Caption v2: training-window comparison</strong> (400 renders)</summary>
695
 
 
706
 
707
  > **Run note:** Checkpoint coverage differs by leg; empty table cells were not rendered.
708
 
709
+ ### No trigger word / off-genre descriptors
 
710
 
711
  Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean electric piano, airy male vocals, smooth verse-chorus songwriting, restrained dynamics, polished intimate mix.`
712
 
 
758
 
759
  </details>
760
 
761
+ ### Trigger word only
 
 
 
762
 
763
  Caption: `system of a down, soad style music`
764
 
 
810
 
811
  </details>
812
 
813
+ ### Matching descriptors, no trigger word
 
 
 
814
 
815
  Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
816
 
 
862
 
863
  </details>
864
 
865
+ ### Trigger word + matching descriptors
 
 
 
866
 
867
  Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
868
 
 
916
 
917
  </details>
918
 
 
 
919
  <details>
920
  <summary><strong>Caption v2: singer-tag comparison</strong> (480 renders)</summary>
921
 
 
932
 
933
  > **Run note:** Checkpoint coverage differs by leg; empty table cells were not rendered.
934
 
935
+ ### No trigger word / off-genre descriptors
 
936
 
937
  Caption: `Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean electric piano, airy male vocals, smooth verse-chorus songwriting, restrained dynamics, polished intimate mix.`
938
 
 
984
 
985
  </details>
986
 
987
+ ### Trigger word only
 
 
 
988
 
989
  Caption: `system of a down, soad style music`
990
 
 
1036
 
1037
  </details>
1038
 
1039
+ ### Matching descriptors, no trigger word
 
 
 
1040
 
1041
  Caption: `Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
1042
 
 
1088
 
1089
  </details>
1090
 
1091
+ ### Trigger word + matching descriptors
 
 
 
1092
 
1093
  Caption: `system of a down, soad style music, lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.`
1094
 
 
1142
 
1143
  </details>
1144
 
 
 
1145
  ## Evaluation lyrics
1146
 
1147
  The original comparison lyrics are shared by the first four tournaments. The latest singer-tag