10-19-2023, 08:56 PM
3. Find out best model by making them participate in tournament.
'The Tournament' is simple. I will generate 12 topics - 6 from the learned ones, and 6 from completely new ones, and repeat it with 6 seeds each.
Then I compare each and every lora type with each other - like, only LOCONs, only LOHAs8, only LOHAs16...
In comparsion, I will judge both how close on the style LORA got, and how much broken images it generate.
From each comparsion, up to two best matches receive 60pt lime colored circle in the corner.
Winning model is the one with most circles.
If the winning model is Baseline (built-in model), then all LoRAs is disqualified from tournament.
After all I will reach the situation where I have exactly one Locon, Loha8, Loha16, and Full. They will fight each other then, with different 12 seeds. The winner in the tournament is winner overall, and it's LoRA will be used as example for training better ones.
Tournament will be composed of:
XYZ grid
X parameter - Prompt S/R - (first half of items),by bahnbahn,(second half of items)
Y parameter - Prompt S/R - "zoroark, feral","mew, feral","princess luna \(mlp\), feral","floatzel, feral","dragon, feral","figdet, feral","typhlosion, feral","espeon, feral","owl, anthro","lapras, feral","orca, anthro","quaquaval, feral"
Z parameter - Seed - 142, 242, 342 | 442, 542, 642 (there will be 2 images, since one image would be 350mb)
Of course, Floatzel, Dragon and Fidget were already drawn by Bahn, when no Laprases, or Orcas, or Quaquavals were found.
3.1. Quarter-finals.
Tournament A: Standart LoCONs
Tournament A/1: Normal LR LoCONs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
1. STYLE_BAHNBAHN_V1SKS_LOCON16-000005: 60
2. STYLE_BAHNBAHN_V1SKS_LOCON16L-000007: 37
3. Baseline: 27
Result: while basic Bahn is awesome by itself, the winner is: SKS-style, High Dropout Rate, ADAM, 5th epoch.
Tournament A/2: Slow LR LoCONs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
1. STYLE_BAHNBAHN_V1SKS_LOCON16S-000012: 69
2. STYLE_BAHNBAHN_V1SKS_LOCON16LS-000012: 55
3. Baseline: 13
Result: Slowing down helped a lot, and while others are oversaturated or overtrained, the winner is: Slow Learning, SKS-style, High Dropout Rate, ADAM, 12th epoch.
Tournament A/3: Slowest LR LoCONs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
1. Baseline: 62
... can skip others.
Result: Baseline having a greater quality means that A/3 ends with no victors.
Tournament B: Standart LoHAs
Tournament B/1: Normal LR LoHAs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
1. STYLE_BAHNBAHN_V1SKS_LOHA8-000004: 46
1. Baseline: 46
2. STYLE_BAHNBAHN_V1SKS_LOHA8L-000004: 36
Result: Baseline having a similar quality to lora means that B/1 ends with no victors.
Tournament B/2: Slow LR LoHAs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
1. STYLE_BAHNBAHN_V1SKS_LOHA8S-000012: 61
2. STYLE_BAHNBAHN_V1SKS_LOHA8LS-000012: 22
3. Baseline: 11
Result: Close call (both S and LS were good!), but Slow Learning, SKS-style, High Dropout Rate, ADAM, 12th epoch - wins.
Tournament B/3: Slowest LR LoHAs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
1. STYLE_BAHNBAHN_V1SKS_LOHA8SS-000012: 64
2. Baseline: 52
Result: With barely escaping the foul line, Slowest Learning, SKS-style, High Dropout Rate, ADAM, 12th epoch wins.
Tournament C: Expanded LoHAs
Tournament C/1: Normal LR LoHAs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
Result: Everyone disqualified. Everything is so obviously different from baseline, that you don't need any other proofs.
Tournament C/2: Slow LR LoHAs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
Result: ...again, everyone disqualified for the same reason!
Tournament C/3: Slowest LR LoHAs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
Result: This tournament is by default won by Slowest Learning, SKS-style, High Dropout Rate, ADAM, 12th epoch wins.
Tournament D: Native Training - Slow Learners
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
Result: I honestly started to check those, but in after first seed I realized that everything would simply fail. Everyone disqualified.
Tournament E: Native Training - Slowest Learners
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
Result: Same as D. One glance was enough to see imperfections, and therefore, everyone was disqualified.
3.2. Semi-finals.
I have changed SEED parameter to compare different situations. LORAs must be versatile.
I also tweaked Prompts to match better themes. Fidgets/Nimbats are for some reason not really recognized by network, so I used 'anthro wywern' instead. I also changed Espeon to Vaporeon, Owl to Braixen, and absent Quaquaval to Blaziken - as far as I can remember, none were drawn by Bahn, and none were in dataset.
Finally, this time, there can be only 1 item selected per choice.
Y parameter - Prompt S/R - "zoroark, feral","mew, feral","princess luna \(mlp\), feral","floatzel, feral","dragon, feral","wyvern, anthro","typhlosion, feral","vaporeon, feral","braixen, feral","lapras, feral","orca, anthro","blaziken, feral"
Z parameter - Seed - 1042, 2042, 3042, 4042, 5042, 6042
Tournament A/T: Best of LoCONs
The Tournament A/T
Winner of A/1 - STYLE_BAHNBAHN_V1SKS_LOCON16-000005: 6
Winner of A/2 - STYLE_BAHNBAHN_V1SKS_LOCON16S-000012: 59
Baseline: 7
Result: Best of LOCONs - Slow Learning, SKS-style, High Dropout Rate, ADAM, 12th epoch.
Tournament B/T: Best of LoHAs (standart)
The Tournament B/T
Winner of B/2 - STYLE_BAHNBAHN_V1SKS_LOHA8S-000012: 66
Winner of B/3 - STYLE_BAHNBAHN_V1SKS_LOHA8SS-000012: 4
Baseline: 2
Result: Best of LOHAs - Slow Learning, SKS-style, High Dropout Rate, ADAM, 12th epoch.
Tournament C/T: Best of LoHAs (expanded)
Winner by default (no competitors!) - STYLE_BAHNBAHN_V1SKS_LOHA16SS-000012
3.3. Finals.
I have changed SEED parameter again and tweaked Y parameter once more.
Changed Wyvern to Dragon; added Deer (Deerling is a deer, after all) and Lucario, also Goodra (feral, not in dataset) and Night Fury (also not in dataset or anywhere for that matter).
Y parameter - Prompt S/R - "zoroark, feral","mew, feral","princess luna \(mlp\), feral","floatzel, feral","dragon, feral","dragon, anthro","deer, feral","lucario, feral","typhlosion, feral","vaporeon, feral","braixen, feral","lapras, feral","orca, anthro","blaziken, feral","goodra, feral","night fury, toothless, anthro"
Z parameter - Seed - 10042, 20042, 30042, 40042, 50042, 60042, 70042, 80042, 90042, 100042, 110042, 120042
Tournament Ω: One True LoRA
The Tournament Ω part 1
The Tournament Ω part 2
1st place: STYLE_BAHNBAHN_V1SKS_LOHA8S-000012: 66
2nd place: STYLE_BAHNBAHN_V1SKS_LOCON16S-000005: 21
3nd place: STYLE_BAHNBAHN_V1SKS_LOHA16SS-000012: 9
And so, the Ultimate Leader, and therefore, First Version, and therefore First Available Soul is: LoHA of 8 dimensions, Slow Learning, SKS-style, High Dropout Rate, ADAM, 12th epoch.
'The Tournament' is simple. I will generate 12 topics - 6 from the learned ones, and 6 from completely new ones, and repeat it with 6 seeds each.
Then I compare each and every lora type with each other - like, only LOCONs, only LOHAs8, only LOHAs16...
In comparsion, I will judge both how close on the style LORA got, and how much broken images it generate.
From each comparsion, up to two best matches receive 60pt lime colored circle in the corner.
Winning model is the one with most circles.
If the winning model is Baseline (built-in model), then all LoRAs is disqualified from tournament.
After all I will reach the situation where I have exactly one Locon, Loha8, Loha16, and Full. They will fight each other then, with different 12 seeds. The winner in the tournament is winner overall, and it's LoRA will be used as example for training better ones.
Tournament will be composed of:
XYZ grid
X parameter - Prompt S/R - (first half of items),by bahnbahn,(second half of items)
Y parameter - Prompt S/R - "zoroark, feral","mew, feral","princess luna \(mlp\), feral","floatzel, feral","dragon, feral","figdet, feral","typhlosion, feral","espeon, feral","owl, anthro","lapras, feral","orca, anthro","quaquaval, feral"
Z parameter - Seed - 142, 242, 342 | 442, 542, 642 (there will be 2 images, since one image would be 350mb)
Of course, Floatzel, Dragon and Fidget were already drawn by Bahn, when no Laprases, or Orcas, or Quaquavals were found.
3.1. Quarter-finals.
Tournament A: Standart LoCONs
Tournament A/1: Normal LR LoCONs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
1. STYLE_BAHNBAHN_V1SKS_LOCON16-000005: 60
2. STYLE_BAHNBAHN_V1SKS_LOCON16L-000007: 37
3. Baseline: 27
Result: while basic Bahn is awesome by itself, the winner is: SKS-style, High Dropout Rate, ADAM, 5th epoch.
Tournament A/2: Slow LR LoCONs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
1. STYLE_BAHNBAHN_V1SKS_LOCON16S-000012: 69
2. STYLE_BAHNBAHN_V1SKS_LOCON16LS-000012: 55
3. Baseline: 13
Result: Slowing down helped a lot, and while others are oversaturated or overtrained, the winner is: Slow Learning, SKS-style, High Dropout Rate, ADAM, 12th epoch.
Tournament A/3: Slowest LR LoCONs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
1. Baseline: 62
... can skip others.
Result: Baseline having a greater quality means that A/3 ends with no victors.
Tournament B: Standart LoHAs
Tournament B/1: Normal LR LoHAs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
1. STYLE_BAHNBAHN_V1SKS_LOHA8-000004: 46
1. Baseline: 46
2. STYLE_BAHNBAHN_V1SKS_LOHA8L-000004: 36
Result: Baseline having a similar quality to lora means that B/1 ends with no victors.
Tournament B/2: Slow LR LoHAs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
1. STYLE_BAHNBAHN_V1SKS_LOHA8S-000012: 61
2. STYLE_BAHNBAHN_V1SKS_LOHA8LS-000012: 22
3. Baseline: 11
Result: Close call (both S and LS were good!), but Slow Learning, SKS-style, High Dropout Rate, ADAM, 12th epoch - wins.
Tournament B/3: Slowest LR LoHAs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
1. STYLE_BAHNBAHN_V1SKS_LOHA8SS-000012: 64
2. Baseline: 52
Result: With barely escaping the foul line, Slowest Learning, SKS-style, High Dropout Rate, ADAM, 12th epoch wins.
Tournament C: Expanded LoHAs
Tournament C/1: Normal LR LoHAs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
Result: Everyone disqualified. Everything is so obviously different from baseline, that you don't need any other proofs.
Tournament C/2: Slow LR LoHAs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
Result: ...again, everyone disqualified for the same reason!
Tournament C/3: Slowest LR LoHAs
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
Result: This tournament is by default won by Slowest Learning, SKS-style, High Dropout Rate, ADAM, 12th epoch wins.
Tournament D: Native Training - Slow Learners
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
Result: I honestly started to check those, but in after first seed I realized that everything would simply fail. Everyone disqualified.
Tournament E: Native Training - Slowest Learners
First part (seeds 142, 242, 342) + Second part (seeds 442, 542, 642)
Result: Same as D. One glance was enough to see imperfections, and therefore, everyone was disqualified.
3.2. Semi-finals.
I have changed SEED parameter to compare different situations. LORAs must be versatile.
I also tweaked Prompts to match better themes. Fidgets/Nimbats are for some reason not really recognized by network, so I used 'anthro wywern' instead. I also changed Espeon to Vaporeon, Owl to Braixen, and absent Quaquaval to Blaziken - as far as I can remember, none were drawn by Bahn, and none were in dataset.
Finally, this time, there can be only 1 item selected per choice.
Y parameter - Prompt S/R - "zoroark, feral","mew, feral","princess luna \(mlp\), feral","floatzel, feral","dragon, feral","wyvern, anthro","typhlosion, feral","vaporeon, feral","braixen, feral","lapras, feral","orca, anthro","blaziken, feral"
Z parameter - Seed - 1042, 2042, 3042, 4042, 5042, 6042
Tournament A/T: Best of LoCONs
The Tournament A/T
Winner of A/1 - STYLE_BAHNBAHN_V1SKS_LOCON16-000005: 6
Winner of A/2 - STYLE_BAHNBAHN_V1SKS_LOCON16S-000012: 59
Baseline: 7
Result: Best of LOCONs - Slow Learning, SKS-style, High Dropout Rate, ADAM, 12th epoch.
Tournament B/T: Best of LoHAs (standart)
The Tournament B/T
Winner of B/2 - STYLE_BAHNBAHN_V1SKS_LOHA8S-000012: 66
Winner of B/3 - STYLE_BAHNBAHN_V1SKS_LOHA8SS-000012: 4
Baseline: 2
Result: Best of LOHAs - Slow Learning, SKS-style, High Dropout Rate, ADAM, 12th epoch.
Tournament C/T: Best of LoHAs (expanded)
Winner by default (no competitors!) - STYLE_BAHNBAHN_V1SKS_LOHA16SS-000012
3.3. Finals.
I have changed SEED parameter again and tweaked Y parameter once more.
Changed Wyvern to Dragon; added Deer (Deerling is a deer, after all) and Lucario, also Goodra (feral, not in dataset) and Night Fury (also not in dataset or anywhere for that matter).
Y parameter - Prompt S/R - "zoroark, feral","mew, feral","princess luna \(mlp\), feral","floatzel, feral","dragon, feral","dragon, anthro","deer, feral","lucario, feral","typhlosion, feral","vaporeon, feral","braixen, feral","lapras, feral","orca, anthro","blaziken, feral","goodra, feral","night fury, toothless, anthro"
Z parameter - Seed - 10042, 20042, 30042, 40042, 50042, 60042, 70042, 80042, 90042, 100042, 110042, 120042
Tournament Ω: One True LoRA
The Tournament Ω part 1
The Tournament Ω part 2
1st place: STYLE_BAHNBAHN_V1SKS_LOHA8S-000012: 66
2nd place: STYLE_BAHNBAHN_V1SKS_LOCON16S-000005: 21
3nd place: STYLE_BAHNBAHN_V1SKS_LOHA16SS-000012: 9
And so, the Ultimate Leader, and therefore, First Version, and therefore First Available Soul is: LoHA of 8 dimensions, Slow Learning, SKS-style, High Dropout Rate, ADAM, 12th epoch.