model test settings table.xlsx

	A	B	C	D	E	F	G	H	N	O	R	U
1	4dimbase1	AdamW8bit	4	10	bf16	0.0003	2	4	equirectangular oil on canvas painting by Van Gogh, The bank of the Hudson River with New York City skyscrapers distant in the background, detailed brushstrokes
2	8dimbase1	AdamW8bit	4	10	fp16	0.0003	2	4
3	v8	Adafactor	3	50	fp16	0.0003	1	4		equirectangular photo, centered, architecture, design, aesthetically pleasing, a psychedelic, colorful groovy, 1960's living room
4	v10	Adafactor	3	3	fp16	0.0003	1	8			Steps: 35, Sampler: Restart, CFG scale: 7.5, Seed: 8489645495946, Size: 1448x724
5	v5150	Adafactor	3	3	fp16	0.0003	1	8
6	v6	Prodigy	3	50	fp16	1	1	8
7	dim8alpha1	AdamW	3	120	fp16	0.00001	2	16
8	v126	AdamW8bit	6	10	fp16	0.00008	8	16
9	v135	AdamW8bit	6	10	fp16	0.0005	8	16
10	v18	Prodigy	1	15	fp16	1	4	16
11	v9largedataset	Prodigy	4	1	fp16	1	8	16
12	32dimbase1cont	Prodigy	4	10	fp16	1	16	32
13	d32a1version3juggernaut40	Prodigy	3	50	fp16	1	1	32
14	modeltestv3plus50	AdamW	3	100	fp16	0.00003	32	32
15	v11	AdamW8bit	6	60	fp16	0.0002	8	32
16	v115	Adafactor	1	3	bf16	0.0003	1	32
17	v145	AdamW8bit	6	70	fp16	0.0004	8	32
18	v146	Prodigy	6	15	fp16	1	2	32
19	v15	Prodigy	6	15	fp16	1	1	32
20	v16	Prodigy	6	15	fp16	1	4	32
21	v17	Prodigy	6	90	fp16	1	2	32
22	v560	Prodigy	3	15	fp16	1	1	32
23	v590	Prodigy	3	30	fp16	1	1	32
24	v7	Prodigy	3	100	fp16	1	16	32
25	File Name	Optimizer	Batch Size	Epochs	Save Precision	Learning Rate	Network Alpha	Network Dim				<lora:v15:1>, <lora:4dimbase1:1>, <lora:8dimbase1:1>, <lora:v8:1>, <lora:v10:1>, <lora:v5150:1>, <lora:v6:1>, <lora:dim8alpha1:1>, <lora:v126:1>, <lora:v135:1>, <lora:v18:1>, <lora:v9largedataset:1>, <lora:32dimbase1cont:1>, <lora:d32a1version3juggernaut40:1>, <lora:modeltestv3plus50:1>, <lora:v11:1>, <lora:v115:1>, <lora:v145:1>, <lora:v146:1>, <lora:v16:1>, <lora:v17:1>, <lora:v560:1>, <lora:v590:1>, <lora:v7:1>
26
27
28
29	Note: I want to explore adafactor more, a learning rate of .0003-.0004 seems good.
30	Network dim 32 alpha 8 followed the prompt the best and preserved the equirectangular
31	projection for the van gogh landscape scene
32
33	Dim 32 is consistently the best
34
35	Note; epochs tab is incorrect for some models, v145, 146, 115, 560, 590, and 5150 are trainings
36	resumed from an existing checkpoint. v145 = v14 + 5 more epochs
37
38	11/2	V21. and V22 use network dim of 128 with alpha of 64. This improved the
39	quality of the model greatly to render images in different styles than the training set
40	My next step will be to reduce the size of the training set, leaving only 100 or so of the
41	most asthetically pleasing images, hopefully this will improve
42
43
44
45
46
47
48
49
50
51
52
53
54
55	Above, an image created with v23j, which also uses a different base model from civitai, juggernaut v6
56	equirectangular oil on canvas painting, by Van Gogh, a decaying town square in Chernobyl, apocalyptic, brushstrokes <lora:v23j:1>
57	Steps: 60, Sampler: DPM++ 3M SDE Karras, CFG scale: 7, Seed: 940668922, Size: 1448x724, Model hash: 1fe6c7ec54, Model: juggernautXL_version6Rundiffusion, Denoising strength: 0.45, Hires sampler: Restart, Hires upscale: 2, Hires steps: 10, Hires upscaler: SwinIR_4x, Lora hashes: "v23j: 42febccca6d2", VAE Encoder: TAESD, Version: 1.6.0
58
59	Time taken: 2 min. 23.2 sec.
60
61
62
63
64
65
66
67	equirectangular oil on canvas painting, by Van Gogh, a decaying town square in Chernobyl, apocalyptic, brushstrokes <lora:v23j:1>
68	Steps: 60, Sampler: DPM++ 3M SDE Karras, CFG scale: 7, Seed: 940668922, Size: 1448x724, Model hash: 1fe6c7ec54, Model: juggernautXL_version6Rundiffusion, Denoising strength: 0.45, Hires sampler: Restart, Hires upscale: 2, Hires steps: 10, Hires upscaler: SwinIR_4x, Lora hashes: "v23j: 42febccca6d2", VAE Encoder: TAESD, Version: 1.6.0
69
70	Time taken: 2 min. 23.2 sec.
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100