Deforum Gradient Conditioning Spreadsheet

	A	B	C	D	E	F	G
1		Variable used with			Gradient_wrt		Remember: These are guidelines, not rules! You can experiments with much higher values and sometimes negative values.
2					x (slow)	x0_pred (fast)	These are recommended settings if you want to use ONE of the variables named like "*_scale".
3			min	max	recommended		These recommendations are for a 512x512 image at scale=7,steps=50
4	Exposure/Contrast Conditional Settings
5	mean_scale		0	400	50	100
6	var_scale		-200	200	50	100
7	exposure_scale		0	400	50	100
8	exposure_target	exposure_scale	-1	1
9
10	Color Match Conditional Settings
11	colormatch_scale		0	1000	400	500	Colors are more accurate with decode_method=autoencoder
12	colormatch_image	colormatch_scale
13	colormatch_n_colors	colormatch_scale	1	6	4	4
14	ignore_sat_weight	colormatch_scale	1	5	1 or None	1 or None
15
16	CLIP\Aesthetics Conditional Settings
17	clip_name	clip_scale and aesthetics_scale	['ViT-L/14', 'ViT-L/14@336px', 'ViT-B/16', 'ViT-B/32']		ViT-L/14	ViT-L/14
18	clip_scale		0	200	50	not recommended	clip_scale may depend on which clip_name is being used, these recommendations are for ViT-L/14
19	aesthetics_scale		0	10	3	not recommended	aesthetics_scale may depend on which clip_name is being used, these recommendations are for ViT-L/14
20	cutn	clip_scale	1	6	1	1	Higher values use more VRAM if clip_scale > 0
21	cut_pow	clip_scale	0.000000001	10000000	0.001	0.001
22
23	Other Conditional Settings
24	init_mse_scale		0	1000	1000	500
25	init_mse_image	init_mse_scale
26
27	Conditional Gradient Settings
28	gradient_wrt	all *_scale variables	["x", "x0_pred"]		x	x0_pred	x is much slower, but uses the whole model to affect the content
29	gradient_add_to	all *_scale variables and cond_uncond_sync	["cond", "uncond", "both"]		both	both	values other than "both" only work when cond_uncond_sync==False
30	decode_method	all *_scale variables	["autoencoder","linear"]		linear	linear	linear is much faster and uses much less vram, but may not have the best quality in some cases. If the results look noisy, try autoencoder.
31	grad_threshold_type	all *_scale variables	["dynamic", "static", "mean", "schedule"]		mean	dynamic
32	clamp_grad_threshold	all *_scale variables	0	1	0.2	0.2
33	clamp_start	all *_scale variables and grad_threshold_type	0	1	0.2	0.2	only used when grad_threshold_type=="schedule"
34	clamp_stop	all *_scale variables and grad_threshold_type	0	1	0.01	0.01	only used when grad_threshold_type=="schedule"
35	grad_inject_timing	all *_scale variables			1	1
36
37	Speed vs VRAM Settings
38	cond_uncond_sync		True/False		TRUE	TRUE	False uses less VRAM, but is slower
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
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