Hugging Face's logo

AbstractFramework
/
wan2.2-ti2v-5b-diffusers-bf16

Text-to-Video
Safetensors
MLX
Wan2.2
mlx-gen
mflux
apple-silicon
bf16
wan
video-generation
image-to-video
Model card Files
xet
 Community
lpalbou commited on
Commit
6738b2c
·
verified ·
1 Parent(s): 0853dcc

Add TI2V-5B validation assets

Browse files
Files changed (1) hide show
  1. validation/ti2v5b-clean/metrics.json +128 -0
validation/ti2v5b-clean/metrics.json ADDED
@@ -0,0 +1,128 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "validation_date": "2026-06-04",
3
+ "machine": "Apple M5 Max, 128 GiB unified memory",
4
+ "profile": {
5
+ "prompt": "A short cinematic video of a glowing orange glass sphere floating above calm teal water, soft reflections, gentle camera movement",
6
+ "negative_prompt": "",
7
+ "width": 1280,
8
+ "height": 704,
9
+ "frames": 17,
10
+ "steps": 20,
11
+ "guidance": 5,
12
+ "fps": 24,
13
+ "seed": 321
14
+ },
15
+ "layouts": {
16
+ "base_source": {
17
+ "repo": "Wan-AI/Wan2.2-TI2V-5B-Diffusers",
18
+ "storage_gib": 31.85,
19
+ "mp4": "base-source.mp4",
20
+ "total_seconds": 216.15181183815002,
21
+ "generation_seconds": 215.164,
22
+ "physical_peak_gib": 102.67981120198965,
23
+ "rss_peak_gib": 13.703689575195312,
24
+ "resident_peak_gib": 13.702316284179688,
25
+ "mlx_peak_gib": 58.507674764841795,
26
+ "mlx_active_after_generate_gib": 10.348156608641148,
27
+ "mlx_cache_after_generate_gib": 0.48403943702578545,
28
+ "logical_model_gib": 10.62639182060957,
29
+ "logical_transformer_gib": 9.313806891441345,
30
+ "logical_vae_gib": 1.3125849291682243,
31
+ "model_bits": null,
32
+ "health": {
33
+ "source": "validation_outputs/ti2v_memory_profile_2026_06_04/base_source.mp4",
34
+ "frame_count": 17,
35
+ "width": 1280,
36
+ "height": 704,
37
+ "fps": 24.0,
38
+ "luma_min": 7.228399753570557,
39
+ "luma_max": 246.07899475097656,
40
+ "luma_mean": 174.08145500631895,
41
+ "mean_temporal_delta": 0.7045322582125664
42
+ }
43
+ },
44
+ "prepared_bf16": {
45
+ "repo": "AbstractFramework/wan2.2-ti2v-5b-diffusers-bf16",
46
+ "storage_gib": 21.22,
47
+ "mp4": "prepared-bf16.mp4",
48
+ "total_seconds": 261.56354999542236,
49
+ "generation_seconds": 260.648,
50
+ "physical_peak_gib": 102.583978459239,
51
+ "rss_peak_gib": 14.515884399414062,
52
+ "resident_peak_gib": 14.515884399414062,
53
+ "mlx_peak_gib": 58.472518514841795,
54
+ "mlx_active_after_generate_gib": 10.348156608641148,
55
+ "mlx_cache_after_generate_gib": 0.48403943702578545,
56
+ "logical_model_gib": 10.62639182060957,
57
+ "logical_transformer_gib": 9.313806891441345,
58
+ "logical_vae_gib": 1.3125849291682243,
59
+ "model_bits": null,
60
+ "health": {
61
+ "source": "validation_outputs/ti2v_memory_profile_2026_06_04/prepared_bf16.mp4",
62
+ "frame_count": 17,
63
+ "width": 1280,
64
+ "height": 704,
65
+ "fps": 24.0,
66
+ "luma_min": 7.228399753570557,
67
+ "luma_max": 246.07899475097656,
68
+ "luma_mean": 174.08145500631895,
69
+ "mean_temporal_delta": 0.7045322582125664
70
+ }
71
+ },
72
+ "mixed_q8_bf16": {
73
+ "repo": "AbstractFramework/wan2.2-ti2v-5b-diffusers-8bit",
74
+ "storage_gib": 16.86,
75
+ "mp4": "mixed-q8-bf16.mp4",
76
+ "total_seconds": 243.38612413406372,
77
+ "generation_seconds": 242.664,
78
+ "physical_peak_gib": 103.74846588075161,
79
+ "rss_peak_gib": 13.800796508789062,
80
+ "resident_peak_gib": 13.79901123046875,
81
+ "mlx_peak_gib": 54.187850546091795,
82
+ "mlx_active_after_generate_gib": 6.063488639891148,
83
+ "mlx_cache_after_generate_gib": 0.48403943702578545,
84
+ "logical_model_gib": 6.3417238518595695,
85
+ "logical_transformer_gib": 5.029138922691345,
86
+ "logical_vae_gib": 1.3125849291682243,
87
+ "model_bits": 8,
88
+ "health": {
89
+ "source": "validation_outputs/ti2v_memory_profile_2026_06_04/mixed_q8_bf16.mp4",
90
+ "frame_count": 17,
91
+ "width": 1280,
92
+ "height": 704,
93
+ "fps": 24.0,
94
+ "luma_min": 7.866199493408203,
95
+ "luma_max": 246.94140625,
96
+ "luma_mean": 174.77739491182214,
97
+ "mean_temporal_delta": 0.6711972616612911
98
+ }
99
+ }
100
+ },
101
+ "comparisons": {
102
+ "base_vs_prepared_bf16": {
103
+ "frame_count": 17,
104
+ "byte_identical_decoded_frames": true,
105
+ "mean_frame_mae": 0.0,
106
+ "max_frame_mae": 0.0,
107
+ "max_abs_channel_delta": 0
108
+ },
109
+ "base_vs_mixed_q8_bf16": {
110
+ "frame_count": 17,
111
+ "byte_identical_decoded_frames": false,
112
+ "mean_frame_mae": 1.6606264709363858,
113
+ "max_frame_mae": 1.7536328864820077,
114
+ "max_abs_channel_delta": 133
115
+ },
116
+ "prepared_bf16_vs_mixed_q8_bf16": {
117
+ "frame_count": 17,
118
+ "byte_identical_decoded_frames": false,
119
+ "mean_frame_mae": 1.6606264709363858,
120
+ "max_frame_mae": 1.7536328864820077,
121
+ "max_abs_channel_delta": 133
122
+ }
123
+ },
124
+ "interpretation": {
125
+ "prepared_bf16": "Storage/download reduction only for this profile; MLX-Gen already casts upstream Wan transformer and VAE weights to BF16 at runtime.",
126
+ "mixed_q8_bf16": "Reduces storage, logical model bytes, active MLX bytes, and MLX allocator peak. This profile did not reduce full-process Darwin physical peak because non-model transient memory dominated the run."
127
+ }
128
+ }