abakanai commited on
Commit
045e37e
·
verified ·
1 Parent(s): c578023

Add 1376x768 simple comparisons

Browse files
.gitattributes CHANGED
@@ -36,3 +36,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
36
  examples/astrella_conservative_16_38g_s20_audio_00001_.mp4 filter=lfs diff=lfs merge=lfs -text
37
  examples/astrella_fast_14_03g_s20_audio_00001_.mp4 filter=lfs diff=lfs merge=lfs -text
38
  examples/minimax_h3_ab_s20_contact.png filter=lfs diff=lfs merge=lfs -text
 
 
 
36
  examples/astrella_conservative_16_38g_s20_audio_00001_.mp4 filter=lfs diff=lfs merge=lfs -text
37
  examples/astrella_fast_14_03g_s20_audio_00001_.mp4 filter=lfs diff=lfs merge=lfs -text
38
  examples/minimax_h3_ab_s20_contact.png filter=lfs diff=lfs merge=lfs -text
39
+ examples/highres/astrella_fast_14_03g_1376x768_s20_simple_audio_00001_.mp4 filter=lfs diff=lfs merge=lfs -text
40
+ examples/highres/astrella_conservative_16_38g_1376x768_s20_simple_audio_00001_.mp4 filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -29,6 +29,12 @@ widget:
29
  - text: Matched-seed A/B contact sheet
30
  output:
31
  url: examples/minimax_h3_ab_s20_contact.png
 
 
 
 
 
 
32
  ---
33
 
34
  # MiniMax H3 Hybrid NVFP4 for NVIDIA Blackwell
@@ -48,7 +54,7 @@ The attention-output and FFN-output projections remain INT8 ConvRot in both vari
48
 
49
  ## Generated examples
50
 
51
- All media below are AI-generated. Both videos use the same reference image, prompt, seed, resolution, duration, sampler, scheduler, and step count.
52
 
53
  <Gallery />
54
 
@@ -85,6 +91,20 @@ All media below are AI-generated. Both videos use the same reference image, prom
85
 
86
  In this single-seed qualitative check, both variants preserved the character's face and glasses across the clip. The conservative variant held finger shapes slightly more consistently, while the fast variant produced a very similar result with somewhat larger changes in background composition and framing. This is one visual sample, not a statistically meaningful perceptual evaluation.
87
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
88
  ## Research summary
89
 
90
  ### Test environment
@@ -174,4 +194,3 @@ The separately distributed Qwen3-VL-32B encoder is licensed under Apache-2.0 by
174
  - Original model: [`MiniMaxAI/MiniMax-H3`](https://huggingface.co/MiniMaxAI/MiniMax-H3)
175
  - ComfyUI single-file release and integration: [`Comfy-Org/MiniMax-H3`](https://huggingface.co/Comfy-Org/MiniMax-H3)
176
  - Hybrid quantization, kernel benchmarking, numerical validation, and A/B generation: **abakanai**
177
-
 
29
  - text: Matched-seed A/B contact sheet
30
  output:
31
  url: examples/minimax_h3_ab_s20_contact.png
32
+ - text: Fast hybrid high-resolution — simple, 1376 × 768, rendered in 568.67 s
33
+ output:
34
+ url: examples/highres/astrella_fast_14_03g_1376x768_s20_simple_audio_00001_.mp4
35
+ - text: Conservative hybrid high-resolution — simple, 1376 × 768, rendered in 598.66 s
36
+ output:
37
+ url: examples/highres/astrella_conservative_16_38g_1376x768_s20_simple_audio_00001_.mp4
38
  ---
39
 
40
  # MiniMax H3 Hybrid NVFP4 for NVIDIA Blackwell
 
54
 
55
  ## Generated examples
56
 
57
+ All media below are AI-generated. The paired videos in each comparison use the same reference image, prompt, seed, resolution, duration, sampler, scheduler, and step count.
58
 
59
  <Gallery />
60
 
 
91
 
92
  In this single-seed qualitative check, both variants preserved the character's face and glasses across the clip. The conservative variant held finger shapes slightly more consistently, while the fast variant produced a very similar result with somewhat larger changes in background composition and framing. This is one visual sample, not a statistically meaningful perceptual evaluation.
93
 
94
+ ### 1376 × 768 `simple` follow-up
95
+
96
+ The two hybrid checkpoints were also tested at 1376 × 768 using the same reference, prompt, seed, 124-frame duration, 20 steps, and `res_multistep` sampler. The scheduler was changed to `simple` for both matched runs.
97
+
98
+ - [Fast 14.03 GiB high-resolution video with audio](./examples/highres/astrella_fast_14_03g_1376x768_s20_simple_audio_00001_.mp4)
99
+ - [Conservative 16.38 GiB high-resolution video with audio](./examples/highres/astrella_conservative_16_38g_1376x768_s20_simple_audio_00001_.mp4)
100
+
101
+ | Checkpoint | Resolution | Video duration | Total render time |
102
+ |---|---:|---:|---:|
103
+ | Fast 14.03 GiB | 1376 × 768 | 5.17 s | 568.67 s (9:28.67) |
104
+ | Conservative 16.38 GiB | 1376 × 768 | 5.17 s | 598.66 s (9:58.66) |
105
+
106
+ Both runs completed on the RTX 5070 Ti 16 GB through ComfyUI DynamicVRAM and produced 124-frame, 24 fps MP4 files with stereo AAC audio at 32 kHz. The fast checkpoint finished about 5% sooner in these sequential single runs. This is a practical observation rather than a controlled end-to-end benchmark because model and cache state were not fully reset between jobs.
107
+
108
  ## Research summary
109
 
110
  ### Test environment
 
194
  - Original model: [`MiniMaxAI/MiniMax-H3`](https://huggingface.co/MiniMaxAI/MiniMax-H3)
195
  - ComfyUI single-file release and integration: [`Comfy-Org/MiniMax-H3`](https://huggingface.co/Comfy-Org/MiniMax-H3)
196
  - Hybrid quantization, kernel benchmarking, numerical validation, and A/B generation: **abakanai**
 
examples/highres/astrella_conservative_16_38g_1376x768_s20_simple_audio_00001_.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:54e3ebc6ce904c4123e79da574069e8ffbfee88bc246a556425ae7913a3de46f
3
+ size 2752589
examples/highres/astrella_fast_14_03g_1376x768_s20_simple_audio_00001_.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:75a86411edc154314af85e62868c4df94ecac4b428bfeae5d1f04ba070b6ddc3
3
+ size 2696509