[ComfyUI]Breakthrough Release: MiniMax H3, the New Leading Open-Source Video Model — Transcript

Full transcript

1. 0:05[whistles]
2. 0:11[whistles]
3. 0:13Wow.
4. 0:18[music]
5. 0:21Fore!
6. 0:23[music]
7. 0:32[snorts]
8. 0:38[snorts]
9. 0:40Foreign! Foreign!
10. 1:00Hi everyone, I'm Jing Chen. It is like
11. 1:03New Year's in the open source community.
12. 1:05Although the release was delayed by a
13. 1:06few hours, that doesn't take away from
14. 1:08the fact that it is open source. Haha.
15. 1:11Today I am bringing you Miniaax H3, a
16. 1:14multimodal video model. It was released
17. 1:16just 2 days ago, so I haven't had time
18. 1:18to test it in depth yet. From this
19. 1:20initial round of testing, I feel it can
20. 1:23already comfortably beat some closed
21. 1:25source models. Breaking the dominance of
22. 1:27closed source models. The consistency is
23. 1:29great and the motion is great, too. This
24. 1:32is the kind of result we never would
25. 1:33have imagined before. So, I'm really
26. 1:36happy about it. Let's start with the
27. 1:38most important thing and what everyone
28. 1:40cares about most, how to speed it up.
29. 1:43Right now, I most recommend the pruned
30. 1:46intate model combined with the MVFP4
31. 1:48clip. This is the fastest combination
32. 1:50and it works well with most people
33. 1:52setups. Even lower machines have a
34. 1:55chance of running it. You only need 8 GB
35. 1:58of RAM. As long as you have enough
36. 2:00system memory, then set aside some RAM
37. 2:02in the startup parameters and give it a
38. 2:04try. Then add such attention after the
39. 2:06model. If you haven't installed Sage
40. 2:08yet, I really recommend installing it.
41. 2:10There is also this prediction node. If
42. 2:12the video doesn't have a lot of motion,
43. 2:14you can turn it on to reduce the amount
44. 2:16of actual sampling in the middle
45. 2:18section. In my test with no acceleration
46. 2:21enabled around 1 megapixel, which is 768
47. 2:25by 1376,
48. 2:27generating a 10-second video took about
49. 2:3030 to 34 minutes on running hub plus
50. 2:32with a 48 GB 490 setup. After adding
51. 2:36such attention, the time was cut in
52. 2:39half. It took around 14 to 16 minutes to
53. 2:42generate. Then with the prediction node
54. 2:44added 768p10.
55. 2:47Second video took about 10 minutes
56. 2:49online. That is already extremely fast.
57. 2:52I also tested the easy cache node
58. 2:54mentioned by other creators.
59. 2:57Although both methods speed things up by
60. 2:59reducing full model computations. After
61. 3:01comparing them, I personally feel the
62. 3:04prediction node gives better quality.
63. 3:06You can try them separately as well.
64. 3:08Let's look at a same seed comparison.
65. 3:11The middle one uses saggot attention
66. 3:13only. The one on the left uses a
67. 3:15detention plus the prediction node.
68. 3:18There actually isn't much difference
69. 3:19between the two. For this level of
70. 3:21motion, the prediction acceleration
71. 3:24performs quite well. The one on the
72. 3:26right uses easy. You can feel that the
73. 3:28distant people look noticeably
74. 3:29blurriier. If you're making a high
75. 3:31motion video, I recommend turning on
76. 3:33prediction acceleration to quickly
77. 3:35search for a good seed. Once you find a
78. 3:37result with suitable motion,
79. 3:38composition, and seed, turn off the
80. 3:40prediction node and run it again fully
81. 3:43with the same seed to preserve as much
82. 3:45final detail as possible. This time I've
83. 3:48prepared three workflows, text to video,
84. 3:51first and last frame video, and full
85. 3:53reference video. The text to video and
86. 3:56first and last frame workflows are
87. 3:57fairly simple. You can download them
88. 3:59locally and explore them yourselves.
89. 4:01Today, I'll mainly cover the full
90. 4:03reference workflow, which is the most
91. 4:04useful one.
92. 4:06Full reference workflow uses the
93. 4:08reference to video model.
94. 4:11It can take reference images, reference
95. 4:13videos, and reference audio at the same
96. 4:15time, which can control the characters,
97. 4:18scene, style, motion, camera movement,
98. 4:21and voice. The size and duration
99. 4:23settings are standard.
100. 4:26Just compare them with the table on the
101. 4:27side. In general, 0.4 megapixels should
102. 4:31work on most machines. That is roughly
103. 4:33480p. The 0.8 and 1.0 megapixel settings
104. 4:38look sharper, but they also require more
105. 4:40resources and take longer. You can keep
106. 4:42adding image, video, and audio inputs on
107. 4:45the reference node. The node expands
108. 4:47automatically. According to the official
109. 4:50documentation, you can add up to nine
110. 4:52images, three audio clips, and three
111. 4:54videos. That is more than enough for
112. 4:56normal use. Every additional reference
113. 4:59makes the generation take longer,
114. 5:01especially reference videos. They can
115. 5:03increase the runtime significantly. Just
116. 5:06add references according to your actual
117. 5:08needs.
118. 5:09The workflow itself isn't complicated.
119. 5:12The really difficult part is the prompt.
120. 5:14The full reference workflow handles
121. 5:15images, videos, audio, character
122. 5:18relationships, motion, camera work, and
123. 5:21sound all at once. If you only write a
124. 5:23few simple sentences, it is easy to make
125. 5:25the roles of the references unclear. It
126. 5:28might even say things you can't
127. 5:29understand. So this time I've prepared a
128. 5:31system prompt template. It automatically
129. 5:34identifies whether you're providing
130. 5:36text, images, videos, or audio. Then
131. 5:39completes it into a prompt suitable for
132. 5:41H3.
133. 5:43When using it, just put this system
134. 5:46prompt into a web- based vision model.
135. 5:48Then send it your video requirements and
136. 5:50all your reference materials. Briefly
137. 5:52describe the story and result you want,
138. 5:54and it will organize a prompt you can
139. 5:56copy directly into the H3 workflow.
140. 6:03This template was organized and refined
141. 6:05based on several official prompt guides.
142. 6:08If you're interested, you can also open
143. 6:10the official documentation to learn
144. 6:12more. I've put the links in the
145. 6:14description. The download package will
146. 6:16include the pruned model and the
147. 6:18smallest clip. For local use, I
148. 6:21recommend starting with the smaller
149. 6:22clip. If you have plenty of RAM online,
150. 6:25you can also try the intake clip. ASA
151. 6:28I'll share the three workflows with
152. 6:29acceleration nodes already connected
153. 6:32along with the system prompt template
154. 6:33and the prediction acceleration node.
155. 6:36I'll share all of them with you. That is
156. 6:38all for now. Go ahead and start playing
157. 6:40with it. Open sourcing a video model at
158. 6:43this level significantly speeds up the
159. 6:45open-source community's efforts to catch
160. 6:47up with closed source models and clearly
161. 6:49lowers the cost of video generation. I
162. 6:52think this already counts as a major
163. 6:53generational leap for open source. I'll
164. 6:56continue testing the models techniques
165. 6:57and prompt writing methods in depth and
166. 7:00share more with you later. That is it
167. 7:01for this video. If you enjoyed it,
168. 7:04remember to like and subscribe. See you
169. 7:06next time.