MINIMAX H3 IS THE NEW FREE VIDEO AI KING! — Transcript

Full transcript

1. 0:00Minemax H3 is the best free video AI
2. 0:03king ever, and it's not even close.
3. 0:08>> [screaming]
4. 0:09>> Hello humans, my name is Kay Ovrlod, and
5. 0:11boy oh boy, do I have some mind-blowing
6. 0:12stuff for you today. Because yes, you
7. 0:14heard it right, we have a brand new
8. 0:16video AI model that was released called
9. 0:19Minemax H3. That is simply the best free
10. 0:23video AI model ever made, capable of
11. 0:26generating videos from text, images, as
12. 0:29well as from multiple other references
13. 0:32at the same time. It is just incredible.
14. 0:35So, today I'll show you how to install
15. 0:37it, how to run it locally, and on
16. 0:39RunPod, and show you how to get the best
17. 0:41results possible. So, that being said,
18. 0:43sit back, relax, and let's go. And to
19. 0:45install Minemax H3, you have two ways.
20. 0:48The first is of course by using my
21. 0:49one-click installer that is available
22. 0:51for my Patreon supporters. Just
23. 0:52double-click on the installer, and then
24. 0:54it will automatically install ComfyUI
25. 0:56and all the models and nodes that you
26. 0:57need to run Minemax on your computer.
27. 1:00And the second way is to rent a GPU on a
28. 1:02website like RunPod, and use my special
29. 1:04RunPod installer, and run Minemax as if
30. 1:06this was running on your local computer.
31. 1:08And then once you have ComfyUI up and
32. 1:10running, for this video I prepared a
33. 1:11special Minemax Ultra workflow that you
34. 1:14can find on my Patreon. Then you're
35. 1:15going to drag and drop inside ComfyUI.
36. 1:17And now we can finally have some fun.
37. 1:19Now, before we begin, try to explain the
38. 1:21workflow, what makes this Minemax H3
39. 1:24model so special, and why is everybody
40. 1:27and their grandma talking about it.
41. 1:29Well, Minemax H3 is a 33 billion
42. 1:31parameter model that can generate videos
43. 1:33from text or images in high resolution
44. 1:36and with native audio, and that can run
45. 1:39on your local computer. But that is not
46. 1:41all. This model is extremely powerful
47. 1:44and versatile thanks to its ability to
48. 1:46do video editing and using multiple
49. 1:48references at the same time to generate
50. 1:51your final video. Now, obviously,
51. 1:53Minemax H3 is a very thick boy, and you
52. 1:56really need a lot of VRAM to generate
53. 1:59videos at high resolution at a decent
54. 2:01speed. But don't worry, even if you have
55. 2:048 GB of VRAM, the model will work as
56. 2:06well. Okay, so enough talking, let's
57. 2:08actually show what this model can
58. 2:10actually do and see how good the video
59. 2:12generations are really are. Okay, so
60. 2:14first, let's start with the
61. 2:15text-to-video workflow, which is by far
62. 2:17the easiest to understand. Here, once
63. 2:19again, all of the most important part
64. 2:21will be located in the first column
65. 2:23right there. And before we start
66. 2:25generating, we need to look at the
67. 2:27available speed-up options that are
68. 2:29available for the model right now, which
69. 2:31are stage attention and spectrum. Now,
70. 2:34basically, all of these two allows you
71. 2:36to generate your video much faster at
72. 2:39the expense of a small hit in video
73. 2:42quality, especially if you're using
74. 2:44spectrum. Now, there's also plenty of
75. 2:46other nodes that can do that, but stage
76. 2:48attention and spectrum are by far the
77. 2:50best. Oh, and by the way, speaking of
78. 2:52speed-ups, this is entrepreneur in the
79. 2:54future because as I'm editing this
80. 2:56video, we actually already have the
81. 2:58release of the MiniMax Turbo Laura.
82. 3:01That's right. Now, obviously, right now,
83. 3:03it is still a Laura in training. It is
84. 3:05still not done. As of right now, the
85. 3:07only version available is the one at 500
86. 3:11steps. And to make it work, you need to
87. 3:13use the special comfy UI version. And
88. 3:15I've already tried it and it is very
89. 3:17good for a Laura that is not even done
90. 3:19training. And although in this video,
91. 3:21you will only see videos generated
92. 3:24without the Laura, just know that if
93. 3:26you're watching this video right now,
94. 3:27the workflow on Patreon will already be
95. 3:30updated to support this Laura. And you
96. 3:32can either choose to run it without the
97. 3:34Laura. And basically, the differences
98. 3:35between the workflow is very simple.
99. 3:38Without the Laura, we are basically
100. 3:40using the res multi-step sampler with
101. 3:43simple scheduler and 20 steps. Whereas,
102. 3:45if we are using the MiniMax Turbo Laura,
103. 3:49we will be using the Euler sampler with
104. 3:52beta scheduler at eight steps. That's
105. 3:54it. That is pretty much the only
106. 3:56difference. So, instead of generating a
107. 3:58video with 20 steps, you are now able to
108. 4:00generate your video in around eight
109. 4:03steps or even less. And of course, as I
110. 4:05said, this Turbo Laura is not even done
111. 4:07training, so in the upcoming days, we
112. 4:09should have an even better Turbo Laura
113. 4:12that we can use to generate absolutely
114. 4:14amazing videos. So, yeah, there you go.
115. 4:16Okay, so then right here, this is where
116. 4:18you're going to input your prompt for
117. 4:20the video generation. However, keep in
118. 4:22mind that MiniMax H3 has a very, very
119. 4:25specific prompting style. And I'm not
120. 4:27going to go into details on how it works
121. 4:30and everything. If you want more info,
122. 4:32you can simply either read the note I
123. 4:34have input right there, or if you want
124. 4:36to simplify your life, you can simply go
125. 4:38there and then copy and paste this
126. 4:40entire note inside software like
127. 4:42ChatGPT, and there you go. And now you
128. 4:44can simply just ask ChatGPT to write you
129. 4:47a prompt for whatever video you want.
130. 4:49Oh, and also, right before you generate,
131. 4:51let's also have a talk about the
132. 4:53resolution, because here you have
133. 4:55actually two options to choose the
134. 4:57resolution of your video. Either you use
135. 4:59the resolution selector that will
136. 5:01automatically select the resolution
137. 5:03following the aspect ratio guide right
138. 5:06there, or you can simply just leave this
139. 5:08option enabled and then choose yourself
140. 5:10the resolution that you want. Now, if
141. 5:12you don't want to type anything
142. 5:13yourself, it is much easier to just use
143. 5:16the resolution selector. And also here,
144. 5:18I have highlighted all the resolutions
145. 5:20that I recommend you to use, and it is
146. 5:22simply either 480p, 720p, and 1080p. So,
147. 5:27depending on your GPU, you might want to
148. 5:29choose either one of these three,
149. 5:31because these resolutions are more
150. 5:32standard, so they are much easier to
151. 5:34upscale. But obviously, if you have like
152. 5:36a 5090, you can definitely go to the
153. 5:381080p and then generate your huge video
154. 5:41from scratch. And then here, finally,
155. 5:43this is where you input the video
156. 5:44duration in seconds. And of course, the
157. 5:47longer the video, the longer it will
158. 5:49take for the video to be generated. So,
159. 5:51yeah, I mean, very simple stuff. If
160. 5:53you've done it once, you've done it a
161. 5:54thousand times. So, let me actually just
162. 5:56generate a video. Let me write a prompt.
163. 5:59Then I'm going to leave the 0.9
164. 6:00megapixel resolution, use 10 seconds
165. 6:03instead, and then I'm going to click
166. 6:04run, which in the end will give us
167. 6:06something like this.
168. 6:07>> I'm dead, Jerry. The city says I'm dead.
169. 6:10Well, that should clear up the parking
170. 6:12tickets.
171. 6:13Good news. I got you a box.
172. 6:16>> So, yeah, I mean, it's pretty good. Now,
173. 6:18as you can see, like this is very, very
174. 6:20decent. Now, this was generated in one
175. 6:22single shot. This is not cherry-picked
176. 6:24or anything. I mean, it is really,
177. 6:26really good. Actually, I even generated
178. 6:28another version with the same exact
179. 6:30prompt, but this is not at 1080p. Take a
180. 6:32look.
181. 6:34>> I'm dead, Jerry. The city says I'm dead.
182. 6:37Well, that should clear up the parking
183. 6:38tickets.
184. 6:39>> [laughter]
185. 6:40>> Good news.
186. 6:41I got you a box.
187. 6:43>> So, as you can see, in terms of quality,
188. 6:45except resolution, there aren't that
189. 6:47many differences between the two
190. 6:49generations. Hence why it's probably
191. 6:51better to generate at a lower
192. 6:53resolution, like 720p, and then upscale
193. 6:56it to a higher resolution later. And of
194. 6:58course, the model can do a lot of very
195. 7:00cool stuff, like a Breaking Bad parody,
196. 7:02for example, or memes, kind of like
197. 7:04this.
198. 7:06>> Jesse,
199. 7:07we have to cook.
200. 7:15>> Yeah, there you go. Like, super, super
201. 7:16funny
202. 7:18uh Breaking Bad, Baking Bread parody. We
203. 7:21all know that. It's not new, but it's
204. 7:23still really, really fun. And of course,
205. 7:25it can do more than just show parody or
206. 7:28real-life stuff. It can generate anime
207. 7:30parody as well. Like, take a look.
208. 7:34>> I'm going to become the king of the hook
209. 7:36again.
210. 7:43>> Yeah, I mean, this is really, really
211. 7:44incredible. Like this is insane. Like
212. 7:47you can make your own anime parody or
213. 7:49anime stuff already immediately right
214. 7:52now. Like we've just a simple text
215. 7:54prompt. Now this video was generated at
216. 7:571080p, so you get like the best quality
217. 7:59possible. But if you don't want to
218. 8:01generate at such a high resolution, you
219. 8:03can also generate at 720p. And if you
220. 8:06do, you get something like this.
221. 8:07>> I'm going to become the king of the
222. 8:09hook. Okay.
223. 8:17>> So yeah, I mean as you can see, very
224. 8:18decent as well. Maybe not as good as the
225. 8:211080p resolution, but it's still really,
226. 8:24really fantastic. I mean it's really,
227. 8:26really cool. So yeah, I mean I could
228. 8:27spend hours and hours generating new
229. 8:30videos to show you how good it is. But
230. 8:32I'm sure that by this time that you're
231. 8:34watching this video, you've already seen
232. 8:36multiple examples anyway. So let's move
233. 8:39on and see what else this model can do.
234. 8:41Now the second thing that this model can
235. 8:43also do is that instead of generating
236. 8:45videos from text, you can generate a
237. 8:47video from a image, which is definitely
238. 8:49something that I prefer. I do like to
239. 8:52generate my images on a separate model
240. 8:54like Create 2, and then using an image
241. 8:56to video model to bring it to life. I
242. 8:59find this much, much better. So here
243. 9:01basically the principle is exactly the
244. 9:03same, except that now you have the
245. 9:05choice of inputting two different
246. 9:07images. Either you input one single
247. 9:10image and you just make that image into
248. 9:13a video, like I'm going to show you
249. 9:14right now. Let's say I input this image,
250. 9:17this selfie of a warrior on a
251. 9:19battlefield, then I'm going to write my
252. 9:20prompt. Then here I either have the
253. 9:22possibility of using the original image
254. 9:25size, or if I don't want that because
255. 9:28it's going to take too much time, I can
256. 9:29just use the normal resolution selector
257. 9:32and then generate at the resolution
258. 9:34written right there, which is going to
259. 9:35be 720p instead. Input the video
260. 9:38duration and now if I click run, which
261. 9:40in the end will give us something like
262. 9:41this.
263. 9:47>> Well, I guess I'm going to be late for
264. 9:49bingo tonight.
265. 9:52>> And here, I mean as you can see, like
266. 9:53this is fantastic. This is really,
267. 9:55really good. This is great. I mean,
268. 9:57huge, amazing quality. The generation is
269. 10:00fantastic. The people behind him
270. 10:03walking,
271. 10:04you know, sitting down and everything. I
272. 10:06mean, this is really super, super
273. 10:08impressive. Super realistic, super good.
274. 10:11I mean, yeah, I mean, this is really,
275. 10:12really good. And of course, this is not
276. 10:14the only thing that you can do with this
277. 10:16particular workflow. And that is because
278. 10:17you also have the ability to input a
279. 10:20second image. Because this workflow can
280. 10:22also work as an in-between video
281. 10:25generation. So, if you put a first frame
282. 10:28and a last frame, you will be able to
283. 10:30generate a video between those two
284. 10:33beginning and end frames. So, like for
285. 10:35example, if I enable this last frame and
286. 10:38then I input frame one with the warrior
287. 10:40sitting on the ground, the last frame
288. 10:43where he is shouting, if now I write my
289. 10:45prompt and I click run, which in the end
290. 10:48will give us something like this.
291. 10:58>> [screaming]
292. 10:59>> So, yeah, I mean, as you can see, this
293. 11:00is really fantastic. I mean, we used
294. 11:02only two separate frames, one beginning
295. 11:06and one end, and with one single prompt,
296. 11:08Minimax made a video in between those
297. 11:12two frames. And I mean, I mean, what do
298. 11:14you want me to say? It looks really,
299. 11:15really good. It's it's fantastic. I
300. 11:17mean, LTX could do it, too, but Minimax
301. 11:20is really just on another level. It is
302. 11:23really super, super good. But of course,
303. 11:25once again, this is not the only thing
304. 11:27that this model can do. Because not only
305. 11:30can do all of that, but more, because
306. 11:32Mini Max also has a very specific video
307. 11:35model that allows you to use several
308. 11:38different references to make your video.
309. 11:40And the way it works is very, very
310. 11:42similar to the image to video workflow,
311. 11:45except that this time you have three
312. 11:47different types of references. You can
313. 11:49actually use image reference, video
314. 11:52reference, and also audio reference,
315. 11:55which really makes this workflow and
316. 11:58model extremely powerful and extremely
317. 12:00versatile. Now, the way you do this is
318. 12:02the same exact principle as what we did
319. 12:05with the image to video, except that
320. 12:07this time it's kind of up to you to
321. 12:08decide how many references you want to
322. 12:11use. And I think the limit is around
323. 12:13nine references that you can use at one
324. 12:15single time, which is really just
325. 12:17insane. Now, by default, I have input
326. 12:20here like two different references for
327. 12:22the images, one reference for the video,
328. 12:24and one reference for the audio, but if
329. 12:26you want to add more or disable
330. 12:28different references, you can do so very
331. 12:30easily as well. So, like for example,
332. 12:32let's say that I want to add an
333. 12:33additional reference image, all I have
334. 12:35to do is just click on one node, then
335. 12:38press control C on your keyboard, then
336. 12:40control V to copy and paste that node,
337. 12:43and then what you're going to do is that
338. 12:44you're going to click and then connect
339. 12:46it to that node right there. And you're
340. 12:48going to connect it to the reference
341. 12:50image node right there. And as you can
342. 12:52see, now we have three different
343. 12:54reference images that we can use. And
344. 12:56you can do the same thing with video,
345. 12:58audio, it's the same exact principle.
346. 13:01And if you don't want to use a certain
347. 13:02reference, you can just click on the
348. 13:04node, and then click on this little
349. 13:05button to bypass the reference, as well
350. 13:08as right there. So, like for example,
351. 13:09let's say that I upload an image of this
352. 13:12character sheet right there, then
353. 13:14another character for reference image
354. 13:17two, and then for the third image, the
355. 13:19interior of this castle right there. And
356. 13:21now, if I write my prompt, which once
357. 13:23again, I made it with ChatGPT, I'm not
358. 13:26going to write all of that myself. Like,
359. 13:28no way. And now, if I click run, which
360. 13:30in the end gives us something like this.
361. 13:32>> Wow. So, it can generate multiple
362. 13:35characters inside the video?
363. 13:37>> Yep, it looks like it.
364. 13:41>> Amazing.
365. 13:42>> Yeah, it sure is, buddy. I mean, this is
366. 13:44really just incredible. I mean, it's
367. 13:46insane. I mean,
368. 13:49>> [laughter]
369. 13:49>> I mean, I'm I'm kind of shocked. I'm
370. 13:51going to say it's It works so good. It
371. 13:54worked better than I thought it's going
372. 13:55to be. It really like took the
373. 13:57characters from the sheet, the character
374. 14:00sheet, all the images, and then made
375. 14:03video like this with full dialogue
376. 14:06interacting with two different
377. 14:07characters, two different voices, and I
378. 14:11mean, it looks so good. It's It's
379. 14:13amazing. I mean, yes, it was generated
380. 14:15at a lower resolution, so you can have
381. 14:17even higher quality if you generate at a
382. 14:19higher resolution, or if you upscale it.
383. 14:22But, I mean, just like that already, it
384. 14:25is It is really, really good. I mean,
385. 14:27it's It's incredible. I got to say, it's
386. 14:29insane. It's fantastic. You don't need
387. 14:31any LORAs, you don't need any training.
388. 14:34Everything is already done for you. Just
389. 14:36Just incredible. And also, of course,
390. 14:38there's plenty of other ways of using
391. 14:40this, using different types of
392. 14:42references. Like, for example, let's say
393. 14:44that I want to use a video reference.
394. 14:47Let's say I upload this video of this
395. 14:49woman kind of like turning around, and I
396. 14:51want to like edit this video. I want to
397. 14:54replace that woman with a different
398. 14:56character. Let's say I want to replace
399. 14:57it with that girl. I'm actually going to
400. 14:59upload that girl into image zero. Then,
401. 15:02I'm going to disable all the other
402. 15:04references. I'm going to write my
403. 15:06prompt, check the video direction. And
404. 15:08now, if I click run, which gives us
405. 15:10something like this.
406. 15:16I mean, listen. I mean, is it amazing or
407. 15:19what?
408. 15:20What do you want me to say? It's It's
409. 15:21incredible. I mean, we literally just
410. 15:24like took a video reference and then
411. 15:27replace it with a character from an
412. 15:31image. No Laura training, no anything.
413. 15:35We just like put everything together and
414. 15:38it just worked. It just works. All
415. 15:40right, it's
416. 15:44Huh? It It's I'm I'm blown away. I'm
417. 15:46blown away. If you can tell, I'm blown
418. 15:48away. It is incredible. I mean
419. 15:51And obviously, this is once again not
420. 15:54the only thing that we can do. We can do
421. 15:57much, much more than that. We can do so
422. 15:59much more. I can spend hours and hours
423. 16:02and hours showing you everything that
424. 16:05you can do. Like for example, you can
425. 16:07even like simply use like one single
426. 16:09video as a reference. If I use like this
427. 16:11video from, you know, Soprano with
428. 16:13Paulie talking, which I'm not going to
429. 16:15play because I kind of want to avoid
430. 16:18any, you know, copyright issues if there
431. 16:21is any. But what I can do instead is
432. 16:23just use that video as a reference and
433. 16:25then make him say something completely
434. 16:26different with my prompt. And because
435. 16:28the video not only can reference the
436. 16:32video itself, but also the audio, we
437. 16:34should also be able to copy his voice
438. 16:37fairly well. So now if I click run, so
439. 16:40that in the end we get something like
440. 16:41this.
441. 16:42>> Wow, this model is incredible. If Tony
442. 16:45hears about this, he's going to be
443. 16:46pissed.
444. 16:47>> So yeah, I mean, what what do you want
445. 16:49me to say? It's It's great. It's
446. 16:51fantastic. It's
447. 16:53It's everything that we've ever wanted
448. 16:54from a video model. It's It is that
449. 16:57simple, guys. It is that simple, okay?
450. 17:00Like Why Why Why are you still here?
451. 17:03Just Just stop listening to me and, you
452. 17:06know, like and and do your thing. Do
453. 17:08your thing. Try it out. It It's It's
454. 17:10amazing. It's amazing. Oh, and also one
455. 17:12little trick. If you want to upscale
456. 17:14your video from a lower to a higher
457. 17:16resolution, I highly recommend using my
458. 17:18LCX 2.3 ultra workflow version 3 because
459. 17:22in that workflow I have added a video
460. 17:24enhancer upscaler, which is actually
461. 17:27really really good at enhancing and
462. 17:30upscaling at the same time your video.
463. 17:32So, you can literally just upload a
464. 17:34video made with Minimax. This is a video
465. 17:36that was made and generated at 720p, and
466. 17:40then you can like input a resolution of
467. 17:42full HD, for example, and then if you
468. 17:45click run, it will use LCX 2.3 to take
469. 17:48this video and then upscale it at a
470. 17:50higher resolution and enhance it at the
471. 17:53same time. So, that in the end we go
472. 17:55from this, a 720p video that looks like
473. 17:58this,
474. 18:05there we go, to a 1080p video.
475. 18:13Yeah, I mean, there is really a huge
476. 18:15difference in quality between those two
477. 18:18videos. Hence why this workflow is
478. 18:20really insanely good because not only we
479. 18:23are upscaling, but we are also enhancing
480. 18:27the original video. And obviously, it
481. 18:29takes way less time to generate at 720p
482. 18:32and then upscale it at 1080p rather than
483. 18:35generating the full video at 1080p
484. 18:37instead. So, yeah, I definitely
485. 18:39recommend you to use this upscaler
486. 18:41whenever you want to upscale one of your
487. 18:43videos made with Minimax. Now, the one a
488. 18:46little bit of a weird stuff with this
489. 18:48model is the fact that the license is
490. 18:51very very strange. Basically, it says
491. 18:53that you cannot really use the model if
492. 18:56you are in the European Union, United
493. 18:59Kingdom, Korea, or the United States,
494. 19:02which is, you know, very very strange.
495. 19:04But, that is because Minimax is
496. 19:06currently in a lawsuit against Disney,
497. 19:09and they kind of want to avoid any
498. 19:10issues with the model. Now, if you want
499. 19:13to make sure that everything is fine,
500. 19:14you can simply click on this little link
501. 19:16and just like sign a waiver giving you
502. 19:19the right to use this model however you
503. 19:21want. It takes like 30 seconds to do and
504. 19:24then you're good to go. Now, obviously,
505. 19:26if you are like a normal person, you
506. 19:28don't necessarily need to do that. I you
507. 19:30know, you're fine. But in my case and a
508. 19:32lot of people's cases, since we do have
509. 19:34companies and we do represent companies
510. 19:37ourselves, just to make sure that
511. 19:38everything is fine, we need to kind of
512. 19:40go through this, sign that waiver, and
513. 19:43everything is going to be fine. So, yes,
514. 19:45YouTube, if you're watching this video,
515. 19:47just know that I did receive
516. 19:49confirmation and I do have the right to
517. 19:52use this model. Thank you very much. I
518. 19:54love you. No sarcasm intended. So, yeah,
519. 19:57there you go. This has been Min Max H3,
520. 20:00simply the best open weight video AI
521. 20:03model ever made. An absolute beast of a
522. 20:06model that you can run on your computer
523. 20:08right now. And once again, in this
524. 20:10video, I barely scratched the surface of
525. 20:13what this model can do. And even with my
526. 20:15version one of the Min Max workflow, you
527. 20:17can pretty much do absolutely insane
528. 20:19stuff. But don't worry, this is
529. 20:21definitely not the last time that I'll
530. 20:22be making an update video about this
531. 20:25model. There is really too much to say
532. 20:27and too much to do. So, yeah, like once
533. 20:30again, stop wasting time, download this
534. 20:32workflow, download this model right now
535. 20:34either locally or on RunPod, and just
536. 20:37use it. It is incredible.
537. 20:41Insane. So, that you and I can finally
538. 20:44make our dreams come true.
539. 20:51>> [music]
540. 20:56>> And there we have it, folks. Thank you
541. 20:58guys so much for watching. Don't forget
542. 20:59to subscribe and smash the like button
543. 21:01for the YouTube algorithm. Thanking also
544. 21:03so much to my Patreon supporters for
545. 21:04supporting my videos. You guys are
546. 21:06absolutely awesome. You people are the
547. 21:07reason why I'm able to make these
548. 21:09videos, so thank you so much, and I'll
549. 21:11see you guys next time. Bye-bye.