How To Create Ultra-Realistic AI Videos (Full Workflow) — Transcript

Full transcript

1. 0:00This isn't realistic UGC.
2. 0:01>> thinner. It's actually making you
3. 0:03fatter.
4. 0:04>> And this [snorts] isn't realistic UGC.
5. 0:06>> Supplements are a myth, but I've had a
6. 0:08stupid transformation.
7. 0:09>> I'll tell you more. This isn't realistic
8. 0:11UGC, too.
9. 0:11>> And you only need to do three things.
10. 0:14>> That is one big pile of
11. 0:15>> Okay, so what is realistic UGC? Let me
12. 0:18show you an example.
13. 0:21>> Okay, I just spent the last of my money
14. 0:22on this tiny jar of La Mer.
15. 0:25Was that responsible? Absolutely not.
16. 0:28>> [laughter]
17. 0:28>> But look at it.
18. 0:30I haven't even tried it yet and I'm
19. 0:31already emotionally invested. It better
20. 0:33be good. I'll use it for a bit and tell
21. 0:35you if it's actually worth the hype.
22. 0:37>> There is no plastic skin, no dead eyes,
23. 0:40no robotic voice or desaturation. She
24. 0:42got realistic eyes with emotions behind
25. 0:44them, realistic voice with depth, her
26. 0:46mimic and hand movements are super
27. 0:48natural, and hairs are not plasticky.
28. 0:51So, what I'm going to teach you today is
29. 0:52a completely new method to achieve this
30. 0:54level of realism in your UGC in
31. 0:56literally easiest way possible, so you
32. 0:58can avoid this uncanny valley look. What
33. 1:00uncanny valley even is. The uncanny
34. 1:02valley is a concept proposed by a
35. 1:03Japanese roboticist called Masahiro Mori
36. 1:06in 1970. His idea was that as a robot
37. 1:08starts looking more human, we can feel
38. 1:10more comfortable with it. But, when it
39. 1:11gets very close to looking human, while
40. 1:13still being slightly wrong, that feeling
41. 1:16can suddenly turn into discomfort. That
42. 1:18drop is the valley. He was talking about
43. 1:20robots, but it's a useful way to think
44. 1:21about AI people on screen, too. A
45. 1:23cartoon character gives you one set of
46. 1:25expectations. A face that looks like a
47. 1:26real person gives you a different set.
48. 1:28And when that face starts moving in a
49. 1:30strange way, those little mismatches
50. 1:32become much harder to ignore. With AI
51. 1:34UGC, there are a few things I always
52. 1:36look at. The skin can be too smooth, no
53. 1:38variation in texture, the eyes can feel
54. 1:40fixed like a person is staring through
55. 1:41the camera. The mouth might follow the
56. 1:43words, but the rest of the face barely
57. 1:45responds. Then there's the voice, the
58. 1:47pacing, the pauses, the emphasis. You
59. 1:49can also have a voiceover that feels
60. 1:50completely disconnected from the person
61. 1:52and the room you're looking at. Sound
62. 1:54still needs to fit the video. Like if
63. 1:55she's in bathroom, then we want to hear
64. 1:57an echo.
65. 1:58>> Okay, so I'm literally filming this in
66. 2:00my bathroom right now. Can you hear how
67. 2:02echoey it is in here?
68. 2:03>> If she's outside, we want to hear cars
69. 2:05driving by and a bit of wind.
70. 2:07>> Okay, it is so windy out here today. I'm
71. 2:10just walking to get coffee.
72. 2:12>> And lighting is another big one. If
73. 2:14someone is sitting in a normal kitchen,
74. 2:15but their face looks like it's being lit
75. 2:17for a beauty campaign, the whole scene
76. 2:19feels staged. Before we go any further,
77. 2:21you don't need perfect realism to make
78. 2:23an ad that sells. I still see horrible
79. 2:25looking ads converting like crazy every
80. 2:27single day.
81. 2:28>> Bro, what are you talking about, man?
82. 2:29>> They work really well. The idea can be
83. 2:31entertaining, the product can be easy to
84. 2:32understand, and the offer can still make
85. 2:34sense. So, don't spend days chasing
86. 2:36perfect skin texture if realism isn't
87. 2:39important to the concept. In that
88. 2:40format, believable movement, lighting,
89. 2:42and delivery matter to me because I want
90. 2:44people watching the video to feel like
91. 2:46they're watching a really good TikToker.
92. 2:48Because if it looks even a little bit
93. 2:50off, it feels like a really bad actor
94. 2:52playing a role.
95. 2:53>> Oh, hi, Mark.
96. 2:54>> But with AI character, this feeling is
97. 2:56multiplied by a hundred X. All right, so
98. 2:58now I'm going to show you the whole
99. 2:59workflow. We'll be doing it in Max
100. 3:00Fusion. You can use whatever site you're
101. 3:02on right now, but honestly, Max Fusion
102. 3:04is one of the best tools I've ever used,
103. 3:05so I'm sticking with it. First thing we
104. 3:07need is a reference image for our video.
105. 3:09So, as you can see, we've got our prompt
106. 3:11here. I'll explain how it works a bit
107. 3:12later. Then we've got the reference
108. 3:13image, the reference image of our
109. 3:15product, and the reference video. And
110. 3:17this one is super, super important. The
111. 3:19reference video is the video we're
112. 3:21trying to imitate. The style, the
113. 3:23emotions, the movements, we basically
114. 3:25want to copy all of that onto our video.
115. 3:27I always find mine on TikTok. You can
116. 3:29also go on Pinterest, type UGC video,
117. 3:32and you'll get tons of stuff like this.
118. 3:34But they're pretty generic, and you
119. 3:35really have to be careful because
120. 3:36Pinterest is full of AI videos now.
121. 3:39So, I definitely recommend going on
122. 3:41TikTok and finding them there. And now
123. 3:42pay attention because this is the
124. 3:44important part. The reference image is
125. 3:45like 50% of your video. Actually, forget
126. 3:48that.
127. 3:49It's more like 80%. It decides whether
128. 3:51your video comes out real, whether it
129. 3:53comes out good or not. So, I'd spend at
130. 3:55least half of your time just perfecting
131. 3:56the first frame. Let me click here and
132. 3:58show you what it looks like. This is our
133. 4:00first frame. There are a few ways to
134. 4:01make a perfect first frame. The first
135. 4:03one, go to Pinterest or TikTok. Type
136. 4:05something like UGC video, girl yapping,
137. 4:08whatever, and find a girl that's not AI.
138. 4:11Screenshot it. So, let's say we want to
139. 4:13use this one. We just screenshot it,
140. 4:14bam. We go to GPT and paste it in
141. 4:17together with an image of our character,
142. 4:19and we type something like change the
143. 4:20girl from image one to the girl from
144. 4:22image two. Keep everything else,
145. 4:24including the background, exactly the
146. 4:26same as in image one. Hit enter and wait
147. 4:28a couple of seconds. All right. As you
148. 4:30can see, the first time it generated the
149. 4:32image with these overlay icons on top
150. 4:34because I forgot to put that in the
151. 4:35prompt.
152. 4:36So, I added it right here, and now our
153. 4:38image is done. But, if you look really
154. 4:40closely, you can see the eyes are a bit
155. 4:42over-polished, a bit plasticky, and
156. 4:44overall the face is kind of mid in terms
157. 4:46of quality. You can clearly tell it's
158. 4:48AI. So, now I'm going to show you the
159. 4:49best method to create super, super
160. 4:51realistic first frames for your videos.
161. 4:53So, stay with me.
162. 4:55All right. So, the second method is to
163. 4:57just screenshot the first frame from the
164. 4:58video you downloaded from TikTok. Upload
165. 5:00it right here and tell GPT, "Give me a
166. 5:02detailed JSON prompt for this image."
167. 5:05Down below, GPT writes a really, really
168. 5:08detailed JSON prompt, as you can see. We
169. 5:10copy it, go back to Max Fusion, click
170. 5:12images right here, and choose Nano
171. 5:14Banana Pro. Why not GPT image two?
172. 5:17Because GPT image two was trained on
173. 5:19photos from newer iPhone cameras, so its
174. 5:21images come out over-polished, way too
175. 5:23perfect, and that's exactly what we want
176. 5:25to avoid.
177. 5:26So, we're using Nano Banana Pro because
178. 5:28it was trained on photos from older
179. 5:30phones, like the iPhone 8, 10, and 11.
180. 5:33Those look really natural, and it gives
181. 5:35you an image that actually looks like
182. 5:37it's from TikTok. You paste the prompt
183. 5:39right here. As you can see, I've already
184. 5:41done it a couple of times. Here's the
185. 5:42JSON prompt, and our image is way, way
186. 5:44better than the one from GPT after
187. 5:46swapping the faces. It's incredible. The
188. 5:49details here are out of this world. The
189. 5:51depth on this image, there's just so
190. 5:53much going on. You can even see the
191. 5:55lights from her TikTok lamp in her eyes.
192. 5:57This is just crazy. And the third
193. 5:59method, which is just as good as this
194. 6:01one, you find a video on TikTok or
195. 6:03Pinterest, screenshot the first frame,
196. 6:05same as before, upload it to GPT and
197. 6:08turn it into a detailed JSON prompt. So,
198. 6:10it starts exactly the same, but now you
199. 6:12copy that JSON prompt, go to Max Fusion,
200. 6:15pick C Dance 2.5, references 9x16, 4
201. 6:19seconds, 1080p. Paste the prompt right
202. 6:22here and you'll get videos like this.
203. 6:24Then you just screenshot one frame from
204. 6:26this video and you get an even better,
205. 6:28even more realistic first frame. Oh, and
206. 6:31one more thing. Above that JSON prompt,
207. 6:33you can type something like, "Give me a
208. 6:35rapid still motion photo slideshow of
209. 6:37different people per your reference,
210. 6:39hard cuts every fraction of a second."
211. 6:41Then paste the prompt and you'll get a
212. 6:43video like this with four different
213. 6:45people every second. And as you can see,
214. 6:47every single one of them looks super,
215. 6:49super realistic. Then you just
216. 6:51screenshot the frame you like and use it
217. 6:53as the reference image for your video.
218. 6:55Let me show you right here. This is the
219. 6:56reference video we'll be using. We
220. 6:58downloaded it from TikTok. Okay, so now
221. 7:00let me explain how this prompt works and
222. 7:02how to make one. The template for this
223. 7:03prompt is linked down below in the
224. 7:05description in the free resources. It's
225. 7:07completely free. You just click it and
226. 7:08copy it. Then you paste the template
227. 7:10into GPT, explain what you want to see,
228. 7:12and upload your reference image, your
229. 7:14reference video, and your product image,
230. 7:16and you'll get a prompt like this. The
231. 7:18first part of the prompt is format and
232. 7:19visual references. It explains the
233. 7:21length of the video, which image to use
234. 7:23when, which video to use when, the
235. 7:25voice, the accent, and so on. Then it's
236. 7:28body language and movements, the
237. 7:29posture, the expressions, the gestures,
238. 7:31the face movements. Then it's the
239. 7:33dialogue, exactly what we want her to
240. 7:35say. Then voice and emotions. That tells
241. 7:37C Dance what her voice sounds like, her
242. 7:39emotions, her reactions, then product
243. 7:41interaction. For example, as she starts
244. 7:44speaking, she reaches past the side of
245. 7:45the camera and grabs the cream from
246. 7:47right behind it. So, basically, how the
247. 7:49product gets introduced into the video
248. 7:51and where it's coming from. And then
249. 7:52timing and realism, camera and
250. 7:54restrictions. So, it's really, really
251. 7:56simple. You paste my template into GPT
252. 7:59and you get a prompt like this
253. 8:00automatically. Then you upload your
254. 8:01image, your video, your product, paste
255. 8:04your prompt, click generate, and you get
256. 8:06a video like this.
257. 8:08>> Okay, I just spent the last of my money
258. 8:10on this tiny jar of La Mer.
259. 8:12Was that responsible? Absolutely not.
260. 8:15But, look at it.
261. 8:17I haven't even tried it yet, and I'm
262. 8:19already emotionally invested. It better
263. 8:21be good. I'll use it for a bit and tell
264. 8:23you if it's actually worth the hype.
265. 8:26>> It's incredible. Honestly, it doesn't
266. 8:28get much better than this. This is the
267. 8:30best method we have right now to create
268. 8:32realistic AI UGC. So, to sum it all up,
269. 8:35for this kind of realistic UGC, Sea
270. 8:38Dance 2.5 is my first choice right now.
271. 8:41If a client is skeptical about AI, this
272. 8:43is the kind of result I'd want to show
273. 8:44them. And let me be honest with you. It
274. 8:46isn't the cheapest option. But, for a
275. 8:48project where realism really matters,
276. 8:50it's the one I'd start with.
277. 8:52Wuan 3.0 is also worth testing. You can
278. 8:55find it right here. If you want a
279. 8:56cheaper alternative, try the same idea
280. 8:59there and compare the results. For some
281. 9:01shots, you might actually prefer what
282. 9:02Wuan gives you because it's really good
283. 9:04and a lot cheaper. Just check the cost
284. 9:06for the same settings. And honestly,
285. 9:08those are the only two options right
286. 9:10now. Wuan 3.0 and Sea Dance 2.5. With
287. 9:13something like Flux, you can't use a
288. 9:15video as a reference, only images. So,
289. 9:17even though it's cheaper, this method
290. 9:19just doesn't work there. So, quick recap
291. 9:21of the settings. Videos, Sea Dance 2.5,
292. 9:24references, 9 by 16 or 16 by 9 depending
293. 9:27on what you need, and the length of the
294. 9:29video, you pick that right here. 720p is
295. 9:32more than enough, and that's it. That's
296. 9:34all you need to generate super realistic
297. 9:36AI UGC. The part I'd spend the most time
298. 9:39on is getting the right reference, a
299. 9:41really high-quality one. Pick something
300. 9:43that already feels close to the video
301. 9:44you want to make. Give it a clear role
302. 9:46in the prompt, write dialogue that
303. 9:48sounds natural out loud, and review the
304. 9:50result as a moving video with sound.
305. 9:53Then, check if it actually does the job
306. 9:54you made it for. And that's pretty much
307. 9:56it. See you in the next video.