How I Make Realistic AI Videos with Seedance 2.5 - Full Guide — Transcript

Full transcript

1. 0:01Yo, what's good? It's your boy Dan. And
2. 0:03I'm going to show you how to make
3. 0:04realistic
4. 0:08>> [screaming]
5. 0:09>> This guy is insane.
6. 0:13Here is where we like to chill.
7. 0:24LOOK WHAT I HAVE.
8. 0:27LET'S TEST THAT
9. 0:29YEAH!
10. 0:30>> [screaming]
11. 0:33>> YEAH!
12. 0:44>> [screaming]
13. 0:48[music]
14. 1:07[music]
15. 1:20>> MANY PEOPLE ARE STILL STRUGGLING TRYING
16. 1:22TO generate realistic looking AI
17. 1:24footage. And that's why over the last
18. 1:26few weeks I've been putting together a
19. 1:27workflow of how you can achieve this.
20. 1:30We'll be using Higgs Field for this. In
21. 1:32my opinion this is one of the best AI
22. 1:33video platforms out there and they were
23. 1:35also kind enough to sponsor this video.
24. 1:37Okay, let's jump in. The question that I
25. 1:39get the most is which model should I
26. 1:42use? And this goes for both image and
27. 1:44video. And my answer to that really is
28. 1:47it depends. So, let me show you why it
29. 1:49depends on what you're trying to
30. 1:51generate. Because if you go over to
31. 1:53Higgs Field and you click on images,
32. 1:55then you can see we have all of these
33. 1:57different models. Like there's so many
34. 1:58different models you can choose from.
35. 2:00And it might be even a bit of an
36. 2:01overload. So, you might automatically
37. 2:03assume, "Oh, this is new. I should be
38. 2:06using that model." In my opinion, you're
39. 2:08wrong. So, let's take this example. I
40. 2:10have this prompt of a gangster-like
41. 2:13chimp, and he has like diamond grills
42. 2:14and all of that. Let's just hit
43. 2:16generate. I've put the same prompt into
44. 2:18six different AI image models to show
45. 2:20you the difference between them. And you
46. 2:22can choose for yourself which one you
47. 2:24like the most. Because my answer to this
48. 2:26is there are multiple good models that
49. 2:27you can use. You shouldn't stick to just
50. 2:29using one. For example, I really like
51. 2:32GPT image 2.5 flare here, Flux 2.0 Pro,
52. 2:36even though like Flux is a really old
53. 2:38model, and for comparison I put Flux 3
54. 2:40in there, too. Flux 2 is much better in
55. 2:42case in this realism. Nano Banana 2.1,
56. 2:45the newest model, also does a great job.
57. 2:47And SeaDream 5.0 also looks quite good.
58. 2:49XField Soul sometimes surprises me, but
59. 2:52that's why you don't want to stick to
60. 2:53just using one model and call it a day.
61. 2:56Images are cheap. Just generate many
62. 2:58different models all together at the
63. 2:59same time. Here we have another example.
64. 3:01Again, for all of them I used the same
65. 3:03prompt. And here you can see which one
66. 3:05you like the most yourself. Flux 3 feels
67. 3:07very plastic again. I really like the
68. 3:09look of GPT image 2.5 flare, and I think
69. 3:12Soul 2 here gives like an interesting,
70. 3:14like more realistic look to it as if it
71. 3:17was actually taken on like an iPhone at
72. 3:19like the morning sunrise. Next example,
73. 3:22again, we're using the same prompt. In
74. 3:24my opinion, GPT 2.5 flare looks best,
75. 3:27but I also like the look of SeaDream 5.0
76. 3:29Pro here, and I think the other ones are
77. 3:31not too bad, either. Here's example
78. 3:33number four, and here we are really
79. 3:34going for that disposable camera type of
80. 3:37look. I also asked it to include this
81. 3:39timestamp, so we have that in the image,
82. 3:41and Soul 2 and Flux 2.0 Pro didn't do
83. 3:44that because they're bad at generating
84. 3:45text on your images. But the look of
85. 3:48Flux 2 Pro is really good. I also like
86. 3:50GPT image 2.5 flare here, but it feels a
87. 3:53bit too blurry if you were to zoom in.
88. 3:56And then the last example here, I like
89. 3:58Flux 2.0 Pro the most. The only thing it
90. 4:00didn't do is follow up my prompt. I told
91. 4:02the AI to give it the finger and it just
92. 4:05put up a finger instead of putting up
93. 4:06the middle finger, but it looks insanely
94. 4:09good. Like it looks like it was actually
95. 4:10taken from a camera. So, what I'm trying
96. 4:13to say with this is that these models
97. 4:15are trained on like different data sets.
98. 4:17Some of them are better at certain
99. 4:19things. For example, for realism, Flux
100. 4:212.0 Pro is still quite good. Nano Banana
101. 4:24Pro and Nano Banana 2.1 also are still
102. 4:27quite good. I maybe even prefer the Nano
103. 4:29Banana Pro look over Nano Banana 2.1,
104. 4:32but GPT image 2.5 flare surprised me
105. 4:35over using a tool like GPT 2.5 sunburst.
106. 4:38And that's why throughout this video, I
107. 4:40don't want you to stick to just using
108. 4:42one model. Try out different
109. 4:44generations. Image generation is cheap.
110. 4:47And especially these older models like
111. 4:48Flux, like Soul, like Sea Dream, they
112. 4:51are super cheap. GPT 2.5 and Nano Banana
113. 4:54may be a little bit more expensive, but
114. 4:55still compared to videos, it is cheap.
115. 4:58So, don't limit yourself to just
116. 5:00sticking to one image model and think
117. 5:02that that is the reason your image looks
118. 5:04good or bad. Let's talk about AI video
119. 5:06models because here it makes a bit more
120. 5:09of a difference which model you're
121. 5:10choosing for realism. I'll be showing
122. 5:12you four different generations and I
123. 5:14will tell you why. We have Sea Dream 2.5
124. 5:161080p. We have Minimax HD Max. We have
125. 5:20Flux Tree and we have Sea Dream 2.5 480p
126. 5:23draft mode. We're using draft mode
127. 5:25because it's cheap and sometimes even
128. 5:27for realistic footage, the draft mode,
129. 5:29because it's lower quality, helps with
130. 5:31that realistic aesthetic. But you could
131. 5:33also be using something like Kling 4.0,
132. 5:35which I don't have access to as of
133. 5:37recording this, but test it out. Don't
134. 5:39limit yourself to these models. So, for
135. 5:41the first generation, I don't use any
136. 5:43references because that's the test that
137. 5:45I want to do. And then the second test
138. 5:46will include references. For all of the
139. 5:48prompts that I'm using, including all of
140. 5:50the files, all of the assets, everything
141. 5:52I share in this video, I will leave a
142. 5:54link to my school community where you
143. 5:56can find everything per video, so you
144. 5:58don't have to take notes. You can just
145. 6:00go to it. You can copy my prompts. You
146. 6:02can analyze my style, and you can learn
147. 6:04from that. So, the goal for this prompt
148. 6:06is that we are doing a raw found footage
149. 6:09feel compilation of a gang of street
150. 6:11monkeys causing chaos in everyday human
151. 6:13places. It's kind of like Jackass style.
152. 6:15That is the video that we're going to
153. 6:17do, 10 seconds long. And then I've
154. 6:18described like all of the different
155. 6:19things that I wanted to do. Again, this
156. 6:21is not about the prompt, but I want to
157. 6:22show you what we're generating. So,
158. 6:24here's the result for Cidence 2.5 1080p.
159. 6:36And honestly, this is very good. Like,
160. 6:38you might already know Cidence 2.5 is
161. 6:41one of the best AI video models. If you
162. 6:43want that realistic look, I would use
163. 6:45this model if you can afford it because
164. 6:48it's quite expensive. Then, let's
165. 6:50compare it to Minimax H3 Max.
166. 7:02Unfortunately, Minimax H3 is way worse
167. 7:05compared to Cidence when it comes to
168. 7:06generating realistic-looking footage
169. 7:08with no references in it. Then I have
170. 7:11Flux 3.
171. 7:20>> [laughter]
172. 7:22>> Now, this one is quite good because this
173. 7:23one will get you like 80 or 90% of the
174. 7:26result of Cidence 2.5, but it is way
175. 7:29cheaper. So, for these type of videos,
176. 7:31Flux 3 works quite well. Then we have
177. 7:33Cidence 2.5 draft mode. This is in 480p.
178. 7:47>> Again, this is very similar to the 1080p
179. 7:49version, but the difference here is if
180. 7:51you're using this draft mode, you can
181. 7:53like regenerate over and over again
182. 7:55while using like the cheapest type of
183. 7:57generation, which is the 480p version.
184. 7:59And then once you really like the result
185. 8:00that you have, then you can upscale this
186. 8:02in 1080p. And then you only then pay
187. 8:05like a one-time fee to upscale your
188. 8:07video, which is used often to save
189. 8:09yourself some credits. Okay, so to
190. 8:11summarize, for image models, you just
191. 8:13want to play around, test out multiple
192. 8:15different image models. For video
193. 8:17models, seed and stupid 5 remains on
194. 8:19top. There's nothing that's going to
195. 8:20beat it, but you might want to try out
196. 8:22flux tree and also minimax H3 if you're
197. 8:24after that realistic look. And also try
198. 8:27out clink 4, but from my first look at
199. 8:29it, I found it less realistic than seed
200. 8:31and stupid 5. So, not too sure about it,
201. 8:33but I haven't tried it out myself yet.
202. 8:35Okay, now let's move over to the next
203. 8:37segment, which is learning how to prompt
204. 8:39for realistic-looking images. For this,
205. 8:41we'll be using the following method.
206. 8:43It's called the CTLT method. If you have
207. 8:46no idea what it means, let me explain.
208. 8:47The C stands for camera, the T stands
209. 8:49for tone, the L stands for lighting, and
210. 8:52T stands for texture. I've created this
211. 8:54template for you, which again, you can
212. 8:56find in all of the assets that are
213. 8:57linked in this video. So, why are we
214. 8:59using this formula? Basically, it is to
215. 9:01teach yourself to think about all of
216. 9:04these different parts of what goes into
217. 9:06your prompt. So, for example, you want
218. 9:07to give the camera settings such as the
219. 9:09lens that's been used, the framing. You
220. 9:11want to set the mood, the atmosphere,
221. 9:13and the emotional energy first. Then you
222. 9:15want to describe who's in the image,
223. 9:16what they're doing, and where it is
224. 9:18happening. Then you can also describe
225. 9:20the lighting and the texture, where you
226. 9:22specify physical details that interact
227. 9:24with the environment. Okay, let's do
228. 9:26this together as an example. So, for the
229. 9:28camera, I want to have a candid 35-mm
230. 9:31film photograph with a 12-mm fish-eye
231. 9:34lens. You could ask your LLM to help you
232. 9:36with this. And no worries, at the end of
233. 9:38this template, I have this LLM which I
234. 9:41we will go over in a second, which will
235. 9:43help you write this prompt. But we're
236. 9:45going to do a candid 35 mm film
237. 9:48photograph with a 12 mm fish eye lens.
238. 9:51It's slightly elevated side view with a
239. 9:53wide full body framing, tilted horizon,
240. 9:55and a slight motion blur. Then for the
241. 9:56tone, I want something weird. So I'm
242. 9:58going for a frantic, absurd, and darkly
243. 10:01comic, like an unexpected chase caught
244. 10:03by a passerby. So basically what I'm
245. 10:05describing is, we weren't meant to see
246. 10:07this. It's just like someone drove by or
247. 10:09someone passed by and took a picture of
248. 10:11it. Then next up, we're kind of like
249. 10:12describing the action of what we're
250. 10:14seeing. So I want to do a skeleton that
251. 10:16is wearing a dress made of raw meat that
252. 10:18is sprinting across the streets of Los
253. 10:20Angeles at night with three dogs chasing
254. 10:23close behind him. For the lighting, I
255. 10:25want to do a harsh direct on camera
256. 10:26flashlight on the skeleton. And then for
257. 10:28the texture or the details, we're adding
258. 10:30in matte ivory bones, glossy marbled
259. 10:33meat with visible fat and uneven edges
260. 10:35flapping. Like this kind of get
261. 10:36disgusting to describe this, but it's
262. 10:38just describing what we want to see in
263. 10:41terms of the details. Okay, so now if we
264. 10:43combine this prompt, like all of these
265. 10:45different segments, and we copy that, we
266. 10:48go over to image, and then we create an
267. 10:50image, and then we drop in that prompt
268. 10:52right here. Then we can select our
269. 10:54model. Let's do something funky. I think
270. 10:57Flux 2 might be good for this. I'm going
271. 10:59to use Flux 2 Pro, and now we can select
272. 11:02our dimensions. Let's do 16 by 9, and
273. 11:04let's do 2K quality. Then we hit
274. 11:06generate, and that gave us this result.
275. 11:09Like this looks quite cool. This is what
276. 11:11I wanted, and this is how we prompt
277. 11:14simplified. Now I want you to know that
278. 11:15you don't have to be a filmmaker or
279. 11:17something like that to get prompts like
280. 11:19this. You could use an LLM, and I go
281. 11:22over to my ChatGPT or my Claude. So I've
282. 11:25put together this prompt for you. I
283. 11:27paste it in, and then I scroll down, and
284. 11:29here I paste in my idea. Or you just
285. 11:32whisper flow what your idea is. So, here
286. 11:34in this case, I could say like, "It's a
287. 11:37bystander taking a photograph with a
288. 11:39hard flash of a skeleton that is covered
289. 11:42in a meat dress being chased by three
290. 11:44dogs, and it's kind of like a image that
291. 11:47I'm not supposed to see." That is my
292. 11:48idea. You hit send, and then I will help
293. 11:51you flush out that idea using that
294. 11:53structure. There are certain words and
295. 11:55sentences that you can use to create
296. 11:57unique styles. For example, this is a
297. 12:00close skate camera type of image. If I
298. 12:03change up the wording a little bit, I
299. 12:05can also make some kind of like intimate
300. 12:07documentary. I can make a motion blurred
301. 12:10photo dump. I can do a low curb side
302. 12:12shot. We can do disposable camera look
303. 12:15or a concealed observation. As you can
304. 12:17see, these all have a totally different
305. 12:20visual identity and style to them. The
306. 12:22important thing is is that you just have
307. 12:24to understand how to come up with these
308. 12:26kind of different styles. That's why I
309. 12:28put together this document that is seven
310. 12:30ways to capture an idea, where we go
311. 12:32into different structures that you can
312. 12:35look at. And then you can analyze these
313. 12:37prompts to create shots like these. For
314. 12:40example, this phone snapshot is also
315. 12:42quite cool. So, I filled in the details
316. 12:44for you, so you can just look at it, and
317. 12:45you can copy these structures for your
318. 12:48own images. Let me just give you one
319. 12:49more demo of how I would do this with my
320. 12:51LLM. So, we start up by copying this
321. 12:54prompt again. Then we go into ChatGPT,
322. 12:57and then we add in the reference image
323. 12:59that we want to use. For example, I have
324. 13:02this reference image of Bruiser the
325. 13:04robot that I want to use for my video.
326. 13:07So, for my idea, I'm going to whisper
327. 13:09flow Bruiser sitting outside a
328. 13:11laundromat at night holding some fries,
329. 13:14and he is looking away. Really close low
330. 13:16angle skate photo, 12 mm fish eye lens,
331. 13:19and it's a bit crooked with some visible
332. 13:22film grain. Then we're going to hit
333. 13:24enter, and now it is starting to write
334. 13:27the prompt for us using this formula.
335. 13:29Then we copy this, we go back into Higgs
336. 13:32Fields, we paste in the link, and then
337. 13:34make sure you add in that reference
338. 13:35image. I'm using now Pro for this as
339. 13:38this works just as well. I just want to
340. 13:39show you that this also works. And then
341. 13:41we hit generate. And there we have it,
342. 13:44Bruiser sitting outside on the curb.
343. 13:45It's looking quite cool. Like I quite
344. 13:47like this one. Now, here's what the
345. 13:48style can do for this image. So, for
346. 13:50example, here I got the same image, but
347. 13:52I changed up the style. So, this time
348. 13:54I'm using a disposable camera wall flash
349. 13:56style, and we got a slightly different
350. 13:58image, but I might even like this one
351. 14:00more. Here again, now we're going for
352. 14:02the low quality camcorder style, and
353. 14:04it's badly compressed. So, I'm also
354. 14:06using here the lowest resolution. So,
355. 14:08I'm using the 1K resolution. Like on
356. 14:10purpose, I'm making this look worse
357. 14:12because imperfect looks more realistic
358. 14:15in this case. You can do this for any
359. 14:16type of style. So, here I made a iPhone
360. 14:19selfie type of style. Again, you can see
361. 14:21the prompt that I used beneath that to
362. 14:23get that type of style. And knowing
363. 14:25these different styles, you can make
364. 14:27some very unique-looking images. If you
365. 14:29don't know again how to prompt this, I
366. 14:31even put together this MD file for you.
367. 14:33I'm giving away like all of the assets
368. 14:35here. So, I'm hoping that this is useful
369. 14:37to you. Please let me know in the
370. 14:38comments if sharing these assets is
371. 14:40useful to you because I put in a lot of
372. 14:42extra time trying to make them universal
373. 14:44so you can use them too even if you have
374. 14:46no idea what you're doing. One more
375. 14:47thing I want to share with you, if you
376. 14:49have made an image like this one for
377. 14:50example, and you like the style of that,
378. 14:53you can even ask your LLM to make an MD
379. 14:56file, which is a skill file, for you.
380. 14:58So, for example, here I ask, "Can you
381. 15:00take the camera style based on the
382. 15:01prompt that gave me this image and turn
383. 15:04it into a reusable markdown file?" So, I
384. 15:06gave it the prompt, I gave it the image,
385. 15:07and then it made me this MD file. That's
386. 15:10how I make these type of things. You can
387. 15:12do that too. If you have made something,
388. 15:14you've no idea what you have done, you
389. 15:15just put in your prompt, put in your
390. 15:17image, and ask it to make that MD file
391. 15:19for you so it's using that type of style
392. 15:21next time you use your LLM. Now,
393. 15:23creating the images is actually the easy
394. 15:25part. The harder part is coming up with
395. 15:28creative ideas. And there are levels to
396. 15:30this. Let me explain to you. We're going
397. 15:31to talk about contrast because
398. 15:33cool-looking images like these are like
399. 15:35optimized for, I would say, social
400. 15:37media. These do amazing on like
401. 15:39Instagram. They all have contrast. You
402. 15:41see something going on there that you
403. 15:43look at and you're like, "What the hell?
404. 15:45What the What is that?" And that's what
405. 15:47I want to explain. I will identify
406. 15:48different type of levels of contrast.
407. 15:50So, level zero is no contrast. It's
408. 15:53nothing that is contrasting with the
409. 15:56thing that's in the image. So, for
410. 15:57example, it would be a robot in a
411. 15:59factory. Like nothing here is standing
412. 16:01out because this is the place, let's
413. 16:03say, where it belongs. Level one is in
414. 16:05an unexpected situation. So, here we
415. 16:07have that same robot, but now we put him
416. 16:09in an unexpected situation because he's
417. 16:12sitting down next to a Tesla charging
418. 16:14station. And it almost looks like he had
419. 16:16like one of those drunk nights out and
420. 16:18he's about to just juiced up so he can
421. 16:21continue with what he got. We could add
422. 16:23in even more contrast by going over to
423. 16:25level two, which is also adding in some
424. 16:27unexpected styling. So, now we have the
425. 16:30same robot sitting at the Tesla charger.
426. 16:32He's giving us the finger, and he also
427. 16:34has the I love Dan fully iced out type
428. 16:36of robot. This is culture. This is more
429. 16:38like rapper attitude, and now he has a
430. 16:41bit more personality. If we go one step
431. 16:44further, we're going to add in feeling.
432. 16:46Now, we are taking the example of the
433. 16:47robot again. We're putting him in a
434. 16:49field of flowers. Like you almost get a
435. 16:51feeling of like, "Oh, I feel sorry for
436. 16:53this robot. He didn't make it to that
437. 16:55Tesla charger." Those are the three
438. 16:57levels of contrast explained with
439. 16:59different images. And that is the thing
440. 17:01that you have seen throughout all of the
441. 17:02images and videos that I've shared
442. 17:04throughout this video, which is adding
443. 17:05in contrast. So, you have something cool
444. 17:08that stands out, which can help you go
445. 17:10viral. So, now that you understand
446. 17:12contrast and how to make images, let's
447. 17:14actually talk about how to make videos
448. 17:16cuz that might be even more important.
449. 17:18If you want to achieve realism, you want
450. 17:20to make it seem like someone actually
451. 17:22filmed that stuff, like a human filmed
452. 17:24it, not a robot. With Seid And Sue 2.5,
453. 17:26we can achieve that by making it look
454. 17:28like it filmed something like this.
455. 17:31>> They didn't invite me to the party
456. 17:33because apparently I filmed too much
457. 17:35like a robot. So, he found a
458. 17:37replacement.
459. 17:38>> He's great, actually.
460. 17:45At least I film like a human.
461. 17:48>> HEY, THAT'S MY CAMERA.
462. 17:51>> SO, let me explain camera techniques
463. 17:52that can be used to add in realism to
464. 17:55your generations. First, we have
465. 17:57handheld. This is where the camera has
466. 17:59small shakes. It's just like someone is
467. 18:01filming it using his phone and its
468. 18:03natural shakes. The camera also responds
469. 18:05to what's going on in the scene. Then we
470. 18:07have crash zoom. This is where the
471. 18:09camera is suddenly zooming in fast and
472. 18:12it's almost like the camera is surprised
473. 18:14something is happening and wants to
474. 18:15capture it really quickly. Then we have
475. 18:17voyeuristic. It's almost as if the
476. 18:20moment shouldn't even be captured. Then
477. 18:22we have whip pan. It's a quick turn
478. 18:24towards an unexpected event that makes
479. 18:26it look like the camera is reacting to
480. 18:28something.
481. 18:30>> Next up, we have walking backwards.
482. 18:31Here, it actually feels like the camera
483. 18:33operator is holding that camera and
484. 18:35walking like backwards. You can see them
485. 18:37like move unevenly with the steps.
486. 18:39Another one I love which is not used a
487. 18:42lot is blocking the frame. Here is where
488. 18:44something happens and you block the
489. 18:45frame by having someone like walk in
490. 18:47front of it. You really got to prompt
491. 18:49for this. Like you got to think about
492. 18:51this. And this is what you don't really
493. 18:53see with AI. This is what you can see
494. 18:55with real film things. So, that's why
495. 18:57you got to think like a director.
496. 18:59Another thing you can add in is exposure
497. 19:01adjustments. This is what happens
498. 19:02naturally when you go, for example, from
499. 19:04inside to outside and the camera got to
500. 19:07fix its exposure because it's all of a
501. 19:09sudden super bright. Then lastly, which
502. 19:11is one of my favorite, which is focus
503. 19:13field. Here's where the camera is
504. 19:14struggling to adjust and focus.
505. 19:22I will leave all of the prompts that I
506. 19:23just shared with you into the asset
507. 19:25folder, so you can go through this
508. 19:26yourself at your own pace. The best
509. 19:28thing you can do if you want to create
510. 19:30like real realism is use these different
511. 19:33techniques together, like combine them.
512. 19:35I put a few of them together and I got
513. 19:36this like office style video for it.
514. 20:12Shh.
515. 20:14Now, if you're trying this out yourself
516. 20:16and you're noticing that you're spending
517. 20:17way too much credits, I've been there,
518. 20:19then check out this video that's on the
519. 20:21screen right now, because this video
520. 20:22will teach you exactly how you can save
521. 20:24credits on AI videos.