MiniMax 영상에 섞이는 외계어 없애는 방법 — Transcript

Full transcript

1. 0:00When creating videos with MiniMax,
2. 0:02sometimes the output contains these
3. 0:04strange, nonsensical words.
4. 0:06Let’s take a look. No. Geuni darwo neo
5. 0:10geonjeo seuli nangsi seureo meogi galjan
6. 0:13uro. I won't film it.
7. 0:15I never wasted a single cent of my
8. 0:17father's money. Get out.
9. 0:19>> Yes. Instances like this, where it spits
10. 0:21out strange words that are neither
11. 0:23Korean, Chinese, nor Japanese, happen
12. 0:25quite often. Let’s look at another
13. 0:27example.
14. 0:29No, since you can take the inheritance
15. 0:31whenever...yes.
16. 0:32Sister, do you really have to talk about
17. 0:34that right now?
18. 0:36We're in front of Father's portrait.
19. 0:38>> Yes.
20. 0:39After spending so much time making a
21. 0:41video, getting results like this is
22. 0:43incredibly stressful.
23. 0:45I’ve conducted various tests to resolve
24. 0:48this recurring problem.
25. 0:50I tried using the base diffusion model,
26. 0:53Singularity, Dinomix, and other models.
27. 0:57I also tried switching to Turbo,
28. 1:00changing the resolution, and increasing
29. 1:02the step count, but none of these tests
30. 1:05ultimately solved the issue.
31. 1:09At first, I thought it might be a prompt
32. 1:11issue, so I reviewed the official
33. 1:13prompting guide and had Claude analyze
34. 1:16it in detail to try and fix it, but that
35. 1:18didn't work either.
36. 1:20Frustrated, I wondered if the official
37. 1:23MiniMax commercial service model would
38. 1:25perform any better.
39. 1:28So, I paid for the MiniMax API and
40. 1:30tested it myself, and I was able to
41. 1:33confirm that it worked well.
42. 1:36Wondering why the commercial service
43. 1:38works well while ours doesn't, I
44. 1:40conducted additional experiments to find
45. 1:42the difference.
46. 1:45On the official MiniMax platform, there
47. 1:47is an intermediate agent model that
48. 1:50refines prompts before they go into the
49. 1:52video generation model, and it's a paid
50. 1:55service.
51. 1:57So, I invested about $ 50 to run some
52. 1:59tests.
53. 2:01By testing with roughly a few hundred
54. 2:04prompts, I began to focus on comparing
55. 2:07the differences between the prompts
56. 2:09designed by Claude and those provided by
57. 2:12the official service.
58. 2:14As a result, I discovered one major
59. 2:17difference. That difference was in the
60. 2:19length of the prompts.
61. 2:21Basically, when creating R2V videos,
62. 2:24I’ve always followed the six-step prompt
63. 2:27format.
64. 2:29However, the biggest difference I found
65. 2:31is in the summary description section.
66. 2:36While I was using Claude and entering
67. 2:39very detailed, specific, and long
68. 2:41content, I discovered that the official
69. 2:44MiniMax API prompt never uses long text
70. 2:47, but instead uses a very short,
71. 2:49single-line input.
72. 2:52So, to conduct a comparative test, I
73. 2:54generated about 200 videos.
74. 2:57For the ones where the speech sounded
75. 2:59like garbled Chinese, I kept everything
76. 3:01the same but changed only the prompt by
77. 3:03condensing the summary section, and I
78. 3:05was able to confirm that these issues
79. 3:07disappeared completely.
80. 3:10I don’t want to, you’re dragging me into
81. 3:13this, just leave me alone, you’re being
82. 3:15ridiculous, I won’t film it, I haven’t
83. 3:17wasted a single cent of my father’s
84. 3:20money, get out, I don’t want to, I won’t
85. 3:22film it, I haven’t wasted a single cent
86. 3:24of my father’s money.
87. 3:29Get out.
88. 3:32>> In the past, out of every ten videos I
89. 3:34made, two or three would consistently
90. 3:35have strange pronunciations, but after
91. 3:38changing the prompt format like this, it
92. 3:40was solved perfectly.
93. 3:42That is what I was able to verify.
94. 3:44So, I updated the instructions to
95. 3:46describe only what is shown in the video
96. 3:48in a short, simple, and clear way in the
97. 3:51summary section, and as this issue was
98. 3:53perfectly resolved, I no longer have to
99. 3:55deal with the stress of language-related
100. 3:57problems.
101. 3:59It feels like I spent about 100,000 won
102. 4:01in total to reach these test results,
103. 4:03considering all the many attempts I made
104. 4:06and the costs of paying for tokens
105. 4:08directly for testing.
106. 4:09Since I am sharing this all for free in
107. 4:12this video, I hope those of you who have
108. 4:14been stressed by this issue will find
109. 4:16this know-how helpful and put it to good
110. 4:18use. Yes. Yes. That’s all for today’s
111. 4:21video.