힉스필드는 끝났다? 같은 장면 0원으로 다시 만드는 미친 로컬AI ㄷㄷ — Transcript
Full transcript
- 0:00I created the same scene using both the
- 0:03paid AI Kling 4.0 and the free local AI
- 0:06ComfyUI.
- 0:07One cost over 20,000 won, and the other
- 0:10cost zero won for video generation. This
- 0:13video is a senior-themed short-form clip
- 0:15titled "Mom's Seat." It is a story about
- 0:18the emotions a mother experiences at a
- 0:20wedding. First, please look at the
- 0:22character's expressions and gaze.
- 0:25>> It's my daughter's wedding.
- 0:27Now, let's look at the movement of the
- 0:29daughter leaving.
- 0:33Finally, let's compare the mother's
- 0:34expression.
- 0:36>> Can I sit there?
- 0:38>> Which one feels more natural and conveys
- 0:40the emotion better to you?
- 0:43The left side was made with the paid AI
- 0:46Kling 4.0, which used about 2,040
- 0:49credits for 14 clips.
- 0:51What does this look like if we convert
- 0:53it to Korean won?
- 0:54Based on the actual price we paid to
- 0:56recharge, 1 credit is about 10 to 15 won
- 0:59.
- 1:00So, this one short-form video cost about
- 1:0220,000 to 30,000 won to produce.
- 1:05Looking at the breakdown calculated by
- 1:07the creator, about 45%was spent on
- 1:09remaking or refining the clips.
- 1:12When you aren't satisfied with the
- 1:14result, the costs for re-generation add
- 1:16up. The right side was made in a local
- 1:19environment running AI directly without
- 1:21external services, so the extra cost for
- 1:24video generation was zero won.
- 1:26This local version was made with LTX 2.5
- 1:28, and in a local environment, you can
- 1:31also run video models like MiniMax H3
- 1:33directly on your own computer.
- 1:36This means you don't have to pay for
- 1:38external services to generate videos on
- 1:39a per-clip basis. I will also explain
- 1:41the comparison conditions. The local
- 1:43version used the same starting image as
- 1:45the paid version, and used the same
- 1:47voice and subtitles. Computer,
- 1:49electricity, and some voice-related
- 1:51costs are not included in this "zero won
- 1:53."
- 1:54I will show you why this difference
- 1:56matters and why it eventually leads to
- 1:59the story of AI companies.
- 2:01There are so many more people watching
- 2:03short dramas and video content these
- 2:05days. According to a survey, the number
- 2:08of users of major domestic short-form
- 2:10drama apps increased 3.3 times in one
- 2:13year, from about 900,000 to 2.98 million
- 2:16. The market outlook is also significant
- 2:18.
- 2:18The industry expects the global
- 2:20short-form drama market to grow from
- 2:22about 22 trillion won this year to about
- 2:2535 trillion won by 2030.
- 2:27The trend of creating stories into
- 2:29videos using AI is also accelerating.
- 2:32However, just because the market is
- 2:33growing doesn't mean everyone is making
- 2:35money.
- 2:36That's because, in addition to
- 2:38production costs, expenses for
- 2:39advertising to people are also involved
- 2:41.
- 2:42If you upload one video a day for 30
- 2:44days, it costs about 600,000 to 900,000
- 2:46won per month. If you add the monthly AI
- 2:49subscription fees you're already paying
- 2:51to that, the burden grows even larger.
- 2:55If production and marketing costs are
- 2:56higher than the revenue the video earns
- 2:58, losses accumulate the more you make.
- 3:00Ultimately, a business must move toward
- 3:02a structure that reduces costs and
- 3:03increases profits to survive for a long
- 3:05time. If you can make the same video for
- 3:08less, that difference becomes monthly
- 3:11savings.
- 3:13Yet, many videos recommend signing up
- 3:16for paid AI video services like Pika or
- 3:19sponsored AI tools.
- 3:22But signing up doesn’t automatically
- 3:24create videos, and the fees keep
- 3:26recurring every month.
- 3:28As you saw today, you don't necessarily
- 3:30have to do that.
- 3:32However, installing AI in a local
- 3:34environment and asking it to "make a
- 3:36video" doesn't immediately yield the
- 3:38desired result.
- 3:40To run these models in ComfyUI, you have
- 3:43to connect nodes one by one and build
- 3:45the workflow yourself.
- 3:47Since even a single scene is tangled
- 3:49with so many nodes, it is truly
- 3:50difficult for a person to adjust them by
- 3:52hand every time.
- 3:54That’s why methods are emerging where
- 3:56trending AIs like GPT-6 Astra operate
- 3:59ComfyUI on your behalf. But every time
- 4:02you fix a scene, it consumes a large
- 4:04amount of tokens, so now those costs
- 4:06pile up instead. Because you have to
- 4:08match the character's face, movement,
- 4:10the emotion of the scene, and the flow
- 4:12between them.
- 4:14The left is the first local result
- 4:16before editing, and the right is the
- 4:18result after I provided feedback six
- 4:20times.
- 4:21We used the same starting image and
- 4:23audio, but the character’s movements and
- 4:25expressions have changed.
- 4:30>> May I sit there? The company saved that
- 4:34feedback as a reference document to be
- 4:36used in future projects. So that we
- 4:38don't have to explain it from scratch
- 4:40every time.
- 4:41That is why we delegated this complex
- 4:44work to AI employees. Employees inside
- 4:46the game are sitting there. But these
- 4:48employees actually work. I assigned them
- 4:51the task of finding products to use for
- 4:53shopping shorts.
- 4:55The AI employee researches candidates
- 4:57and reports the results along with files
- 5:00.
- 5:01Even without a person holding onto the
- 5:03nodes, once you assign tasks via the
- 5:04game screen, the AI employees handle the
- 5:07rest.
- 5:08As an AI-native company, we connected
- 5:10the entire process from planning to
- 5:13scene breakdown, creation, review, and
- 5:15revision. A great result is not just the
- 5:18product of a single model, but the
- 5:20result of managing this process. A
- 5:22regular company would have to assign a
- 5:24person to every one of these node tasks
- 5:26.
- 5:26Our company runs itself like an
- 5:28autonomous vehicle.
- 5:31So, at our company, there is no need for
- 5:34people to learn Comfy or use expensive
- 5:37AIs like GPT-6 Astra.
- 5:40It’s not just automation; we tested and
- 5:43refined the same scenes multiple times
- 5:45to produce actual results. A short-form
- 5:47factory that produces quality while
- 5:50automating. That is what makes us
- 5:52different from others. This method is
- 5:55not used only for senior storytelling
- 5:57videos.
- 5:58The shopping shorts and architecture
- 6:00shorts we made were created by the same
- 6:02AI company.
- 6:04This shopping short explained the
- 6:06history of masks through images and
- 6:08narration.
- 6:12It is a video assembled into a single
- 6:14story by adding movement to the screen
- 6:15and layering subtitles.
- 6:17The additional cost to create one
- 6:19shopping Short in this style was about
- 6:2130 won.
- 6:23This architectural video shows the role
- 6:25of earth retaining walls in the process
- 6:28of building an underground parking lot.
- 6:32Both cases differ in their production
- 6:34methods from the senior-focused video
- 6:35you just saw.
- 6:37The shopping video assembled images and
- 6:39audio, while the architectural video
- 6:41utilized scenes made with external video
- 6:43generation tools.
- 6:46Even if the materials and tools differ,
- 6:48the process of planning, creating scenes
- 6:50, reviewing, and refining continues.
- 6:52If you try using an AI video tool once,
- 6:55you get results quickly.
- 6:57But if you don't like it, you have to
- 6:59give new instructions, recreate it, and
- 7:01move results between multiple windows.
- 7:04As the number of videos increases, that
- 7:06work increases as well. We divided this
- 7:09process as if it were a company.
- 7:11The writer creates the story, the
- 7:13producer creates the scenes, and the
- 7:15reviewer checks the results.
- 7:17We made it so that everyone's results
- 7:20lead to the next task.
- 7:22Humans determine the direction and
- 7:24results, and AI employees record and
- 7:27connect those decisions to tasks based
- 7:29on that judgment.
- 7:31It becomes a standard for work that
- 7:32accumulates in the company, not just a
- 7:34one-time disposable answer.
- 7:36I view the way of using AI in four
- 7:38stages. The first is the stage of asking
- 7:40one AI a question.
- 7:42The second is the stage where humans
- 7:44directly connect multiple AIs.
- 7:47The third is the stage of creating an AI
- 7:49company with divided roles, and the
- 7:51fourth is the stage where the company
- 7:53checks its status while working and
- 7:55organizes what it has learned.
- 7:58Our AI company checks its operational
- 8:00status like a weekly health checkup and
- 8:02organizes unnecessarily long memories
- 8:05like going on a diet. We leave important
- 8:08knowledge and feedback so they can be
- 8:09used for the next task.
- 8:11If you only look at one result, the
- 8:13difference might seem small. However,
- 8:16the method of explaining from scratch
- 8:18every time versus building up standards
- 8:20and experience can result in a growing
- 8:21gap over time. This is precisely the
- 8:24structure I am building. But this
- 8:27structure didn't just end with
- 8:29automatically creating videos. It
- 8:31actually achieved high view counts as
- 8:34well. These are the view counts for my
- 8:36channel over the last 28 days. Among
- 8:37them, the videos with high short-form
- 8:40view counts are like this. A single
- 8:42short-form video generated about 1.8
- 8:44million views and an estimated revenue
- 8:47of about 1.21 million won.
- 8:49As you saw earlier, there are parts you
- 8:51can make for zero won, but that requires
- 8:53a certain level of computer hardware. So
- 8:56, this lecture contains the know-how to
- 8:58create short-form content at a cost of
- 9:00around 30 won, regardless of computer
- 9:02specifications.
- 9:04And I will show you the exact process of
- 9:06how AI employees handle research,
- 9:07creation, and even uploading without you
- 9:09having to click a single thing.
- 9:12The price of this lecture increases as
- 9:14time goes by.
- 9:15Check out the course details in the
- 9:17pinned comment now before the price goes
- 9:20up.
About this transcript
This page contains the full transcript of 힉스필드는 끝났다? 같은 장면 0원으로 다시 만드는 미친 로컬AI ㄷㄷ by 돈버는 온라인자판기, generated from the public captions YouTube serves with the video. The transcript has 1,505 words across 258 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.