Building YouTube Shorts with AI Myself, Then Realizing 'This is Better as a Paid Service'

I had a YouTube Shorts video I really wanted to replicate. I tried to **build it myself** using my **AI API keys** (Gemini/Vertex's **Veo 3.1**) to reach that same level. I started with images, fixed the start and end frames, reviewed frame sheets – I threw in every trick I knew and even spent actual money (₩3,350+). To cut to the chase – **I couldn't reach that benchmark level with Veo 3.1 no matter what, and eventually had to admit that "paid services (like Higgsfield) are more efficient for this."** This post will **fully disclose all the know-how** from that process – what I tried, where I got stuck, and why I concluded that paid is the answer. > I'm a solo developer building services in Korea. I only write about things I've **actually built and experienced**. (I haven't built anything with Higgsfield – this is a story about confirming the **ceiling of self-production** in front of a benchmark.) ## 1. The First Wall — Characters Keep Changing The biggest problem with AI video is **consistency**. When a cut changes or movement becomes significant, the protagonist's face and clothes subtly shift. I made a 4-cut video, and while running, it was fine, but the moment of **jumping** turned gray pants black and the goggles on the forehead into glowing eyes. The protagonist became "someone else" within a single video. The cause is clear – **when you generate video directly from text (text-to-video), the AI imagines a new character for each cut.** ## 2. The Method That Actually Worked — Not "Video First," But "Image First" Through trial and error, I found two key principles: - **① Create an image first, then animate that image.** By fixing the character with a **single master still (I used the Nano Banana image AI)**, and then animating that image (image-to-video), it becomes much more consistent. - **② Fix both the start and end frames.** If you only provide the starting image, the AI will improvise the middle and it can get messy. **By creating and fixing an ending image with the same character**, the AI knows the "start and end points" and only fills in between, leading to less variation. Applying these two principles, the clothes and goggles remained consistent for the entire 8 seconds. And importantly, **failures are caught at the image stage (₩80) — fixing them in the video (₩600+) is expensive.** ## 3. Rules Bought for ₩3,350 — Still Remaining Walls Completing one 4-cut, 32-second video cost a **total of ₩3,350** (5 stills + 4 video cuts). Cuts 3 and 4 were good, but cuts 1 and 2 had **morphing** (a brush disappearing in mid-air, hands teleporting) and **jump cuts**, so they were put on hold. The rules learned from this expense: - i2v motion must stay **within the "action radius" of the still pose** (e.g., a still with a hand on a desk + motion of pushing a window = physically incompatible → morphing). - **One action per clip**, and prohibit instructions like "ends with new element" (the model interprets this as a jump cut). - Add `no scene change / cut / transition` to the negative prompts. - **QC stills not just for aesthetics, but for logical action** (pose-motion compatibility, direction of movement). - **No text within the screen** (AI still writes text poorly – use subtitles in editing). - Review using **frame sheets (1-second intervals) instead of single frames** — morphing and jump cuts are only visible in sequence. - **Videos for minors are completely impossible** (Google AI blocks generation of underage content by policy, no workarounds). ## 4. Five Episodes Actually Rendered — Not Once, But Five Times It wasn't a case of "tried it once and it didn't work." To **empirically test** different hypotheses, I rendered **5 episodes**. Each tested something different, costing a total of **₩13,960**. | Episode | What Was Tested | Cost | Result | |---|---|---|---| | ① Past Exam POV (30s) | First completion — assembling 4 cuts from still→i2v | ₩3,350 | Cuts 3-4 OK / Cuts 1-2 morphing → On hold | | ② Anime Live-Action Stage (28s) | Mimicking a benchmark (anime fancam) + anchor frames | ₩4,000 | **Look achieved** (still below) | | ③ Seoul Street Interview (8s) | Veo's strength — single person lip-sync | ₩2,320 | Confirmed Veo's best area | | ④ Anime OP Rooftop Chase | One-take chase (continuous spectacle) | ₩2,400 | Limitations in long-take consistency | | ⑤ Snowy Mountain Sunrise (Emotional Cut Editing) | Close-up on single subject's emotion | ₩1,890 | **Frame consistency** (verify sheet below) | | **Total** | **5 structural experiments** | **₩13,960** | Mapping the tool's boundaries | **Stills and looks can be produced at a benchmark level.** For episode ② (mimicking an anime fancam), I repeated the master still 6 times to achieve this look – prompts for fire and water effects, magic circles underfoot, burning Taisho-era streets, and even audience phone fancams were followed:

▲ Master Still for Episode ② (Nano Banana). The look can be achieved to this extent — the problem is in the stage of connecting it into a '30-second moving video'.

And for verification, you **must look at frame sheets, not single frames**, to spot morphing and jump cuts. In episode ⑤ (Snowy Mountain Sunrise), since it's a single subject, the frames are consistent — Veo's **best area**:

▲ Frame verification sheet for Episode ⑤. Top = Climber's emotional close-up (frame consistency), Bottom = Snowy mountain sunrise aerial view. For single subjects and contained actions, Veo performs well.

**A pattern emerged** — single person/single action (③, ⑤) worked well, but as soon as I tried to connect multiple cuts like a benchmark (①, ④), it fell apart. A single still (₩80-640) is excellent, but the iterative cost of turning it into a "benchmark-level 30-second moving video" accumulated at ₩2,000-4,000 per episode — leading to the following conclusion. ## 5. In Front of the Benchmark — Admitting Veo 3.1's Ceiling By this point, you might think "AI video works," but looking again at the **YouTube Shorts I tried to replicate**, the gap was clear. That benchmark was a **8-12 second continuous spectacle of connected shots** (a seamless action one-take), but Veo 3.1's method of **stitching together 8-second clips** couldn't achieve that "continuous flow of a single breath." Even by matching start and end frames, it was **an approximation, not a perfect match**. That "progressive action one-take" is the domain of models like **Seedance 2.0** or **Gemini Omni**, which are the engines used by **paid services like Higgsfield**. My tool (Veo 3.1) can approximate it, but **if the goal is that level, paid services are overwhelmingly more efficient**. (For reference, Seedance 2.0 is only accessible via Dreamina/fal, and OpenAI has exited video – within my 3 keys, Veo is virtually the only option for video.) ## 6. What I Gained — And Why I Have No Regrets By investing time and money, I learned the **realistic ceiling of self-production**. My mindset was this: > **Try it first. If the best (free self-production) doesn't work, use the next best (paid service) first. However, technology is constantly evolving, so I'll track the model landscape, document it, and try again later.** This isn't a failure, but **having paid for the realistic limitations**. Now I know: - Character consistency is achieved by **image first + start/end frames** (this principle is valid wherever you produce). - **8-second quality** can be achieved with my current tools. - However, **continuous spectacle in one breath** requires Gemini Omni/Seedance 2.0 level, and if that's the goal, **paid is the answer**. - The video AI model landscape **changes monthly** — today's ceiling could be tomorrow's basic feature, so I'll keep tracking it. What I gained for ₩3,350 and a few days is not an illusion ("AI can make anything instantly"), but **knowing the actual location of the boundary line**. In the next post, I'll share **how the "Riel Chatbot" has grown through all these trials and errors.**