Building a Bulk AI Video Generator from Scratch | Path to 100 Paying Users

Hey everyone,

I've never committed myself to finishing a product until it's marketable or makes me any money. So I feel like maybe it's time I put myself out there.

I’m currently working full-time in management, but my goal is to scale my custom desktop AI video generator to 100 recurring subscribers so I can build a 12-month savings runway and transition into full-time software development. Right now it's only good enough for my personal use but working towards monetizing it.

What I'm Building​

A lightweight desktop app designed for batch short-form video creation (TikTok/Reels/Shorts). Instead of spending hours manually editing or paying expensive cloud rendering fees, the app lets users prompt a topic, auto-generates scripts and 6 scroll-stopping hooks, fetches AI visuals, and stitches everything locally with dynamic subtitles and watermarks.

Current Tech Stack:

• Python
• Gradio UI
• FFmpeg (local rendering)
• Gemini 3.5 Flash and 3.1 Flash Lite (for AI script generation)
• Fal.ai (for AI image generation using Flux).

How the Software Evolved (v1 to v5 Log summary)​

• v1 (Core MVP): Built the base script—takes a text topic, generates the script + AI images, and stitches them into a video using local FFmpeg.
• v2 (Formatting & Subtitles): Added customizable prompt parameters and dynamic auto-subtitles. Includes background audio option.
• v3 (Branding & Retention): Added Hormozi-style word highlighting, custom logo overlays, and watermark protection. Removes background audio option because it's probably better to use TikTok audio.
• v4 (Marketing Engine): Integrated social caption generation and built an AI feature that outputs 6 distinct video hooks per script.
• v5 (Current Build): Added the Ken Burns pan/zoom motion effect to give static images a cinematic feel.

Real-World Testing: Real Estate TikTok​

I’ve started running a real estate niche TikTok account using videos generated 100% by this tool.

• Goal: Test production efficiency and hook retention in the wild.
• Current Status: I’m able to generate 6 videos in under 7 minutes (most of the time is basically the video generation). Views are currently sitting at 100–300 per clip while I refine the script prompts, but the key is that zero manual editing time was spent on my end. I was able to spend a few minutes prompting 3 ideas and having 18 video outputs that I've scheduled to post on my account 3 times a day.
• Analytics: I'll be sharing analytics updates as the account grows.

Roadmap & What's Next​

1. Version Plans: Refining the FFmpeg rendering pipeline so the Ken Burns effect runs smoothly across all CPU/GPU setups (currently the output looks literally shaky). Probably speed up video generation as well. The subtitle highlight also doesn't match the voiceover, but that's probably the least of my worries for now.
2. Backend Proxy & Bundled API: Setting up a FastAPI backend server to handle license validation and bundle $5/mo of AI generation tokens into the subscription. I've never gotten this far into building anything so this would probably be the biggest mental hurdle for me.
3. Packaging: Compiling the script into a clean .exe so non-tech users don't need Python installed.
4. Beta Testing: Recruiting a few Windows 10/11 users to stress-test the installer on different hardware specs.

Maybe if I post this I would be more inclined to get out of my comfort zone and push myself to at least try. Either I fail or I don't, at least I could say that I genuinely tried.

Question for the community: For those running faceless TikTok or Shorts channels, which feature matters most to you—fast batch rendering, custom watermarks, or AI hook generation?

Would love to hear your thoughts on this.

The idea is very good bro, but i'd change gemini for deepseek v4, flash or pro. It's many times cheaper, and multiples time more inteligent. I'm building something like this myself, and a breakthrough i had is that you always gotta look for chinese ais, because almost always there's a smarter one that's cheaper. Hope you have success in your jorney!

I'm curious how people are handling the local rendering step — most similar tools I've seen just offload that to a cloud API and eat the cost per render. Doing it through FFmpeg locally is the right call long-term for margins.

That's a solid amount of work you've put into it. Looking forward to seeing future updates.

I would say, any kind of A/B testing features would be very useful. I'd be interested, please DM me when it's ready.

itsworth said:

The idea is very good bro, but i'd change gemini for deepseek v4, flash or pro. It's many times cheaper, and multiples time more inteligent. I'm building something like this myself, and a breakthrough i had is that you always gotta look for chinese ais, because almost always there's a smarter one that's cheaper. Hope you have success in your jorney!

Click to expand...

Thanks for the tip on Chinese AI APIs. DeepSeek/Qwen for script generation is definitely on my radar once I deploy the backend proxy to keep API costs near zero. I have been experimenting a bit with Gemini first since they gave a generous free tier for me to test out personally at least. Will try and check out the DeepSeek/Qwen to see whether their script generation is engaging enough or not.

Appreciate your support on this!

nightJoe said:

I'm curious how people are handling the local rendering step — most similar tools I've seen just offload that to a cloud API and eat the cost per render. Doing it through FFmpeg locally is the right call long-term for margins.

Click to expand...

I hope it goes well though, but it probably depends a lot on user's spec as well. I have checked the option to offload to cloud API as you said, and yea the cost per render is absolutely mind staggering for me.

Thanks for the support!

Adult-HTML said:

That's a solid amount of work you've put into it. Looking forward to seeing future updates.

Click to expand...

Thanks man, I appreciate the support. Hopefully I can get some new updates by either this weekend or early next week.

Gnomical said:

I would say, any kind of A/B testing features would be very useful. I'd be interested, please DM me when it's ready.

Click to expand...

Thats great to hear, thanks for the support! I'll drop you a DM later so we can stay connected when the beta testing is ready.

Update 1: 3/8/2026

Quick Update & Responses to Everyone!

First off, huge thanks to everyone who stopped by and commented. The feedback has been really encouraging!

️ Tech & Build Update (v5.0.1 - The Good, The Bad, & The Regression)

Hitting my first real wall with optimization, but I guess that's part of the process.

1. Ken Burns Motion: Still working on it. The pan/zoom effect is still outputting shaky video.
2. Prompt / Image Count Bug: Found a weird bug in v5.0.1 where setting the front end to 3 images per scene leads to dynamic unpredictability. Normally it would default to output 3 images per scene. But it's not doing that. I can reduce it to 2 or 4 images, but not 3. This bug happens when I'm using Gemini 3.5. When I switch to using a different Gemini model (3.5 flash lite), the 3rd image changes into a text overlay.
3. Performance Spike (Major Issue): Generation time for a batch of 6 videos spiked from 7 minutes all the way up to 20 minutes! Going from ~1 min/video to over 3 mins/video is unacceptable for a local batch tool.
4. Action Plan: I'm likely going to regress back to v5.0.0 (where rendering was stable and fast at 7 mins) and re-integrate changes modularly so I can isolate what broke the pipeline.

---

Real Estate TikTok Test Update

I've hit 18 total videos posted over the past 6 days (3x per day), 100% generated with zero manual editing time. But I did had to spend some time to:

1. Schedule the videos.
2. Add TikTok audio.
3. Set cover photo.
4. Paste the caption and hashtags.
5. Tick the AI disclaimer.

This did take some time, but it's probably an automation for another time.

The goal here isn't viral numbers right out of the gate, but to test the workflow velocity and baseline retention. Seeing steady baseline impressions across all uploads proves the automated formatting, subtitles, and hooks pass TikTok's quality checks without getting flagged as low-effort spam.

I've attached some screenshot analytics below showing the current overview for the last 7 days against the previous 7 days before that. My videos went from getting 200 - 300 views per video (when I posted once a day) to getting 100+ views per video (when I posted 3 videos per day). Obviously seems like the views are divided per day. Haven't been able to breach to 400 views on a video yet, but I'll try and adjust the scripts and hooks.

Next focus: Fix the FFmpeg speed regression, roll back to v5.0.0 baseline, and fix the prompt logic.

More updates coming soon!