Video summary
I made an AI YouTube channel to see if it's "ez money"
Main summary
Key takeaways
Technological concepts & workflow (main focus)
- Goal for the channel: create self-improvement YouTube videos where every visual is AI-generated, but in a style intended to look hand-made enough that at least one viewer believes it’s not AI.
- Visual style decision: avoid the “hyperrealistic/fake” look; use a static, Microsoft Paint–style aesthetic (simple illustrations, redrawn every ~3 seconds).
- Core planned “pipeline” (as advised by ChatGPT):
- Generate scene ideas
- Generate simple illustrations
- Convert images into Microsoft Paint-like outputs
Image generation tooling (and problems encountered)
Stable Diffusion direction (installation confusion + GPU issues)
- ChatGPT recommendation: use Stable Diffusion.
- Main blocker: installing and configuring the system quickly became confusing (engine vs model).
- Local approach: used Stable Diffusion Web UI (Automatic1111) via a looked-up GitHub setup, targeting an AMD GPU path.
- Repeated issues:
- Torch not able to use GPU
- Wrong DirectML build for AMD
- Torch installation failures tied to Python version mismatch
Free/hosted fallback: Amuse AI
- Moved to Amuse AI when Stable Diffusion installation and costs were unclear.
- Learned about:
- Models
- LoRAs
LoRA/style training concept
- LoRAs were understood as style/subject adaptations.
- Training guidance: typically needs ~20–100 images for a custom LoRA.
- But: platform limitations prevented direct training inside Amuse.
Switching to pre-made style LoRAs
- Used “Microsoft Paint Portraits” (safe tensors model) and other LoRAs.
- Major issue: outputs didn’t match Microsoft Paint styling closely; prompts produced inconsistent style.
Final breakthrough: Google Whisk
- Discovered Whisk can copy a style nearly 1:1 from an uploaded style reference/design.
- Workflow used:
- Created a Microsoft Paint–style reference in MS Paint
- Uploaded both content + style to Google Whisk
- Result: much closer pixel-style results
- Drawback: anti-aliasing/blurry edges.
- Tried Photoshop fixes, but concluded most viewers won’t notice.
- Kept edge artifacts.
Automation additions for speed
“Auto Whisk” style automation
- Used a Chrome extension / Auto Whisk approach:
- Generated many images automatically from prompts (using a TXT list)
- Downloaded results in bulk
- Impact: reduced repeated prompting, but editing still required lots of manual correction.
Script and scene description generation
- ChatGPT-generated scripts were “not good,” so the creator heavily edited them.
- The edited script was then used to generate sentence-by-sentence scene descriptions, enabling image generation per segment.
Voiceover generation
- Used ElevenLabs with a cloned voice option (Scottish accent).
- Avoided full custom voice cloning later due to complexity.
- Attempted free alternatives (e.g., RVC) but quality was poor → returned to ElevenLabs.
Audio post-processing automation
- Wrote/used a Python script (via ChatGPT) to remove silence:
- Trimmed pauses
- Reduced editing time
Editing pipeline and remaining bottlenecks
Video assembly in Premiere
- Images were placed based on voiceover timing.
- Even with automation, images sometimes failed, requiring:
- Regeneration
- Masking
- Failure examples included:
- dismorphed faces
- floating objects
- other generation errors
Face replacement attempt (major time sink)
- Attempted automation with Python to replace AI-generated faces with a preferred expression.
- Result: inconsistent placement, causing issues such as:
- faces on wrong regions
- double mouths
- faces in black areas
- After days and credit limits, the full automation was abandoned.
Scaling attempt with AI “agentic workflows” (Claude Code)
- Found a tutorial about building “Claude skills” and agentic workflows for more consistent AI output.
- Implemented a “grand slam” automation concept:
Input topic →
- generate script
- review/approval
- generate scene descriptions
- create multiple image options per scene
- choose best
- generate voiceover
- remove silences
- assemble video
- Practical result:
- Heavy learning curve
- High usage/credit consumption
- Also tried:
- Gemini image generation via API
- But: Gemini couldn’t generate images on the free tier, and API setup used most usage.
Image library/index idea
- Built a reusable library of previously generated images with manual descriptions.
- Added an AI-generated index to support future workflow reuse.
Results / reviews / analysis provided (outcomes)
- Channel start: “Paint Wizard” began near zero views and with no thumbnail upload (possibly related to an unverified channel limit).
- Performance improvements:
- Second video: “The Psychology of Disappearing to Come Back Unrecognizable”
- ~56 subscribers
- ~570 views
- Meaningful comments praising audio/visuals and reducing AI suspicion
- Third video: “How to Win If You Have High IQ”
- Initially dominated: ~400 views in 3 days
- ~59 subscribers
- Crossed 1,000 views later
- Second video: “The Psychology of Disappearing to Come Back Unrecognizable”
- Key success metric stated: achieved the goal of getting at least one viewer to believe it wasn’t AI, supported by comments.
Product/platform sponsor (mentioned)
Akiflow (calendar/time blocking + AI assist)
- Features described:
- Time-blocking
- Integrations (e.g., Slack/email)
- Command bar (universal search)
- Optional AI that learns from project tasks and assigns new tasks
- Claimed privacy/safety:
- Data used only to train a private model
Main speakers / sources (as referenced)
- Primary narrator: the YouTube creator (narrates their own experiments; references “Ryan” in an example, but narration is by the channel creator)
- Tools/AI systems consulted:
- ChatGPT
- Google Gemini
- Google Whisk
- Stable Diffusion (plus Automatic1111)
- Amuse AI
- Claude (Claude Code)
- ElevenLabs
- Akiflow (sponsor)
- External tutorial/source mentioned:
- FE Engineer (AMD fix tutorial for Automatic1111)