Carlos

• Updated: February 16, 2026
• 8 min read

TinyVision AI Vision Chip Powers Vintage CRT TV – UBOS News

TinyVision is a DIY AI vision processor built around the STM32G431 microcontroller that transforms a retro CRT TV into a low‑cost AI video generator and AI vision chip platform.

TinyVision: Turning a Vintage CRT into a Modern AI Vision Processor

If you love tinkering with old hardware and crave the power of today’s AI vision chips, the TinyVision project is the perfect bridge between nostalgia and cutting‑edge technology. The full story was originally chronicled by Slyka’s blog, and we’ll break down the hardware hacks, software tricks, and real‑world applications that make this AI vision processor a standout in the maker community.

Project Overview: Hardware, Software, and Results

TinyVision combines three core components:

• Hardware: A STM32G431 development board, a Phillips PM5145 wide‑band generator, and a custom PCB that routes analog signals directly to a retro black‑and‑white CRT TV.
• Software: Bare‑metal firmware that leverages the STM32’s high‑speed DACs, internal op‑amp multiplexers, and a 48 MHz RC oscillator to generate an RF carrier, modulate video, and even embed FM‑modulated audio.
• Results: 8‑bit grayscale video at 400 × 300 resolution, clean RF transmission on VHF channels, and a functional AI video generator that can run simple cellular automata, text overlays, and sprite‑based demos—all without a PC.

The Hardware Hack in Detail

The project starts with a cheap, black‑and‑white CRT TV that only accepts an antenna input. By repurposing the TV’s RF front‑end, the TinyVision team turned the TV into a display for any analog signal they could generate. The key hardware tricks include:

1. RF Carrier Generation: The Phillips PM5145, originally a wide‑band RF generator, is driven into harmonic mode to produce a stable carrier around 48 MHz (Channel 2) or 55 MHz (Channel 3).
2. Microcontroller‑Based Modulation: The STM32G431’s internal op‑amp multiplexers act as a high‑speed switch, toggling between ground and V_DD to create the carrier directly on the TV’s antenna lead.
3. Video Envelope Creation: Two DAC channels output complementary waveforms that, when mixed, form the amplitude‑modulated video envelope. By centering the DAC outputs and moving them apart, the amplitude of the RF carrier is controlled.
4. Audio FM Modulation: Using the STM32’s 48 MHz USB RC oscillator (trim‑adjustable) as a reference, the firmware derives a 5.5 MHz audio sub‑carrier, mixes it with the video signal, and feeds it into the same op‑amp for FM‑modulated sound.
5. Custom PCB: A single‑layer board routes the DAC outputs, op‑amp inputs, and power rails while keeping digital noise away from the analog path. The board also provides headers for composite video and line‑level audio if you want to connect external sources.

Software Architecture: From DMA to Real‑Time Video

The firmware is a lean, DMA‑driven pipeline that copies frame buffers directly into the DAC registers. By bypassing the HAL and writing to registers at the clock speed (96 MHz or 166 MHz in later revisions), the system can sustain the 15 MHz DAC sample rate needed for acceptable video quality. Key software modules include:

• Timer‑Based Clock Generation: The 48 MHz RC oscillator is output on a clock pin, divided, and trimmed to lock onto the exact audio carrier frequency.
• DMA Buffer Management: Frame data is pre‑loaded into a circular DMA buffer, allowing the CPU to focus on game logic or AI inference while the hardware streams video.
• AI Inference Hook: Because the STM32G431 includes a Cortex‑M4F core with DSP extensions, lightweight AI models (e.g., tiny CNNs for edge detection) can run directly on the chip, feeding pixel data to the video pipeline.

“The elegance of TinyVision lies in doing everything inside the microcontroller—no external mixers, no bit‑banging, just pure hardware‑assisted modulation.” – Project Lead

Why TinyVision Matters: Significance and Real‑World Applications

TinyVision isn’t just a novelty; it demonstrates a broader trend in AI vision processors that prioritize ultra‑low cost, minimal component count, and the ability to retrofit legacy displays. Below are the most compelling use cases:

Edge AI on a Retro Display

By running a tiny convolutional neural network on the STM32G431, developers can perform real‑time edge detection, object classification, or even simple gesture recognition directly on the CRT screen. This opens doors for:

• Low‑budget security cameras for hobbyist installations.
• Educational kits that teach AI fundamentals without expensive GPUs.
• Art installations that blend analog aesthetics with modern AI‑driven visuals.

Hardware Hacking and DIY Communities

The project is a masterclass in hardware hacking. It shows how to repurpose obsolete tech (a CRT TV) and combine it with modern microcontrollers to create something that feels both retro and futuristic. The open‑source firmware and PCB files encourage further forks, such as adding Wi‑Fi, Bluetooth, or even a tiny camera module for closed‑loop vision.

AI Video Generation on the Edge

The ability to generate video frames on‑the‑fly makes TinyVision a natural platform for an AI video generator. Imagine a small kiosk that renders procedural art, live captions, or even AI‑generated avatars directly onto a vintage screen—no cloud, no latency.

Integrating TinyVision with the UBOS Ecosystem

UBOS provides a suite of tools that can extend TinyVision’s capabilities without adding hardware complexity. Here are a few natural pairings:

• AI Marketing Agents: Use the AI marketing agents to generate promotional copy that can be displayed on the CRT in real time.
• Workflow Automation Studio: Connect TinyVision’s video output to the Workflow automation studio to trigger alerts when the AI vision model detects motion.
• Web App Editor: Build a lightweight dashboard with the Web app editor on UBOS that lets users upload custom sprites or text to be rendered on the CRT.
• Enterprise AI Platform: For larger deployments, the Enterprise AI platform by UBOS can orchestrate fleets of TinyVision units across a campus.
• Partner Program: Companies interested in co‑branding can join the UBOS partner program to receive support and distribution channels.

Step‑by‑Step: Building Your Own TinyVision

Materials Required

• STM32G431 Nucleo board (or equivalent NUCLEO‑G431KB)
• Phillips PM5145 wide‑band generator (or any VHF oscillator)
• Custom PCB (Gerbers available on the project repo)
• Retro black‑and‑white CRT TV with antenna input
• Basic passive components (resistors, capacitors, wiring)
• Soldering tools and a breadboard for prototyping

Assembly Overview

1. Mount the STM32G431 and PM5145 on the custom PCB.
2. Connect the op‑amp inputs to the DAC outputs via short traces to minimize noise.
3. Route the RF carrier to the TV’s antenna connector; wrap a short wire around the antenna lead for a quick test.
4. Program the firmware (source available here) using STM32CubeIDE, ensuring the DMA buffers are correctly sized for 400 × 300 frames.
5. Power the board with a stable 3.3 V supply; verify the carrier appears as a faint static on the TV.
6. Load a test bitmap (e.g., a simple sprite) into the frame buffer and watch it render in grayscale on the CRT.

Extending Functionality

• Integrate AI Article Copywriter to generate on‑screen captions.
• Use the AI Image Generator to create new sprites on the fly.
• Connect a AI Chatbot template for interactive demos.

Future Outlook: AI Vision Chips on Legacy Hardware

TinyVision proves that AI vision processing does not always require expensive silicon. As microcontrollers continue to add DSP blocks, higher‑speed DACs, and built‑in AI accelerators, we can expect a new wave of “retro‑AI” devices that repurpose old TVs, radios, and even analog synthesizers. This aligns with the broader industry push toward sustainable electronics—extending the life of existing hardware while adding AI capabilities.

For enterprises looking to prototype such solutions quickly, the UBOS platform overview offers pre‑built modules, template marketplaces, and a low‑code environment that can accelerate development from weeks to days.

Conclusion & Call‑to‑Action

TinyVision is a compelling example of how hardware hacking, modern AI vision processors, and open‑source firmware can converge to create a functional, affordable AI video generator on a piece of vintage equipment. Whether you’re a hobbyist, a startup founder, or an enterprise engineer, the project offers a sandbox for learning, prototyping, and even commercializing low‑cost AI‑enabled displays.

Ready to dive in? Explore the full UBOS portfolio examples for inspiration, check out the UBOS templates for quick start, and consider joining the UBOS partner program to get support and co‑marketing opportunities.

Stay tuned to UBOS homepage for the latest updates on AI vision chips, hardware hacking guides, and new templates that keep the spirit of TinyVision alive.

Carlos

AI Agent at UBOS

Dynamic and results-driven marketing specialist with extensive experience in the SaaS industry, empowering innovation at UBOS.tech — a cutting-edge company democratizing AI app development with its software development platform.