This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
What I Built
Most outdoor AI apps demand your attention on a screen: tracking GPS routes, grinding streak counters, or squinting at maps. Wayside does the opposite. The screen is the shortest part of the experience; the walk is where the project actually happens.
Wayside turns a simple walk into an interactive specimen collection and illustrated storybook:
- Head outside and observe: You go on a walk and photograph small, overlooked things along the way—a scarlet maple leaf stuck in moss, an acorn with its cap askew, an abandoned candy wrapper, or a quiet neighborhood cat.
- Curate your specimens: When you return, Wayside removes the background clutter with an on-device cutout model and identifies each find with a vision-language model. Each item is mounted on a physical-style specimen card inside an interactive wooden shadowbox.
- Add your field notes: Jot down personal sensory memories (what it smelled like, where it was hiding, or what you were thinking).
- Weave the narrative: Choose which specimens to include, set an atmosphere (mystery, cozy, adventure, quiet, or spooky), and pick a location. A local language model weaves your finds and handwritten notes into a coherent short story.
- Illustrate the cover: A local diffusion model paints a custom 512×512 watercolor storybook cover based on your specimens in their setting, binding your walk into a digital keepsake.
Wayside is built for hikers, neighborhood wanderers, naturalists, and anyone who wants a reason to put their phone in their pocket, look at the trail, and turn real-world observations into imaginative fiction.
Demo
Video: https://youtu.be/2X3cnlWAxoQ
Deployment: https://wayside-46at.onrender.com/
(sometimes deployment timeouts :( trying my best to keep it up, couldn't get the render free credits)
Local Run: Follow the Quick Start steps below to run it on your own laptop in under two minutes.
Screenshots:
Code
Check it out here: Github Repository
How I Built It
Wayside is built entirely around four open-weight models running locally on everyday consumer hardware (tested on a standard Windows laptop with an Intel Core i7 CPU and no dedicated GPU). There are zero closed AI APIs, zero subscriptions, and zero paywalls.
The Open Model Stack
- Background Removal (rembg with u2netp): At just 4.4 MB, the lightweight u2netp ONNX model isolates leaves,flowers, animals and curiosities from sidewalk and grass backgrounds in under 3 seconds on CPU.
- Vision Labeling (HuggingFaceTB/SmolVLM-256M-Instruct): A compact 494 MB vision-language model that inspects the isolated find, names it, and writes an atmospheric one-line description without bloating system memory.
- Narrative Generation (Qwen/Qwen2.5-0.5B-Instruct): A 953 MB instruct model that weaves real find labels and human field notes into an imaginative story. It avoids generic lists, treating your items as clues and characters in a continuous narrative.
- Storybook Cover Illustration (stabilityai/sd-turbo via diffusers): Operates in a two-step sequence inside a single lock:
The Qwen story model distills the story and finds into a visual scene prompt of under 28 words.
SD-Turbo generates a rich 512×512 watercolor storybook cover in just 2 inference steps (~25s on CPU).
Frontend: Built with vanilla HTML, modern CSS variables (featuring six themes: Moss, Tide, Dusk, Blossom, Honey, and Night), and native JavaScript ES modules. Zero build steps, zero npm bundle bloat.
Why Does Open Innovation Matter?
Open innovation is what made Wayside possible:
- True Privacy for Real Walks: Your walk locations, outdoor memories, and personal field notes never leave your computer. Nothing is sent to commercial cloud AI vendors, data scrapers, or third-party servers.
- Runs on Everyday Consumer Hardware: You don’t need an expensive cloud GPU instance or high-end graphics card. By selecting high-efficiency open models (u2netp, SmolVLM-256M, Qwen2.5-0.5B, sd-turbo), the entire pipeline runs smoothly on standard laptop CPUs.
- Zero Marginal Cost: Walks should be free. With open models running on your own machine, there are no monthly token subscriptions, API rate limit paywalls, or per-photo charges.
- Full Transparency & Auditability: Every model checkpoint and line of code is open-source. Anyone can inspect how their data is processed, tune prompts, or swap model checkpoints.
Prize Categories
- Best Use of Render Track (deployment)
- Best Use of ElevenLabs (voiceover)
Top comments (1)
just made my first submission omg yay!