This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass Touch Grass Tales turns a short walk into an illustrated children's story.
You step outside, photograph 5 to 10 random things (a park bench, a pigeon, a weird sign, a lamp post), and an open-weight model, Gemma, turns them into a story, a rhyming poem or a comic strip. The pictures in the finished tale are your own photos. You pick the mood: funny, spooky (never gory) or mystery.
The screen is the shortest part of the experience on purpose:
Who it's for:
Live app: https://touch-grass-flutonp.onrender.com/ No API key? The app has a built-in demo (a canned walk with a park bench, a lamp post and a pigeon) and a gallery of sample tales.
How a tale gets made, step by step:
Go outside, photograph 5 to 10 random things, and Gemma AI turns them into a children's story, a rhyming poem or a comic strip (funny, spooky or mystery), illustrated with your own photos. Everything is written in simple words: short sentences, no idioms, about CEFR A2, so young readers and beginner English learners can follow it.
/setup.html walks through both.
The open-source AI core: Gemma, an open-weight model. Every word in a tale comes from Gemma. I use two Gemma 4 models, each for the job it's best at:
| Task | Model | Why |
|---|---|---|
| Look at each photo | Gemma 4 26B-A4B | Fast mixture-of-experts model; one small (384px) photo at a time |
| Plan the plot and pick a photo for each moment | Gemma 4 26B-A4B | Fast, structured JSON planning | | Write each scene or verse, then proofread and fix it | Gemma 4 31B | The best prose and rhyme | | Write comic panel captions and speech bubbles | Gemma 4 26B-A4B | Fast; a caption plus up to two bubbles per panel |
Many small calls, never one long reply. Instead of asking for a whole story at once, the app builds it piece by piece:
Small calls keep the model focused, make each reply cheap to retry, and let the page show progress the whole time.
The rest of the stack:
1. The same open model runs in the cloud or on your own machine.
Because Gemma's weights are open, the app has two interchangeable back ends: Google AI Studio (free key, nothing to install) and LM Studio, which runs Gemma locally. Switching is one click in the app, with no code changes. With LM Studio, your photos never leave your computer: the model that looks at your walk runs on your own hardware. A closed, API-only model can't offer that choice.
2. Your walk stays yours.
Photos of a walk can reveal where you live. So:
3. I pick the right model for each job.
Open weights come in many sizes, so the app uses a fast mixture-of-experts Gemma for looking and planning, and the large dense Gemma for writing and editing. In LM Studio you can swap in any Gemma build you've downloaded, per task, from a dropdown.
4. It costs nothing to run.
Gemma is free to use through a free AI Studio key, and completely free (and limitless) when run locally. A family can make a tale every weekend without a subscription.
Where open worked better than closed: <!-- TODO: in one or two sentences, add a concrete moment from building this, e.g. swapping models in LM Studio, testing locally for free, or keeping photos on your machine -->
Best Use of Gemma: every tale is written, planned and proofread by Gemma 4 (26B-A4B and 31B), and the photos are read by Gemma too. LM Studio runs Gemma locally.
Best Use of MongoDB Atlas: every tale's pipeline state lives in MongoDB Atlas, which is what lets a failed step resume exactly where it stopped. We don't save user generated tales.
Best Use of Render: the app is deployed as a single Render web service (render.yaml in the repo).