MediaMend：本機 AI 媒體清理助理

Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady 🦥

T*This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady

What I Built

MediaMend is a local AI assistant for people who record far more media than they will ever review: blurry photos, black video frames, silent recordings, empty files, duplicates.

The lazy way to clean that up is "delete everything below a threshold", and that is also the way to lose the one blurry, unrepeatable shot that mattered. MediaMend takes a safer lazy path:

1. Scan a folder. Everything runs on your computer. Nothing is uploaded.
2. Review a short queue of flagged files, each with a reason and a confidence level.
3. Move to Bin. Approved files leave your folders but are not deleted.
4. Restore, Archive or Delete permanently from the Bin. Permanent deletion needs a separate confirmation.

You never inspect the whole folder. You only judge the cases the tool is unsure about.

Code

https://github.com/noorulhuda07/mediamend

How I Built It

• Deterministic checks for measurable problems: OpenCV Laplacian variance for blur, resolution checks, sampled video frames for blank-frame share, FFmpeg decoding for silence measured relative to each recording's own level, SHA-256 for exact duplicates.
• A local open-weight vision-language model for judgment: [model name and tag, e.g. via Ollama] looks at flagged photos and answers one narrow question: could this imperfect image still be worth keeping? It returns structured JSON that I validate. If the model is unavailable or answers badly, the app says so instead of faking a result.
• A safety layer: every move records the original path, size and SHA-256 in a manifest and verifies the checksum afterwards. Restore never overwrites an existing file, and nothing outside the scanned folder can be touched.
• Stack: Python, Streamlit, OpenCV, FFmpeg, [Ollama + model].

Technical measurements and AI opinions are stored separately, so a model's guess never overwrites measured evidence.

Results

I tested on [N] labelled files ([N] photos, [N] videos, [N] audio, [N] other). Ground truth was labelled before scanning.

Where it failed: [real examples, e.g. a dark but valid photo flagged as blurry because low contrast lowers the sharpness score; the model's wrong call on ...]. The thresholds are starting points I tuned on this set, not universal values.

Why Open Innovation Matters

• Privacy: personal photos and voice recordings never leave the machine.
• Control: I can swap or compare models ([model A] vs [model B]) and inspect the exact prompt and outputs.
• Cost: no per-image API fees for scanning thousands of files.
• Transparency and reproducibility: the code, prompt, thresholds and test set are public, so anyone can rerun the benchmark and see the failures.

The honest trade-off: [hardware needs, model speed on your laptop, accuracy limits].

Limitations

AI cannot judge creative value perfectly. Deleting from the Bin may not securely erase every copy on SSDs, backups or synced folders. [Anything else you observed.]

devchallenge #hf26challengehis is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady*

What I Built

Demo

Code

How I Built It

Why Does Open Innovation Matter?

My Agent Session

The Lazy Payoff

Prize Categories