@zarazhangrui Love seeing innovation in this space! 🙌 We're tackling a similar problem with LearnLens ( https://t.co/prJu5qbAPb) — a Chrome extension that takes a different approach: ✅ Lives directly in YouTube (side-by-side panel, zero friction) ✅ Knowledge checks that test retention as

Not all hours behind a screen are learning hours💻 Some build knowledge, others slip into distraction. TimeBack makes the difference crystal clear! Fix your anti-patterns, and let every minute count⏳ Been cooking this for weeks—can’t wait to ship 🚀

Know any Gemini video hacks I missed? 🙋 Drop them below 📝👇🏻

Rule of Thumb 📌 1️⃣ Just one video, per prompt. Don’t pass multiple videos at once for best results. (Even though the limit is 10) 2️⃣ Place video before the prompt. Gemini works better when the video is the first element in the prompt, before any questions or instructions.

Flash caveats 🚨 ⚖️ It’s not as instant as old Flash (because it now "thinks"). Turn off the 'thinking' mode if speed is a priority. 🏆 Pro still outperforms for complex reasoning. Consider trying both. If Flash performs well enough, switch. You might not notice the difference,

Still using the 'pro' model for everything? Think again! ⚡ Gemini Flash was once the "fast but shallow" model, but no more! Since Gemini 2.5, 'Flash' has also gained a "reasoning" mode! 🧠 It often delivers comparable results to pro at a fraction of the cost. 🚀 🧩 Reasoning

Want to query specific parts of the video? 📐 ❌ Don’t manually trim, no ffmpeg overhead. ✅ Just use the 'start_offset' and 'end_offset' params. Upload once via Files API, then reference the clip with offsets. No need to re-upload segments — just point to different time ranges

Re-querying the same video or parts of the same video? ♻️ 🚫 Don’t use the inline video passing method. 📁 Use the Files API. Upload once, reuse in all queries. Why this is better? • 🔁 No re-upload: Files are kept up to 48 h, saving subsequent upload times. • ⚡ Faster:

Struggling with fast motion clips? 🏃🏻 Most people don’t know Gemini samples all your videos at 1 FPS by default. So the action in between frames (1 second apart) is lost. 😱 What should we do? Earlier, slowing down or speeding up the video was the way to go. ⏪⏩ But now

Dealing with timestamps? ⏲️ Use the "MM:SS" format ✅ (Like '1:28' instead of '88s') This keeps it consistent with what Gemini internally prefers, hence: • 📍 Improved timestamp accuracy • 🐛 Reduced parsing errors

Analyzing long videos? 📹 Chunk them ✂️ One call per chunk. Why chunk? • 🎯 Smaller context window → less hallucination • ⌚ Avoids the slight drift in timestamps observed beyond 1 hour • 💰 Keeps you in the cheaper Gemini Pro pricing tier (<=200k tokens) Bonus: Overlap

Using Gemini for video analysis?🎥 Want faster, cheaper, better results?✨ Here are 8 pro tips you probably didn't know 🧵👇🏻

This has many use cases: - User activity monitoring - Automated UI Testing - Video content analysis - Attention tracking - Security Systems Has compression ever tripped you up? Let’s talk 👇

The system is now precise and intelligent! Key Learnings: 🔍Always visualize — numbers alone can lie 🫧Video compression in common formats creates invisible differences

But wait—what about non-user changes? Clocks, blinking cursors, loading animations... They do create differences, but they’re not user input Solution: Now that we have clean diffs, we can feed them to vision models 🧠 Let them decide the type of change.

After a bit of threshold tuning: BOOM! 💥 Clean difference detection. Compression artifacts? Gone✅ Real changes? Captured✅ Magic? Nah — just better math.

I needed a better way to detect real activity — not compression noise. New plan: Take absolute pixel difference b/w frames → then denoise 🥅 Step1️⃣Filter out small differences below a threshold. Result -> Weird leftover dots. Step 2⃣ Filter out tiny regions by area. (those

The culprit? 🎭Compression artifacts Most common video formats like .mp4, .mov, and .webm use lossy compression. They don’t preserve exact pixels, just “good enough” approximations. So even if nothing visibly changes, the frames (pixels) aren't identical.

So I plotted the SSIM difference maps. Boom. Weird patches showed up. Pixel changes where there shouldn't be any.

Sometimes the SSIM score was lower even when nothing visibly changed. Lower than when something actually changed ! 🤯 This made zero sense ! What kind of visual sorcery was I dealing with?