𝗜 𝗔𝗜-𝗥𝗲𝗺𝗮𝘀𝘁𝗲𝗿𝗲𝗱 𝗮 𝟮𝟱-𝗬𝗲𝗮𝗿-𝗢𝗹𝗱 𝗚𝗮𝗺𝗲 𝗜𝗻𝘁𝗿𝗼
I spent a long time remastering a game intro from the year 2000. I used AI to turn it into clean 1080p video.
I learned that the source file matters more than the AI model.
Here are my main lessons from the project:
Pick the best source. I had a 360p version with good audio and a 1080p version with bad audio. I chose the 1080p version for the video and added the good audio back later. Always feed your upscaler the highest quality file you can find.
Use small clips to test. Do not run your full video first. I spent 16 hours testing a single 5-second clip. This helped me find the right settings without wasting days of compute time.
Control the noise. Fast motion causes flickering. I used a low latent_noise scale for faces to keep detail. I used a higher scale for fast action to stop the AI from hallucinating new details every frame.
Split by shots, not time. Use tools like PySceneDetect to find cuts. Do not use time-based splits. Time-based splits cause audio sync issues. Use frame-exact cuts to keep your audio and video aligned.
Use the cloud for big jobs. My local hardware was slow. I rented a professional GPU on vast.ai for about $2.70. It turned a 74-hour job into a 2-hour job.
Final upload trick. Upscale your video to 4K before uploading to YouTube. YouTube gives higher bitrates to 4K videos. This keeps your 1080p video looking sharp.
The full code and pipeline are available on GitHub.
Source: https://github.com/andyskw/ig2-solarian-seedvr2-remaster
Optional learning community: https://t.me/GyaanSetuAi