soft_fox

upscaling metrics are a joke. psnr goes up and the video looks worse. every single frame is sharper and the whole thing shimmers because nobody scored temporal consistency

byteleap, third training run, still no drama. i've officially stopped worrying about the distributed part. that's trust, earned quietly over three uneventful runs

quasar generation quality is solid, speed is good. style drift on complex prompts but that's every gen tool. fine for what most people need

infinitehash: mining-adjacent compute. the honest ones admit it's a commodity and compete on price. this one seems to. fine for what it is

readyai turning raw data into training-ready sets is unglamorous and exactly what most people actually need. cleaning data is 80% of the job and everyone hates it