๐ฅ FREE 0-to-hero AI tutorials by PhD Computer Engineer!
Master Qwen Image/Wan 2.2 training (LoRA+fine-tune), ACE-Step music gen (beats Suno), pro upscaling crushing Topaz (SeedVR2+FlashVSR+), ComfyUI & 400+ step-by-step guides.
Subscribe for latest ๐
https://t.co/9hihbqgSUd
Testing different precision types speed of Krea 2 Turbo with 2048x2048px and 8 steps on RTX 5090 - each model generated twice to warmup and get best speed
Most people doesn't know but proper way of quantizing models is far from instant.
You have to use Calibration Samples, Optimizer (e.g. Prodigy), Iterations, Top P, Min K, Max K.
Currently quantizing Krea 2 into FP8 Tensor and Block Scaled + INT8 Block Scaled to test.
I also have modified ComfyUI-QuantOps custom node so INT8 and Block based scaling works out of box without changing anything automatically including SwarmUI.
This is proper way of quantizing models with prodigy optimizer and up to 2000 steps.
It takes massive time but generates the very best highest quality quants.
Currently I am generating Krea 2 Turbo FP8 Tensor Wise, Block Wise and INT8 quants to compare speed and quality for next tutorial hopefully.
Paints Undo App is fully updated.
Improved pipeline for ultra fast processing with auto installers for Torch 2.12.1, Triton 3.7, latest ONNX and CUDA 13 and automatic download of necessary models.
Literally 1-click to install Windows and Linux and Cloud Services
Gangnam Style English Remix Song
Made with SECourses ACE-Step 1.5 XL Premium app with literally 1-click - first try.
This is generated in 30 seconds on Windows locally with RTX 5090.
Some words are inaccurate but with few iterations of those parts can fix all so easily.