Goal is to fix the quality and the bias of Ernie turboEuler simple, cfg 1, 8 stepsVRAM note: If you have 16 GB VRAM or less, I recommend using one of the quantized versions instead of BF16. BF16 can work, but it leaves very little headroom for Forge, prompt enhancers, browsers, Discord, etc.Between the FP8 options, I currently prefer MXFP8. In my testing on a 5060 Ti, it is usually around 4–6 seconds faster than standard FP8 while keeping very similar quality.NVFP4 is the smallest and fastest option, but it is more experimental and may be less consistent in complex scenes.

ET_Center

some more training and new merge bases on the actual turbo model

etCenter_v20Mxfp8.safetensors

Files with SHA256: b261d2e13564fae6f2ee6de359745a07891d05b52e034e6b8ba773488cfa704d