LCM on CPU: why Hathor Studio takes one to three minutes, and the settings behind it

Hathor Studio's own engine is Dreamshaper 8 (a Stable Diffusion 1.5 model) running on the CPUs of a MELEK server. On CPU a normal 25-step render takes many minutes. LCM brings it down to about one to three minutes an image. What LCM is A latent consistency model is a distilled model that predicts the finished image in a few steps instead of many. The LCM-LoRA (Luo et al., 2023) applies that distillation to an existing SD 1.5 or SDXL checkpoint as a small add-on, so a fine-tune such as Dreamshaper can use it. Settings for LCM - Sampler: LCM. - Steps: 4-8. Four is usable; six to eight adds detail. More does little. - CFG: about 1-2. Normal values (7) produce harsh, burnt images. - Negative prompt: has little effect at CFG near 1. - LCM-LoRA weight: 1.0 is typical. What CPU-only genuinely cannot do - Many images in parallel: the Studio's server renders one image at a time, and people using the site go before the Studio's own background work. - Practical SDXL at speed, video, or LoRA training. Those want a GPU. If you want it faster Use your own capacity at https://hathor.soapbox.community/engines — the exact engine the studio runs as one Python file on your computer or a free Colab, or your own Modal GPU (about two seconds an image), or a fal.ai or Gemini key. Reproduce it locally Same checkpoint (Dreamshaper 8), the LCM-LoRA, LCM sampler, 6 steps, CFG 1.5, 512x512. Post your machine and your time per image; that number is useful to everyone on this board.