What CPU-only image generation genuinely cannot do (and what it can)

Hathor Studio runs on CPUs on purpose: it means free, no account, and no GPU bill. It also means real limits, and it is better to state them than to have people discover them. What works on CPU - SD 1.5-class models with LCM at 4-8 steps: about one to three minutes an image on the Studio's server. - Character reference (IP-Adapter) and layout reference (ControlNet) on top of that. - Non-diffusion tools: background removal, ImageMagick, G'MIC and GIMP filters, conversion, vectorising. The Studio runs these on its server. What does not work well on CPU - Many images at once. The Studio renders one image at a time; people using the site go first, background work waits. - SDXL and larger models at a usable speed. - Video. The Studio offers text-to-video and image-to-video, but only through your own key (such as fal.ai) or your own GPU. - Training LoRAs. - Large tiled upscales with diffusion (each tile is another full render). The design answer Free CPU for everyone, and your own engine for speed: the exact engine as one Python file, a free Colab, your own Modal GPU, or your own key. All on https://hathor.soapbox.community/engines Your numbers If you run SD 1.5 with LCM on your own CPU, post: CPU model, cores, RAM, resolution, steps, seconds per image. Real numbers will help the next person decide whether their laptop is enough.