fastsdcpu_lcm_models

A sibling of fastsdcpu_vino_models from the same pack, and a mirror image of it. One dropdown, eight distilled model IDs, one STRING output - and that string is meant to feed the lcm_model input on the main fastsdcpu node. It does no generation on its own, so a canvas containing only this node will render nothing and will look like it's broken. It isn't. It's a picker.

The difference from the OpenVINO list

Where the vino dropdown gives you pre-converted OpenVINO models for Intel CPUs and iGPUs, this list is the "standard diffusers" path: ordinary HuggingFace model IDs that FastSDCPU's server downloads and runs through its regular pipeline. That makes this the portable option - it works on whatever hardware the server happens to be running on, GPU or not, Intel or not - at the cost of the extra OpenVINO speedup on Intel boxes. If your machine is a random CPU with no Intel-specific acceleration, this is usually the list you'll reach for.

The list, and what it's actually made of

• stabilityai/sd-turbo and stabilityai/sdxl-turbo - Stability's adversarial-distillation pair. SD-Turbo is a 512px SD 1.5-class model; SDXL-Turbo is the 1024px one.
• rupeshs/SDXL-Lightning-2steps - ByteDance's 2-step Lightning, widely regarded as the strongest of the SDXL speed tricks.
• rupeshs/hyper-sd-sdxl-1-step - Hyper-SD at its most aggressive: one step.
• rupeshs/sdxs-512-0.9-orig-vae - SDXS-512, a small fast 512px model.
• SimianLuo/LCM_Dreamshaper_v7 - the LCM fusion of DreamShaper v7, a 512px SD 1.5 model.
• latent-consistency/lcm-sdxl and latent-consistency/lcm-ssd-1b - the original LCM line for SDXL and the smaller SSD-1B variant.

These are the models the pack's README is really pointing at when it says FastSDCPU is "optimised around LCM." The KB's distillation essay has the settings for each family: LCM models run 4–8 steps at CFG 1–2, the turbo line runs 1–4 steps at CFG ~2, Lightning and Hyper want their exact step count at CFG ~0–1. The fastsdcpu node's defaults (steps 1, CFG 1) are a fine first attempt, but match them to the model you picked - a 2-step Lightning run at 20 steps artifacts, and CFG 7 on any of these looks like someone left it in the oven.

Installing

Same as the whole pack: ComfyUI Manager, search "ComfyUI-FastSDCPU", or cd ComfyUI/custom_nodes && git clone https://github.com/BetaDoggo/ComfyUI-FastSDCPU, then restart. And same as always with this pack, the real requirement is the FastSDCPU server (clone rupeshs/fastsdcpu, run its install script, start the webserver on port 8000) - the node is just a dropdown that hands a model ID to the server over HTTP. First generation per model pulls it down from HuggingFace, and these are full diffusers checkpoints, so the SDXL ones are multi-gigabyte downloads.

Gotchas

• Wire it into lcm_model, not openvino_model. Wrong input and the server will try the OpenVINO path with a model that isn't converted. The vino dropdown is the one for the OpenVINO input.
• Some of these are 512-class models. SD-Turbo, DreamShaper LCM, and SDXS are SD 1.5-sized; feed them 1024 and you're asking a 512px-trained model to draw outside its training. Keep width/height at 512 for those.
• Steps and CFG are per-model. The dropdown can't set them for you - that's your job on the main node.