fastsdcpu_vino_models

This node does one thing and one thing only: it hands you a dropdown of pre-converted OpenVINO models and returns the one you pick as a STRING. It does zero generation - no prompt, no sampling, no image out. If you drop just this node on the canvas and hit Queue, nothing renders, and that's expected. It's a convenience so you don't have to type rupeshs/FLUX.1-schnell-openvino-int4 into a text field by hand.

It's the companion to the main fastsdcpu node from the same pack. Wire this node's STRING output into the openvino_model input on fastsdcpu, set the generation knobs there, and the server knows which model to spin up. Simple as that.

What "OpenVINO" means here

OpenVINO is Intel's inference toolkit. It converts models into a graph format that runs fast on Intel CPUs, iGPUs, and NPUs - FastSDCPU's creator describes the whole project as optimized around it, and on Intel hardware it's the fastest path this stack offers. The tradeoff, straight from the pack README: OpenVINO models are much faster but require significantly more memory than the standard diffusers versions. "Fast" doesn't mean cheap. The -int8 and -int4 suffixes in the list mean the weights are quantized - smaller and quicker, with the usual small quality cost that quantization always brings. For iteration speed it's a fine deal; just don't expect the int4 Flux to be pixel-identical to the full thing.

The list

The dropdown ships with nine model IDs, mostly from rupeshs and mostly distilled one-to-few-step models:

• rupeshs/sd-turbo-openvino and rupeshs/sdxl-turbo-openvino-int8 - Stability's adversarial-distillation turbo line
• rupeshs/SDXL-Lightning-2steps-openvino-int8 - ByteDance's Lightning, the 2-step variant
• rupeshs/hyper-sd-sdxl-1-step-openvino-int8 - the 1-step Hyper-SD for SDXL
• rupeshs/sdxs-512-0.9-openvino - SDXS-512, a tiny 512px model
• rupeshs/LCM-dreamshaper-v7-openvino and Disty0/LCM_SoteMix - LCM-fused SD 1.5 models
• rupeshs/sd15-lcm-square-openvino-int8 - another 512-class LCM
• rupeshs/FLUX.1-schnell-openvino-int4 - distilled Flux at 4-bit, the heavyweight of the list

Here's where the settings actually matter. These are all distilled models, and distilled models want different treatment than the ones you're used to. The KB's distillation notes are the rule of thumb: LCM models want 4–8 steps at CFG 1–2; the Turbo line wants 1–4 steps at CFG ~2; Lightning and Hyper are trained for an exact step count at CFG ~0–1. The fastsdcpu node defaults to steps 1 and cfg 1, which is a decent starting point but not a magic one - if you pick the 2-step Lightning, run it at 2 steps, not 20. Crank the CFG on any of these and you get the classic overcooked, oversaturated look.

Installing

Same pack as the main node - ComfyUI Manager, search "ComfyUI-FastSDCPU", or cd ComfyUI/custom_nodes && git clone https://github.com/BetaDoggo/ComfyUI-FastSDCPU, then restart. That's it for this node; it has no dependencies beyond the pack. But it's useless without the FastSDCPU server running (clone rupeshs/fastsdcpu, run its install script, start the webserver on port 8000), because the actual inference happens there. First time you generate with a given model, the server downloads it from HuggingFace - and OpenVINO checkpoints are chunky, so budget a decent download.

Gotchas

• The picker alone does nothing. It returns a string; wire it into fastsdcpu's openvino_model input before expecting pixels.
• Watch your RAM. OpenVINO models are memory-hungry, and the int8/int4 variants are the ones that stay reasonable on modest boxes.
• Steps and CFG must match the model you chose. This is the #1 source of "why does it look like garbage" posts with distilled stacks - the model is fine, the settings are from a full-CFG habit.