LCMLoader_SDTurbo

Every other loader in this pack is about LCM. LCMLoader_SDTurbo is the odd one out - it loads SD-Turbo, Stability's adversarial-distillation model, which was the other big "generate in 1–4 steps" technique of late 2023. The author threw it in as a bonus path, and it's a genuinely different animal from the LCM nodes: SD-Turbo wants CFG around 0–2 (the paired LCMGenerate_SDTurbo node hardcodes guidance_scale=0.0), it's an img2img pipeline at heart, and it's noticeably more aggressive about its signature look.

You'd reach for it when you want turbo-style fast img2img with a reference-image option, and you're already living in this pack. It pairs with LCMGenerate_SDTurbo, which is the only generate node that understands this pipeline.

How it works

Unlike the LCM loaders, which assemble a pipeline piece by piece from subfolders, this one does the diffusers equivalent of a one-liner: StableDiffusionImg2ImgPipeline_reference.from_pretrained(...), loading whichever folder you pick from the model_path dropdown (that dropdown lists subfolders of ComfyUI/models/diffusers/ - so it's a chooser, not a free-text path). It then applies ToMe and - unconditionally, with no device toggle - enables xformers and sequential CPU offload.

The reference_only toggle changes which pipeline class it instantiates: disable gives you a plain reference-capable img2img pipeline; enable switches to the reference-conditioning variant so the generate node can take a ref_image plus style_fidelity.

The inputs that matter

Only three:

• model_path - a dropdown of the diffusers folders in models/diffusers. This is where the SD-Turbo model goes; the pack's own workflows ship with a folder in that location. Leave it on whatever you placed there.
• tomesd_value - ToMe merge ratio, default 0.6, 0 = off.
• reference_only - disable/enable. Enable it if you want the generate side to accept a reference image for style borrowing.

The class output wires into LCMGenerate_SDTurbo.

How to install it

Standard pack install - ComfyUI Manager (search "LCM_Inpaint_Outpaint_Comfy") or:

cd ComfyUI/custom_nodes
git clone https://github.com/taabata/LCM_Inpaint-Outpaint_Comfy
cd LCM_Inpaint-Outpaint_Comfy
pip install -r requirements.txt

Restart, then put an SD-Turbo diffusers-format model folder into ComfyUI/models/diffusers/ so it shows up in the dropdown.

Common issues

The big trap is settings, not installation. This pipeline runs at guidance_scale=0.0 - the generate node just doesn't give you a choice. If you've been tuned by the LCM nodes to think "CFG 1.8", fine, but if you came from normal SD and try to raise CFG for quality, you can't. That's correct behavior for SD-Turbo, but it feels like a missing knob if you don't know the model.

Also note there's no device input on this loader - the "offload" branch always runs. On a CPU-only box this will be slow and there's no escape hatch, unlike the LCM loaders which have a CPU mode.

Honest take: this node is the pack hedging its bets. It's not the best way to run SD-Turbo even in its own era (the ComfyUI native path is simpler), and SD-Turbo itself aged worse than LCM did - Lightning largely replaced it for the same job. Keep it around for the pack's img2img-loopback workflows; don't build your whole stack on it.