LCMGenerate_SDTurbo

LCMGenerate_SDTurbo is the odd one out in this pack: instead of the LCM Dreamshaper pipeline, it runs an SD-Turbo model. SD-Turbo is Stability AI's adversarial distillation - a different speed technique from LCM, trained to produce images in 1–4 steps with its own look. The pack is LCM-first, but this node quietly adds a Turbo path with a matching LCMLoader_SDTurbo loader (also in this repo), which pulls a diffusers SD-Turbo model from your models/diffusers folder.

The inputs

It's an img2img-style sampler with a few differences from the LCM generation nodes:

• pipe - from LCMLoader_SDTurbo (not LCMLoader). This matters; the loader selects the model and tomesd settings.
• device - cuda/cpu, lowercase here.
• prompt and negative_prompt - both multiline. Note the SD-Turbo node separates them, unlike the LCM nodes' single text field.
• reference_only (disable/enable) - when enabled it uses reference_image and style_fidelity (default 0.5) to steer toward a reference, mirroring the pack's reference-only trick on the Turbo path.
• image, strength (default 1.0) - the img2img pair. Lower strength for subtler changes.
• Plus seed, steps (4), width/height, cfg, batch (default 1, max 1000).

The interesting detail

Read the source and you'll spot something unusual: the code passes guidance_scale=0.0 no matter what cfg says. The cfg input is there in the UI but effectively ignored - SD-Turbo, like other adversarial distillations, has the guidance baked in and wants CFG at zero. That matches the distillation playbook: don't fight a distilled model with CFG. So when someone tells you this node's cfg slider does nothing, they're basically right.

Setup and honesty

Same pack install (ComfyUI Manager → search LCM_Inpaint-Outpaint_Comfy, or clone + pip install -r requirements.txt). Then download an SD-Turbo model in diffusers format and point LCMLoader_SDTurbo's model_path dropdown at it (that loader reads the subfolders of models/diffusers). LCM_Dreamshaper_v7 isn't needed for this path.

This node is clearly a bolt-on - SD-Turbo dropped in November 2023, right as this pack was active, and the author added it as a variant. It works, and reference-only Turbo generation at 1–4 steps is genuinely quick. But it's niche within the pack: if you're here for Turbo, native ComfyUI's Turbo/Adversarial-Diffusion support is more maintained. Reach for this node if you want the pack's single-node reference-only workflow on a Turbo model specifically.