GU Turbo Lora

Distilled LoRAs are the single best quality-of-life upgrade in local image generation - the LCM, Hyper-SD, and Lightning-style adapters that collapse 25-50 steps down to 4-8 - but they have a sharp edge: every setting that's right for them is wrong for normal generation. Different step count, different CFG, different sampler. GU Turbo Lora is the GU_Nodepack answer to that hassle: it attaches a "turbo" LoRA to your model and switches the step count to match, all driven by one checkbox. Flip it on and the whole graph runs fast; flip it off and you're back to full-quality settings.

The distilled-model background is worth a sentence because it explains what this node is really automating. When a Turbo/LCM/Hyper LoRA is applied, the model is already trained to follow the prompt with very few steps - so you drop to 4-8 steps and cut CFG way down (often 0-2) or you get oversaturated, artifacted garbage. The community's most reliable mistake is keeping normal settings with a fast LoRA. This node automates the step part of that switch, which is the half people forget most. The CFG and sampler you still have to set yourself.

How it works

It's a wrapper around ComfyUI's standard LoRA loader. When enable is on, it loads the LoRA you picked in turbo_lora at full strength (1.0 on both model and clip), and the steps output becomes steps_turbo. When off, no LoRA is attached and steps becomes steps_normal. The model and clip pass through in both cases, and the lora_list output carries a <lora:name:1.0:1.0> string when enabled - handy for dropping into a filename or info block so you can tell which mode an image was made with. The GU example workflow uses a Hyper-SD15-8steps LoRA with 25 normal steps and 8 turbo steps, which is a sensible pairing.

The inputs and output that matter

• turbo_lora - which LoRA file to attach, picked from a dropdown that lists every LoRA in your models/loras folder (recursively, with "None" as an option).
• enable - the checkbox that switches between turbo and normal. This is the one you'll actually toggle.
• steps_normal - step count when the LoRA is off (default 25).
• steps_turbo - step count when it's on (default 8). Match this to your LoRA's trained step count.
• steps - the output you wire into your KSampler's steps input (convert the sampler's steps widget to an input first).
• model / clip - the modified model and clip; wire these into your encoder and sampler.

That last wiring is the whole trick: steps is an actual output, so the sampler's step count follows the checkbox automatically. One toggle, and the same graph runs either as a full-quality 25-step pipeline or an 8-step speed run.

Installing it

Part of GU_Nodepack:

cd ComfyUI/custom_nodes
git clone https://github.com/alexguryev/ComfyUI-GU_Nodepack

Restart ComfyUI, or install via ComfyUI Manager → search "GU_Nodepack" (pip deps: psutil, requests, gu-funclib). No model downloads from the pack itself - you supply the turbo LoRA from Civitai or HuggingFace (LCM-LoRA, Hyper-SD, Lightning, Turbo variants all work as long as they're LoRA files).

Gotchas

The node manages steps, not the whole recipe. It won't lower your CFG - run a turbo LoRA at CFG 7 and you'll still get the washed-out, oversaturated look the distillation docs warn about; you need low CFG (0-2) and, typically, an Euler sampler with a matching scheduler. Match steps_turbo to your LoRA's actual trained step count (8 for most Hyper-SD SD1.5 variants, 4 for Lightning 4-step, etc.) - the default of 8 isn't universal. And remember the dropdown lists every LoRA you own, so pick carefully; there's nothing validating that the file is actually a distilled LoRA. The pack is tested on Windows, best-effort elsewhere, and aimed at single-user installs.