KSampler (LoRA Sigma Inverse)

Here's a scenario you've probably met: a LoRA that's great for fine detail but wrecks composition at full strength, or a distilled-model LoRA that goes unstable if you run CFG too high. The usual fix is a fixed reduced strength and a shrug. KSampler (LoRA Sigma Inverse) is the no-shrug version: a KSampler with the LoRA loader built in, where LoRA strength ramps up over the course of sampling as sigma falls. Composition gets settled early without the LoRA, then the LoRA's influence fades in for the detail work.

The strength at each step follows an explicit schedule: min_lora_strength + (max_lora_strength - min_lora_strength) * (1 - sigma[i] / max_sigma). In plain terms: at high sigma (noisy start), LoRA runs at min_lora_strength; at low sigma (fine detail), it's at max_lora_strength; in between, it eases from one to the other. If you want the LoRA to only appear mid-run, the min_lora_step/max_lora_step gates let you zero it out outside a step window.

How it works

Everything else is a standard KSampler: model, positive, negative, latent_image, steps, cfg, sampler_name, scheduler, denoise, seed. The LoRA is loaded from your loras folder by name and injected via ComfyUI's bypass path when possible (injected once, strength updated per step with multipliers - no per-step repatching), falling back to hook-based scheduling for non-bypass-compatible LoRAs. The clever extras: scale_cfg pulls CFG toward 1.0 as LoRA strength climbs - aimed at distilled-model LoRAs that destabilize at high CFG - and clamp_cfg_after snaps CFG to 1.0 after a given step.

The inputs that matter

• lora_name - pick from your installed LoRAs.
• min_lora_strength / max_lora_strength - the ramp endpoints (defaults 0 → 1, so the LoRA is off at the start and fully on at the end).
• min_lora_step / max_lora_step - step gating; -1 means no bound.
• scale_cfg - enable for distilled-model LoRAs that hate high CFG.
• Everything else - standard KSampler behavior.

Output is a LATENT.

Installing it

Part of Skoogeer-Noise. Manager → search "Skoogeer-Noise", or:

cd ComfyUI/custom_nodes
git clone https://github.com/ttulttul/Skoogeer-Noise

Restart ComfyUI. Deps are torch, numpy>=1.26, einops, pyyaml>=6.0.3 - no extra models.

Common gotchas

Two things trip people. First, this is model-only - there's no CLIP input, so you still need a normal CLIP Text Encode node upstream to produce your conditioning; the sampler just doesn't encode text itself. Second, the defaults ramp from 0 to 1, so if you expected the LoRA to be active the whole time, you're effectively running with it off until mid-run - set min_lora_strength to your usual strength if you want a floor. And remember the LoRA has to be in the folder ComfyUI scans (ComfyUI/models/loras); an empty lora_name dropdown means the file isn't where the node looks.