LLS Simple KSampler

Every sampler in ComfyUI ultimately calls the same native sampling code, but each model family wants different knobs. FLUX wants CFG at 1 and a guidance value. SDXL wants real CFG. Qwen and Z-Image have their own conventions. LLS Simple KSampler is the "stop switching samplers" answer: one node that wraps ComfyUI's native sampler, carries family-aware defaults, and - the part that makes it different - can take a repair_info wire and run a masked repair pass through the exact same node that does your txt2img.

The one-liner from the pack's own docs: it reuses ComfyUI's native sampling pipeline and doesn't rewrite any sampling math. That's the reassuring part - this isn't a reimplementation that'll give subtly different results, it's a wrapper with better defaults.

The knobs that matter

• quality_preset - Family Default, Manual, Fast, Balanced, High Quality. When it's not Manual, the preset overrides steps/cfg/denoise. This is the "don't make me think" dial.
• seed - -1 means "randomize every run." Set a real number to lock it.
• steps, cfg, sampler_name (default euler_ancestral), scheduler (default karras), denoise - the classic sampler controls, with denoise_mode offering manual or auto_from_repair (let the repair chain pick the right denoise for you).
• flux_guidance (default 3.5) - used when the model is FLUX-family; it accepts a plain float or a wired value.
• adapter_mode - auto routes by model family, or force sd_classic, flux, sd3, qwen, zimage.
• model_family - Auto infers from the model, or set by hand.

It outputs latent plus a sample_info JSON string describing the run - feed that to LLS Save Image and your PNG metadata records the actual settings used.

The repair bit

Wire an LLS_REPAIR_INFO into the optional repair_info input and this same node becomes the sampler for a masked repair chain - the pack's Simple workflow runs Load Image → mask → LLS Simple Repair Prepare → this KSampler → VAE Decode → LLS Simple Repair Finish. With denoise_mode on auto_from_repair, the recommended denoise from the repair preparation is used automatically. No second sampler node, no parallel graph.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/Gin3601/LLS-node
# restart ComfyUI

Or ComfyUI Manager → search "LLS-node". No pip dependencies beyond what ComfyUI already has - it imports ComfyUI's own comfy.sample and comfy.samplers.

Gotchas

Because it leans on ComfyUI's internals, it needs to run inside a real ComfyUI environment (it errors clearly if comfy.sample isn't importable - you can't run this node standalone). And the flux_guidance input is typed loosely on purpose (STRING/FLOAT/INT all accepted) so it can be wired from a generation config; if you feed it something nonsense, the node rejects the input before running. It's a capable node, but it's also where beginners most often trip on the same classic ComfyUI trap: that control_after_generate dropdown on the seed. Set it to fixed when you've got a seed worth keeping.