LLS Pro KSampler Bridge

The Pro edit chain's sampler is a bridge in the literal sense: it connects the carefully prepared workspace from LLS Pro Image Edit Prepare to ComfyUI's native sampling, but it's smart about how. It tries the advanced sampling path first - with start_at_step, end_at_step, and return_with_leftover_noise control, which is what you want for img2img-style edits where you're only denoising part of the schedule - and falls back to a plain KSampler path when the advanced one isn't available. Either way, your Pro chain keeps running.

On the surface it looks like a KSamplerAdvanced: model, positive, negative, latent_image, steps, cfg (default 8), sampler_name, scheduler, noise_seed, add_noise. The controls that actually matter for edits are start_at_step / end_at_step (which slice the denoising schedule) and return_with_leftover_noise (whether to leave the latent partially noised, which matters when another pass will finish the job). cfg defaults to 8 here, but if your model is FLUX-family, remember FLUX wants CFG near 1 - this node does not auto-tune CFG the way the Simple KSampler's family presets do.

How it routes

The bridge reads the routing hints baked into the edit chain. When the model is on a native edit path, the sampling is set up to match what that path expects; the details live in the pack's per-backend sampler wrappers. For the user, the practical upshot is: wire Prepare's latent into this node's latent_image, run the decode, and the latent comes out ready for Finish to composite.

Where it fits

Prepare → Bridge → VAE Decode → Finish. The Bridge is the middle of the chain and the least interesting node to look at, but it's the one that makes the Pro chain work on both native-edit models (like FLUX2 Klein) and generic fallback paths without you maintaining two sampler graphs.

Installing it

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

Or ComfyUI Manager → search "LLS-node". It wraps ComfyUI's own samplers, so no extra dependencies.

Gotchas

The cfg default of 8 will give you blobby, oversaturated results on FLUX-family models if you forget to drop it. And because it's a bridge, it's not really meant to be used standalone in a txt2img graph - feed it a latent that came from Prepare, not an empty one, or you're just running a KSamplerAdvanced with extra steps. Keep start_at_step at 0 and end_at_step high unless you're deliberately doing multi-pass edits; those knobs exist for when you know exactly what you're doing.