Chimera (Multi-Sampler) 🦁

"First pass with one sampler, second pass to refine" is one of the most common tricks in ComfyUI, and it usually looks like two KSamplers chained together, each re-adding noise, each with its own step count you're hand-tuning. Chimera is the version of that idea that doesn't waste your time or your denoise. It's a multi-stage sampler: you feed it two (or more) Sampler Settings bundles - the pack's own config node - and it runs them over one image with an explicit step budget, handing off cleanly between stages.

The thing that separates it from "two KSamplers" is the handoff mode. Continuous (the default) builds one sigma schedule and gives each stage a slice of it. Stage two resumes exactly where stage one stopped - no fresh noise added - so the run is one honest traversal of one noise trajectory, just with different samplers, cfg and eta per segment. It's like chaining KSampler Advanced by start/end step, except the boundary is a setting instead of arithmetic you do in your head. Restart is the classic two-samplers pattern: every stage builds its own schedule and adds its own noise, so stage two needs denoise < 1 or it will destroy stage one's work. Different effect, not better - continuous for refinement, restart for genuine img2img re-invention.

Where the boundary falls is its own input, step_split, and you get the same decision in whatever unit you're thinking in:

• stage steps - each stage runs exactly the steps it declares.
• at step - cut a shared total_steps at handoff_step, so you can sweep the boundary without editing the Settings nodes.
• at percent - same cut as a percentage of the steps (60% of 30 → 18/12).
• at sigma - hand over once the noise level crosses handoff_sigma. The one genuinely tricky one: the sigma scale is model-dependent. Flow-matching models (Flux, SD3, Krea) run 1.0 → 0, so the useful range is 0–1; SD/SDXL start around 14.6 and you'd want single digits. The report output prints the curve's actual range, so one run tells you which world you're in.

The stage_a / stage_b inputs take KINBURG_SAMPLER_CFG bundles from the pack's Sampler Settings node - leave stage_b empty and Chimera just runs stage A as a plain single-pass sampler, so it doubles as a fancy KSampler. There are escape hatches if you need them: model_b / positive_b / negative_b for a different LoRA stack or refiner on stage two (must share the main model's noise schedule), model_negative for checkpoints shipped as a model + uncond-model pair (Ideogram), and guider to plug in any custom guider - though wiring one ignores this node's own positive/negative and per-stage cfg, so only do it when you must.

Worth knowing: the sampler measures its own time - time and seconds outputs - which the docs note can't be skewed by how ComfyUI schedules other branches of the graph, unlike Start/Stop Timer nodes. And gen_extra_info carries what it resolved at run time, wireable into a Generation Info node so your settings dump records what actually happened.

Installing

Ships in Kinburg-Nodes: ComfyUI Manager → search "Kinburg-Nodes", or clone into custom_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/Kinburg/Kinburg-Nodes
# restart ComfyUI

You'll also want the pack's Sampler Settings node, which ships in the same pack. No extra dependencies.

Gotchas

In continuous mode the shared curve takes its scheduler and denoise from stage A - a later stage's own denoise is ignored (a different scheduler gets spliced in, and the report tells you). The Settings node's seed_mode / seed_step are Ouroboros-loop dials and are ignored here; only seed is used. And if you set total_steps longer than the sum of the stages' steps, the tail of the curve is left unwalked on purpose - residual noise comes out the other end, which is a feature if you know you want it and a surprise if you don't.