🔶 KSampler Splitted - Expert

This is one denoise stage in Chaosaiart's split-sampler animation system, and you're meant to chain two or more of them together to reconstruct - by hand, with full control - what chaosaiart_KSampler_a1a does automatically in one bundled node. The README describes this "Expert" workflow explicitly as the all-in-one KSampler "dissected into its individual components," which is exactly what this node is: one component, repeatable.

The chaining mechanism is the whole trick worth understanding here. Every instance of this node takes a k_attribut input (type K_ATTRIBUT) and produces a SUB_K_ATTRIBUT output of the same type - so your first stage takes its k_attribut from chaosaiart_Ksampler_attribut's MAIN_K_ATTRIBUT output, and every stage after that takes its k_attribut from the previous stage's SUB_K_ATTRIBUT. That's how a two-, three-, or four-stage split denoise gets built: daisy-chain as many of these as you want, each one running a slice of the total denoise trajectory.

Per-stage fields: end_at_step (default 25) is where this stage's denoise slice stops - set it lower than your total step count if this is an early stage in a multi-stage chain. denoise (default 1, range 0-1) controls how strongly this stage denoises relative to its input latent; on any stage after the first, you'll typically want this well below 1 since you're refining an already-mostly-formed image rather than starting from noise. seed, cfg, sampler_name, and scheduler work like a normal KSampler for this stage - though whether they actually get used or get overridden by the first stage's settings depends on the Synchro_* toggles you set back on chaosaiart_Ksampler_attribut. This is the detail that catches people: change a value here, see no effect, and the reason is almost always that Synchro was set to lock that field to stage one. Also required: model, positive, negative, and latent - standard sampler wiring.

Outputs: SUB_K_ATTRIBUT for chaining to the next stage, SAMPLER (the resulting latent), and Info, a text status readout.

Since this pack is aimed at SD 1.5-class checkpoints doing frame-by-frame animation, stick to DDPM-era conventions here - DPM++ 2M Karras at 20-30 total steps split across your stages is a reasonable starting point, not the low-step Euler-family settings you'd reach for on a flow-matching model like Flux or Z-Image; those don't share the checkpoint architecture this pack was built around.

Install: search "Chaosaiart-Nodes" in ComfyUI Manager, or "Install via Git URL" with https://github.com/chaosaiart/Chaosaiart-Nodes. Manually: cd ComfyUI/custom_nodes && git clone https://github.com/chaosaiart/Chaosaiart-Nodes, then restart. Linux additionally needs pip install opencv-python and pip install tqdm inside your venv.

If you're not already fighting a specific limitation of the bundled chaosaiart_KSampler_a1a, you probably don't need this node - it's real complexity for real control, and it's easy to build a chain where a Synchro toggle silently overrides a value you just spent time tuning on a later stage.