Dazzle Shark Sampler (DazzleNodes)

Here's a sampler with a SAMPLER input socket on it. Dazzle Shark Sampler is the split-brain version of the pack: instead of owning its solver and its execution in one node like the plain DazzleKSampler does, it runs a sampling job but lets the solver come from somewhere else. You configure the run here - steps, schedule, denoise, CFG, mode - and feed it a SAMPLER object (from DazzleKSampler_Advanced, DazzleClownSampler, or anything else that emits one). It's the orchestrator half of the split, and it's ported straight from RES4LYF's SharkSampler.

The payoff is workflow flexibility. Want to try three different solvers on the same prompt, schedule, and seed? Build the three SAMPLER factories, wire them into three Sharks, and A/B them by toggling one connection instead of reconfiguring a monolithic node each time. That's the pattern RES4LYF users adopted for systematic sampler comparisons, and it's why a "split orchestrator" exists at all.

How it works

The full widget set is here: scheduler (default beta57), steps, steps_to_run, denoise, cfg (with the channelwise-CFG note: negative values switch modes), seed, and sampler_mode (unsample / standard / resample). But the interesting plumbing is in the optional inputs. There's a sampler socket for the SAMPLER object, plus a genuinely useful fallback baked into the source: if the incoming latent dict carries model, positive, negative, or sampler keys, the node pulls them out automatically when the corresponding sockets aren't connected. That means a single rich latent dict can carry the whole job forward - which is how this node fits into the context-bus style of workflow plumbing without you dragging every wire by hand.

latent_role behaves like the other Dazzle samplers: auto inspects the incoming latent dict's shape and decides whether the upstream tensor is init, pure noise, or layered noise-plus-init. Same five choices, same console warnings on mismatch.

The inputs and outputs that matter

For a first run, the two sockets that define the job are sampler (the SAMPLER you're executing) and latent_image (the LATENT to run on). Everything else - steps, denoise, CFG, seed - is familiar KSampler territory. Outputs are the pack's standard triple: output (LATENT), denoised (LATENT), and options (OPTIONS).

Installing it

Same pack, one clone:

cd ComfyUI/custom_nodes
git clone https://github.com/DazzleNodes/ComfyUI-DazzleKSampler.git

Restart ComfyUI; the node lands under RES4LYF/samplers. Or search "Dazzle KSampler" in ComfyUI Manager. No pip dependencies, no model downloads.

Where people get burned

Most confusion here is about what's missing: this node has no sampler_name or eta widgets, and people hunt for them. They're not missing - that's the point. The solver and its tuning live in the SAMPLER factory you feed in; this node is pure execution. If you connect a Shark without a sampler input and the latent doesn't carry one, there's nothing to run, and you'll get a failure rather than a graceful fallback. Wire the factory first, then the orchestrator. If you don't actually need to swap solvers mid-graph, the plain DazzleKSampler is simpler and does the same job in one node.