Smart Sampler Settings (Legacy)

Some nodes exist because they're good, and some exist because your saved workflow is from last year. Smart Sampler Settings (Legacy) [Eclipse] is firmly the second kind. It's a deprecated node, kept loadable so pre-4.0 Eclipse workflows don't break, and it emits a single PIPE - a Python dict of sampler settings - that downstream IO nodes read.

If you're building something new, the modern Smart Sampler Settings node in the same pack does the same job with one unified seed and a simplified upscale value. Read this for the wiring you inherited, not as a recommendation.

What it does

One node, one output, a lot of widgets. The features field decides which of them actually make it into the pipe. It's a comma-separated list - sampler,scheduler,steps,cfg,denoise by default - and in the UI a JS combo-chip replaces the raw widget, so you get a row of little toggles instead of a string. A chip that's off means not configured, not configured with the widget's default. That's the design idea the whole pack's pipe system rests on: don't ship keys you didn't set.

The chip names come straight from the source's feature list: sampler, scheduler, steps, cfg, guidance, denoise, seed, noise_injection, upscale and allow_overwrite. Two of them bundle more than one value - noise_injection writes both sigmas_denoise and noise_strength, and upscale writes upscale_steps, upscale_denoise and upscale_value together. allow_overwrite emits an internal flag rather than a sampling value.

The sampling widgets themselves are the usual suspects. sampler_name and scheduler pick the algorithm; steps (30), cfg (5.0) and denoise (1.0) are the standard trio; guidance (3.5) is the separate scalar flow-matching models want; sigmas_denoise (0.45) and noise_strength (0.5) are the noise-injection pair for advanced sampling; upscale_value (1.5) is a plain scale factor for upscale workflows. seed runs from -3 up to 2^64-1, which is a wider range than the stock sampler's seed widget offers - connect a Seed node rather than trusting the low end to mean anything specific. allow_overwrite is a boolean: when it's on, direct inputs on the downstream IO node are allowed to beat the values travelling in this pipe.

Where the pipe goes

The pipe output is a dict consumed by an IO node - Pipe IO Sampler Settings [Eclipse] in the current pack, or one of the legacy v2.1/v2.2/v2.3 nodes if that's what your workflow carries. The IO node merges the pipe with whatever direct inputs you connected, based on the overwrite policy, and re-exports every channel as an individual typed output. That's the whole point of the pipe ecosystem: one wire replacing a dozen, and the values are still addressable at the other end.

This node is also flagged as an output node in the schema, which is why it can sit on the canvas feeding nothing visible and still execute.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_Eclipse
pip install -r ComfyUI_Eclipse/requirements.txt

Or ComfyUI Manager → search ComfyUI_Eclipse → restart. It'll show up under 🌒 Eclipse/ Legacy (the category is literally named Legacy - that's your hint about its future).

Two things to know

First, the documented migration path. Eclipse 4.0.0 removed the deprecated nodes from the codebase apart from the ones it still ships for compatibility, and it has a Workflow Migration Tool - usable as a node, or as python tools/migrate_workflow.py <path_to_workflow_or_directory>, with a search-and-replace mapping in tools/migration_mapping.txt. It converts supported deprecated nodes to current equivalents while keeping saved values and connections, and takes backups. If you're still on a legacy sampler node, that's the one afternoon worth spending.

Second, a behavioural note that costs people afternoons: on any guidance-distilled model - Flux Schnell, Z-Image Turbo, the lightning Qwen builds - cfg should stay at 1, and at CFG 1 the negative prompt does nothing at all. It's not broken; the distillation baked the guidance in. Writing a long negative and wondering why it's ignored is one of the most common traps in modern workflows, and it has nothing to do with this node's age. Words are the tool that still works, and it's cheaper than fighting a sampler.