Universal Input Hub by Steve Lasmin

The whole sampler rig, collapsed into one box

You know the drill: every generation starts with the same ten-node chain - Empty Latent Image, CLIP Text Encode twice, a seed, a steps widget, a CFG widget, a sampler/scheduler picker, a noise source. It's fiddly, it sprawls, and when a shared workflow uses slightly different widget names, your graph turns into a nest of reroutes. Universal Input Hub is a bet that you can collapse most of that into a single node: it takes a model, a CLIP encoder, and a prompt, and hands back the noise, the sigmas, the sampler object, and both conditioning tensors.

The "hub" idea isn't new - rgthree's Context nodes already bundle a dozen connections down one wire. This is the single-node version of that impulse: everything a sampler wants in one place, with tooltips on every input. Whether it beats rgthree for you comes down to taste. It is not a latent generator, which is the first thing to know before you wire anything up.

How it works

The mechanism is honestly pretty clean. You give it a max_size (the longest side) and an aspect ratio - either a preset like 16:9 or a custom width_ratio/height_ratio. It solves for the other dimension, rounds both to your multiplier (64 for SD 1.5, 32 for SDXL, 16 for Flux - the tooltip says so), and hands back width and height as INTs.

From there it builds everything else: a NOISE object from your seed, a SAMPLER object from your sampler picker, standard sigmas computed from the model, scheduler, and steps, and the positive/negative CONDITIONING from your prompt and negative_prompt. The interesting output is flux_sigmas: a dimension-aware schedule computed purely from the output size and step count, using the same time-SNR shift math and empirical constants ComfyUI's own Flux-2 sampler uses. No Flux model loaded? Doesn't matter - that path never touches the model.

The inputs that matter

You'll actually set six of them: max_size, aspect_ratio, seed, steps, cfg, and the two prompts. cfg's tooltip is worth trusting: Flux likes 1.0–4.0, SDXL wants 7.0–8.0. trigger_words is a small nicety - text that gets auto-prepended to your prompt with a ". " separator, handy for sticking a LoRA's trigger in without retyping it. sampler and scheduler are just the standard ComfyUI lists, so no new vocabulary to learn. The preview output is a human-readable summary string of every setting - genuinely handy when you've buried a workflow and need to know what it ran.

What wires where

For a normal KSampler you still need an Empty Latent Image node - the hub gives you width/height to feed it, but not a latent. The README glosses over that; don't let it trip you up. Wire positive and negative into the sampler, model through, and the seed/steps/cfg/sampler_name/scheduler_name outputs straight into the matching widgets. For SamplerCustomAdvanced, grab noise, sigmas, and sampler instead. And if you want to see everything the node computed, the help output is a full in-node manual.

Installing it

ComfyUI Manager → search "Universal Input Hub", or clone it directly:

cd ComfyUI/custom_nodes
git clone https://github.com/Eklipsis/universal_input_hub_by_steve_lasmin.git

Then restart ComfyUI. It shows up under utils. One real plus: pyproject.toml declares zero Python dependencies - no torch pinning, no pip-install dance, nothing to break your environment. It's also a pure backend node with no JavaScript, so it sidesteps the Nodes 2.0 frontend rewrite problems that took down JS-heavy packs.

The honest gotchas

• flux_sigmas are timesteps for Flux-2-style models. Feed them to an SD/SDXL sampler and you'll get garbage. Use the plain sigmas output for normal models.
• The node is marked OUTPUT_NODE, so it always executes on every run, even if nothing downstream consumes it. Not a bug - just know it's there.
• It's tiny and new, with effectively zero community footprint - search as I might, nobody's talking about it. The author is a Boosty (Russian-patreon-style) creator, and the license is CC BY-ND: you can use and share it, but no derivatives without permission. If you hit a wall, expect to debug it yourself.