🌊👌 FlowState Simple Latent

FlowState Simple Latent is the boring one in the suite, and that's its charm. It produces a latent batch for a sampler to denoise - empty canvas or an encoded input image - and nothing else. If you've used Empty Latent Image from core ComfyUI, you already know this node; the difference is a preset resolution list (18 ratios, biased toward the 16:9 and 21:9 sizes the FlowState suite was built around) and an orientation toggle that swaps them vertical. It's the latent node you'd grab when you want fewer widgets and a thumbnail-sized canvas without typing 1920×1080 by hand.

How it works

The mechanism is dead simple and fully visible: resolution picks a preset (or Custom to use custom_width / custom_height), orientation flips the aspect ratio, and latent_type decides between Empty Latent - a torch.zeros tensor shaped [batch, 4, height/8, width/8], the 8s being the VAE's spatial downsampling - and Input Image, which encodes the image on the optional input_img socket through the required vae input. custom_batch_size stacks that many latents in one batch. One output, Latent Image (LATENT), which feeds a sampler's latent input - a KSampler, or the latent socket on this pack's FlowState Flux Engine.

Inputs worth knowing

• resolution - presets run from 512x512 - 1:1 up to 4096x4096, with the production ratios (1920x1080 - 16:9, 2212x948 - 21:9) front and center. "Custom" activates the width/height pair below.
• orientation - Horizontal or Vertical; vertical swaps the chosen preset's dims.
• latent_type - Empty Latent, or Input Image (which needs input_img wired and a vae to encode with).
• vae - technically required even for empty latents; only actually used when you encode an image. Grab the VAE that matches your model family.

The honest take

There is a subtle trap in the "Input Image" mode: pick it but don't wire anything into input_img, and the node silently falls back to an empty latent rather than erroring. That's actually a feature for batch scripts - it won't crash a run - but it will quietly generate from noise when you thought you were doing img2img. If you specifically want to hand a file path to a latent source without wiring an IMAGE, the sibling FlowState_LatentSource adds an "Uploaded Image" mode with a file picker; that's the only functional difference between the two nodes. Otherwise, this is the leaner, calmer choice.

Install

It ships with the FlowState Creator Suite, so installation is the pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/flowstateeng/FlowState-Creator-Nodes

then restart ComfyUI, or install FlowState-Creator-Nodes through ComfyUI Manager. The pack carries no pip dependencies - its requirements.txt is empty, and this node in particular is just a thin wrapper over tensor math and ComfyUI's own VAE. Bring a matching VAE and you're done.