Boyolatent (Empty Latent)

Let's be honest about what this is: Boyolatent is ComfyUI's built-in Empty Latent Image, re-skinned with aspect-ratio presets and an initialization-mode dropdown. That's the entire difference. If you already know how the core node works, you know 90% of this one. What's worth your time is the two additions, because one of them is genuinely useful and the other is a foot-gun hiding in a dropdown.

A latent is the compressed representation the diffusion model actually works in - the VAE's low-dimensional picture of the image, 8× smaller per side. An "empty" latent is just a starting canvas of noise (or zeros) that the sampler denoises into an image. You need one at the head of every text-to-image workflow; it's the shape of the output before the model draws anything.

What the presets get you

Instead of typing width and height as separate numbers and hoping they're compatible, you pick from a dropdown: 1:1 Square, 4:3, 3:2, 16:9 HD, 21:9, 2:3, 3:4, 9:16 Tiktok-Shorts. You set a base width, the node computes the matching height from the ratio, and it rounds both to multiples of 8 (which is what the VAE and model need - non-multiples produce errors or silent cropping). The rounding goes both ways: it adjusts the width too, to preserve the exact ratio. That's a small quality-of-life win if you flip between aspect ratios often and are tired of hand-computing 512×896.

batch_size (1–64) sets how many latents you get in one shot - batch rendering multiple images in a single sampler pass.

The init_type dropdown is the part to understand

• normal (default) - a latent full of random noise (randn). This is what text-to-image needs: the sampler starts from noise and shapes it. Use this for nearly everything.
• zeros - an all-zero latent. This is not a "blank image to fill"; a zero latent fed to a full-strength sampler produces mush. The real uses are the ones where you deliberately skip the noise start: img2img-adjacent tricks, latent upscaling pass-throughs, or masks that expect a neutral starting canvas.

Where people get burned: clicking zeros "to start from a blank canvas" and getting a gray smear, then assuming the node is broken. It's not - you picked the wrong mode. normal is the default for a reason.

Inputs and outputs

Required: width (64–2048, default 512), batch_size, aspect_ratio, init_type. One output: LATENT, wired into your KSampler's latent_image input exactly where the core Empty Latent Image output goes.

Installation

It's part of Boyonodes, no extra dependencies:

cd ComfyUI/custom_nodes && git clone https://github.com/DragonDiffusionbyBoyo/Boyonodes

restart, find it under the Boyonodes category.

Should you reach for it?

If you already use the core node and never think about aspect ratios, honestly - skip it, you're not missing anything. If you're constantly switching between portrait, landscape, and square output sizes, or you want the zeros mode available without hunting for a specialized node, this is a tidy convenience. It's a variant node in the truest sense: same job, nicer input ergonomics, zero magic. That's fine - not everything in a kitchen-sink pack needs to be revolutionary.