CustomLatentImage-NYJY

Core ComfyUI's Empty Latent Image asks you to type width and height and, if you're new, you type 1024×1024 and wonder why nothing else fits. This node is the quality-of-life replacement: instead of raw numbers you pick from a curated list of 29 SDXL-native aspect ratios, each one a label like 13:19 - 832x1216 that tells you both the aspect and the pixel dimensions. It builds the empty latent the same way, but it saves you the mental math and the common mistake of feeding a sampler a resolution the model wasn't trained on. If you generate SDXL or anything built on SDXL conventions, this is the blank-canvas node you'll actually enjoy using.

How it works

The ratio list is the standard SDXL bucket set (the one Stability shipped with the base model) - everything from 1:2 - 704x1408 portrait phone formats up to 3:1 - 1728x576 panoramas, all multiples of 64 that divide cleanly by the VAE's 8× compression. The node takes the ratio, optionally applies overrides and an upscale factor, and creates a zero-filled latent tensor of shape [batch, 4, height/8, width/8] on the right device. A blank latent is a starting condition: the sampler denoises from it, which is exactly what you want for pure text-to-image from an Empty-Latent-style node.

The inputs that matter

• ratio - the dropdown with the 29 presets; default 13:19 - 832x1216 (a nice portrait). This is the one you'll set 95% of the time.
• width_override / height_override - both default 0 (= off). Set them to force specific pixel dimensions instead of the ratio's; values snap to multiples of 8.
• switch_width_height - swaps the width and height of the chosen ratio (portrait ↔ landscape) without hunting for the matching preset.
• upscale_factor - multiplies the resolution, default 1. Think of it as a pre-emptive hires step: 1.5 on 832x1216 gives you a bigger latent so a high-res pass can decode finer detail.
• batch_size - how many blank latents to stack (default 1).

Six outputs: LATENT (the canvas), plus width, height, upscale_factor, upscale_width, and upscale_height - the resolved numbers, handy for feeding a scaling node or a text label so the graph shows you exactly what's being generated.

Installing

Standard NYJY pack install: ComfyUI Manager → Install via Git URL → https://github.com/aidenli/ComfyUI_NYJY, restart. No model downloads - this node only creates tensors.

Common issues & notes

Ratios are SDXL-family. The list is built for SDXL-style models. If you're generating with Flux or Z-Image you don't use this node's preset set the same way - those have their own preferred resolutions (this pack ships a Qwen-latent variant for Qwen-Image, and Flux workflows usually pair with a Flux-oriented empty latent). Also worth knowing: a much larger latent than the model was trained on can produce empty or duplicated regions in SDXL - the overrides and upscale factor are there to keep you sane, not to encourage pushing a 2048-wide latent through a 1024-trained model and expecting magic. Use the ratio defaults and you'll rarely go wrong.