Latent Size Picker 🐲

An empty latent is where every text-to-image run starts, and ComfyUI's core EmptyLatentImage node is fine at that job - except it doesn't help you pick a good resolution, and it doesn't snap to a multiple. Latent Size Picker fixes both: it's an empty-latent generator with a dropdown of 32 predefined ~1MP resolutions tuned for SDXL, plus the pack's scale-to-multiple math bolted on. Pick a shape, pick a multiple, get a latent and the numbers to match.

What it does

It builds an empty latent tensor (all zeros) and hands you three outputs: latent, width, and height. The resolution comes from a dropdown of 32 precomputed ~1MP sizes covering the full aspect-ratio spread around 1024x1024 - from 1728x576 (3:1, ultra-wide) through 1024x1024 (1:1, the default) down to 576x1728 (1:3, ultra-tall). These are the standard SDXL resolutions people actually generate at, which saves you from inventing a weird aspect ratio that a checkpoint wasn't trained on. Generating far off native resolution is a classic source of tiling and duplicated anatomy, per the troubleshooting docs, so starting from a sane ~1MP size is genuinely good hygiene.

On top of that base resolution it applies the pack's scaling logic:

• multiple - snap to a multiple (default 64, step 8). This is the killer feature: EmptyLatentImage will happily give you 1026x1024; this node won't.
• rounding_mode - nearest / floor / ceil.
• scale_factor - prescale before snapping.
• width_override / height_override - set to a nonzero value to override that axis entirely (0 means "use the resolution").
• batch_size - how many empty latents to stack (default 1).

The latent itself is created as [batch, 4, height/8, width/8] - the 8x spatial compression of the VAE - so the values you see are real latent-space dims, not pixels.

The nice UI touch

This is one of the few nodes in the pack with custom frontend code: after a run, it draws the current width and height in blue text right on the node body, next to the output sockets. So you can see at a glance what resolution the last run produced without squinting at a KSampler. Small thing, genuinely useful.

The inputs that matter

• resolution - the dropdown. Default 1024x1024 (1:1). Pick your aspect ratio here.
• multiple - default 64. Set to 8 if your model wants 8-multiples (SDXL happily takes 64; some pipelines want 8).
• batch_size - stack count if you're generating multiple at once.
• width_override / height_override - your escape hatch for non-preset sizes (0 = off).

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/DraconicDragon/ComfyUI-RyuuNoodles

Restart ComfyUI, or install via Manager (search "RyuuNoodles"). No models, no extra dependencies.

Gotchas

The dimension math floors to multiples of 8 internally (min 8x8), so an extreme override might not land exactly where you typed - that's the latent-compatibility safety net, not a bug. And the code credits cubiq's ComfyUI_essentials Empty Latent Size Picker (MIT) as the original - this is a fork with the multiple-snapping added, so if you already use cubiq's, the behavior will feel familiar, just with snapping and overrides. One more: it's meant as the starting latent for txt2img. If you're doing img2img, you want the VAE-encoded latent, not this.