Imagen latente Pro

Every diffusion workflow that generates from pure noise needs a blank latent to start from - that's what ComfyUI's built-in EmptyLatentImage is for. This node is that, with the sharp edges smoothed off. It's a Spanish-labeled utility from the pack's "Selectores Pro" set that gives you a big list of preset resolutions, a batch size, and a rounding option, and hands you a ready-to-sampler LATENT. No API, no keys, nothing paid - it's pure tensor math.

The name means "Pro latent image," and the "Pro" is the preset list. Instead of typing 512×768 by hand and hoping it's a valid size for the sampler, you pick from a curated menu that spans quick-test sizes (256-class), medium (512-class), large (1024-class), and two social formats (720×1280 9:16 and 1280×720 16:9). Each preset is named like 512×512 (1:1) - Medio so you can see the aspect at a glance.

How it works

Internally it builds a zero tensor of shape [batch_size, 4, H/8, W/8] - the four-channel latent space at one-eighth resolution, same as any empty-latent node. The interesting knob is rounding:

• auto_round (default): nudges the width/height up to the nearest multiple of 8 before dividing by 8, so a preset that isn't a clean latent multiple still works.
• strict: refuses to round - if the resolution isn't already a multiple of 8, it raises an error. Use this when you need an exact pixel size and would rather fail loudly than silently get 520px instead of 512px.

The inputs that matter

• size_preset - the resolution menu (the - Test, - Medio, - Grande, - Social suffixes hint at intended use).
• batch_size - 1 to 64, how many empty latents to stack for batch generation.
• rounding - auto_round or strict, described above.

Output: a single latent (LATENT) to feed straight into a KSampler or its modern equivalents.

Installing it

ComfyUI Manager → search "COMFYUI_PROMPTMODELS" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/cdanielp/COMFYUI_PROMPTMODELS

No API key. Requires ComfyUI >= 0.26.0.

The honest take

If you're happy with the stock EmptyLatentImage, this is a lateral move - but the preset list is genuinely nicer for batch experiments, and strict mode is the detail that earns it a place in a graph. One thing to remember: an empty latent is a starting point, not a resolution guarantee. The model's own constraints still apply, and if you feed a strict-sized latent to a sampler that expects a specific base resolution, the model does its thing regardless. When in doubt, use auto_round and let the sampler's preferred dimensions win.