AGSoft Empty Latent

ComfyUI's default Empty Latent node works fine, but it makes you think in latent-space math: pick width and height, and you have to remember that the actual image is 8× larger and that dimensions need to be multiples of 8. AGSoft Empty Latent is the same idea with the arithmetic moved inside. You tell it image dimensions, presets, or megapixels, and it hands back a correctly-shaped latent - plus the pixel and latent sizes as numbers you can wire elsewhere.

If you've ever hit "latent size not divisible by 64" from a sampler, this is the node that quietly prevents that class of error before it happens.

How it works

The node generates a standard 4-channel empty latent - torch.zeros([batch, 4, latent_h, latent_w]) - where the latent dims are your chosen pixel size divided by 8, the standard SD/SDXL VAE downscale factor. Three size_mode options decide how you specify the size:

• Preset - a long dropdown of named resolutions (Square 512→1920, Portrait, Landscape). The full SDXL-flavored menu: 1024×1024, 768×1344, 1344×768, 1152×832, and friends.
• Custom - raw width / height.
• Megapixels - megapixels (1.0 = ~1000×1000) combined with an aspect_ratio dropdown (1:1, 3:2, 2:3, 16:9, 21:9…). The node computes a W×H that lands near your target resolution and rounds to the chosen divisibility.

divisibility defaults to 64 - the SDXL-friendly number - with options from 8 to 128 (the tooltip notes 112 for QwenImage). batch_size makes N identical latents in one go, which is how you drive batch generation from a single node.

The outputs you'll actually use

latent goes straight into your KSampler's latent_image. The rest are informational but genuinely handy: width_px / height_px (the real image size) and width_latent / height_latent (the latent size). When you're building workflows that display or log the resolution, those ints save you a lot of awkward math, and they're the same values a resize node outputs - so wiring them into a text label is painless.

Installing it

Standard pack install: ComfyUI Manager → search comfyui-AGSoft → install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Art-xmaster/comfyui-AGSoft.git

Pure PyTorch, no models, no extra dependencies. It's one of the lighter nodes in the pack.

Common issues

Honestly, there isn't much to break here. The main gotcha is not this node's fault: presets assume SD/SDXL conventions, so if you grab "Portrait - 768x1344" for a Flux 2 or Qwen workflow, the latent channels may not match what that model expects - this is the SD-flavored Empty Latent, and the pack ships Flux.2, QwenImage, and Krea2 variants for those models. Pick the matching one.

The other trap: megapixels mode with a very tall aspect ratio can give you a resolution your VRAM doesn't love. That's a VRAM budget problem, not a node bug - but it's worth knowing the node will cheerfully produce a 9:21 latent at whatever size you ask for.