AGSoft Empty Latent Krea2

Krea 2 is one of those newer architectures that doesn't play by SD rules, and the most visible difference is right where you start: its latent has 16 channels, not the 4 that SD/SDXL/Flux use. Drop a stock Empty Latent into a Krea 2 graph and your sampler will choke on a channel mismatch - the error reads like an obscure tensor-shape complaint, and it takes a while to realize the empty latent you've used a thousand times is the problem. AGSoft Empty Latent Krea2 exists so you start on the right foot.

This is the Krea-flavored member of the pack's Empty Latent family, and if you run Krea 2 locally, it's basically a prerequisite.

How it works

The node generates torch.zeros([batch_size, 16, latent_h, latent_w]) - the 16 channels Krea 2's VAE expects. Beyond the channel count, it's the same friendly shape as the other Empty Latents: three size_mode options.

• Preset - Krea-oriented named sizes (Square 1024×1024 default, plus portrait and landscape sets like 768×1344, 1080×1920, 1216×832, 1920×1080).
• Custom - raw width / height (default 1024).
• Megapixels - megapixels (default 1.0) plus an aspect_ratio dropdown (1:1, 3:2, 2:3, 4:3, 3:4, 16:9, 9:16, 21:9, 9:21), rounded to the divisibility you choose - the dropdown offers 8/16/32/64/128, defaulting to 64.

Outputs are the family standard: latent, width_px/height_px, and width_latent/height_latent. Those last two are nice for logging, and if you're building Krea 2 workflows you'll appreciate having the numbers as ints rather than doing the mental math.

The input that matters most

Honestly, the divisibility default of 64 covers most Krea 2 use. The one setting worth respecting is batch_size if you're doing multi-image generation - the node builds N identical 16-channel latents in one call, which is exactly what a batch sampler wants. Everything else is pick-a-preset.

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

No models, no extra dependencies, pure PyTorch - one of the lightest nodes in a self-contained pack.

Common issues

The classic mistake is the one this node prevents: using the generic Empty Latent (or an old workflow's latent) with Krea 2 and getting a shape error deep in the sampler. If that error ever mentions channel counts, check your latent node before your model. Conversely, don't use this node with non-Krea models - a 16-channel latent is meaningless to an SDXL or Flux checkpoint. Match the node to the model family; the pack's four Empty Latents exist precisely so you never have to guess.

A lighter gotcha: the pack's node is MIT, but Krea 2's model weights run under the Krea-2-Community-License - free under $1M in annual revenue, but it pushes a content-filter obligation onto you and bans stripping safety measures or watermarks. Don't confuse the node's license with the model's; read Krea's terms before shipping anything from it.