🦑 Kraken Flux Empty Latent Image

If you've ever wired Flux into the stock Empty Latent Image node and wondered why your img2img chain came out smeared, here's the thing nobody tells beginners: that default node hands you a 4-channel latent, and Flux's autoencoder doesn't work in 4 channels. Flux runs 16 channels at an 8× downscale per side - four times as much information per unit of latent as SDXL. Feed it the wrong shape and your encode/decode cycles get noisy instead of subtly wrong. That mismatch is exactly what this node exists to fix.

🦑 Kraken Flux Empty Latent Image builds the latent the way Flux actually expects and, more usefully, does the resolution math you were doing on a calculator. This is the front of the KrakenTools workflow: pick a megapixel and an aspect ratio, and it returns a latent plus the resolution and preset strings the pack's other two nodes read.

How it works

You pick a megapixel (how much total canvas you want) and a preset (the aspect ratio). It computes width and height so the area matches - a 1.0 MP 2:3 (Classic Portrait) lands at roughly 832×1216 - then rounds both to a model-friendly multiple. The divisible_by field handles that: Auto uses 64 for canvases over 1.5 MP, 32 for over 0.75 MP, 16 below. Flux is finicky about non-multiples-of-8, so this saves you the silent tile-in-your-generation problem.

The latent it returns is [batch, 16, h//8, w//8] - the 16 channels and the 8× factor are Flux's VAE geometry, not a guess. It also renders a little preview image showing the intended canvas size, so you can see the box before you burn a sampling pass.

The inputs that matter

• megapixel - the total canvas budget (0.5 to 3.0, or Custom to type raw W×H).
• preset - 64 choices. Plain ratios (2:3 Classic Portrait, 16:9 Panorama) scale with your megapixel; the Screen: and Video: presets (like iPhone Wallpaper - 1170x2532 or KlingAI Wide - 1344x768) are fixed sizes that ignore the megapixel field.
• divisible_by - leave it on Auto until you know why you need something else.
• batch_size - number of latents, for when you're generating variations.

Skip use_custom_aspect unless you need a ratio no preset covers. You get five outputs, but the two that matter downstream are resolution (a "736 x 1344" string) and preset_out (a "736x1344" string) - those feed KrakenUpscaleTileCalc and KrakenResolutionHelper so the whole chain agrees on the target. latent goes into your sampler as usual, and width_out/height_out are there for any node that wants plain integers.

Installing it

Through ComfyUI Manager, search Kraken Tools (or ComfyUI-KrakenTools). Or by hand:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/krakenunbound/ComfyUI-KrakenTools

Restart ComfyUI. That's it - no requirements.txt, no model downloads, nothing beyond the torch/numpy/Pillow ComfyUI already ships. The whole pack is MIT-licensed and small enough to read in one sitting.

Common gotchas

• Don't expect exact dimensions from the presets. They're aspect-driven and snap to the divisible grid - at 1.0 MP, 2:3 gives you ~832×1216, not 736×1344. That's fine; the pack's design is "close enough for sampling, exact at the end," which is what KrakenResolutionHelper is for.
• The Screen:/Video: presets don't scale with megapixel. If you pick one of those, megapixel is ignored - the listed W×H is what you get.
• This is a young, small pack (first release late 2025, by "The Kraken" on X/YouTube). Don't expect a big community or a wiki - the source is the documentation.

One thing I'll say for it: it's the rare "empty latent" node that thinks in canvas area instead of raw pixels, which is exactly how Flux users actually talk about resolution.