AGSoft Empty Latent Flux.2

FLUX.2 latents don't follow the SD rules, and that trips everyone up once. With SD/SDXL, the latent is 1/8 of the image on each side. FLUX.2's VAE is twice as aggressive: the latent is 1/16 - a 1024×1024 image becomes a 64×64 latent, not 128×128. Feed it an SD-sized latent and you'll get garbage or a crash, whichever comes first. AGSoft Empty Latent Flux2 exists so you never think about the ÷16 yourself.

This is the Flux-flavored sibling of the pack's generic AGSoft Empty Latent: same structure, FLUX.2's geometry baked in. If you run Flux 2 models in this era of ComfyUI, you'll probably want both in your graph at some point.

How it works

Same three size_mode choices as the base node - Preset, Custom, Megapixels - but the latent it emits divides by 16 and enforces the constraints FLUX.2 needs. In Custom mode the minimum width/height is 16 (because the latent would otherwise be 1 pixel or less), and dimensions step in multiples of 8. Megapixels mode combines megapixels with an aspect_ratio (1:1, 3:2, 2:3, 4:3, 3:4, 16:9, 9:16, 21:9, 9:21) and rounds to the chosen divisibility - 64 by default, and the dropdown also offers 112 if you're ever sharing a workflow with QwenImage nodes.

Outputs are the useful quartet: latent (the tensor), width_px/height_px (image resolution), and width_latent/height_latent (the much smaller latent size - the numbers that look "wrong" if you're used to SD, and the reason this node exists). batch_size covers multi-image generation.

The inputs in practice

For 95% of use you'll set size_mode to Preset and grab a named resolution, or leave it on Megapixels at 1.0 and pick an aspect ratio - which is the friendliest path, since you never have to know whether 1344×768 is a supported Flux size. divisibility only matters when you're mixing this node into a pipeline whose other stages have their own constraints.

Installing it

Same as every node in the pack: ComfyUI Manager → search comfyui-AGSoft → install, restart. Or manually:

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

No downloads, no model files, no dependencies beyond PyTorch. It's a tiny node in a self-contained pack.

Common issues

The realistic failure mode is model confusion, not node failure: if you use this latent with an SDXL checkpoint, or the base Empty Latent with FLUX.2, the channel/layout mismatch will produce broken images - and the error message may not say why. The pack deliberately ships separate latent nodes per architecture, so the rule is simple: match the node to the model. If a shared Flux 2 workflow keeps failing at the sampler with a "divisible by" error, check whether someone replaced this node with the generic one along the way - it's a subtle break that looks like a model problem.