Flux 2 Empty Latent Image MXD

Flux 2 - BFL's 56B flagship that the community bounced off for its VRAM appetite and censorship - is a different animal from Flux.1 under the hood, and one of the first places that bites is the latent. Flux.1 uses the familiar 16-channel / 8x-compressed latent that SD3-style models use. Flux 2's VAE encodes at 128 channels with 16x spatial compression, so a 1024x1024 image becomes a 64x64 latent instead of 128x128. Flux 2 Empty Latent Image MXD just makes sure you create latents with the right shape, so you don't have to think about it.

What it is

It's a resolution-preset empty latent generator, in the same family as the pack's Flux and SDXL versions: pick a resolution from the dropdown, flip vertical if you want it portrait, set a batch_size, and you get a LATENT out. The one output feeds straight into a sampler.

The resolution list is the same 18-entry Flux set as the other MaxedOut latent nodes - High (1408² up to 2176x960), Standard (the 1024² class), Low (down to 576x256). All standard bucket shapes: Square, 4:3, 3:2, 16:9, 21:9.

How it works

Mechanically it's dead simple - it allocates a zeros tensor at [batch, 128, h//16, w//16] on ComfyUI's intermediate device and wraps it as {"samples": ...}. The vertical toggle just swaps width and height before the division. Nothing gets encoded; it's empty noise-initiation space, exactly like the core Empty Latent Image node you're used to.

That 16x compression is the important part: get the shape wrong and your sampler either errors or (worse) silently produces garbage that only shows up after a long wait at high res.

Inputs and outputs, briefly

• resolution - the dropdown. Default is Square (1:1) 1024x1024, which is the sensible 1MP starting point.
• vertical - swaps the dimensions for portrait. It's just a swap, not a separate resolution list, which keeps the preset list short.
• batch_size - 1 to 4096. Leave at 1 unless you know you're batch-generating.
• Output: LATENT - the empty Flux 2 latent batch, wired to your KSampler's latent input.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Maxed-Out-99/ComfyUI-MaxedOut.git

Restart ComfyUI, or use Manager → search "Maxed Out". No extra dependencies; the pack's other latent nodes come along for free.

Troubleshooting

• "… is a header or invalid option" - the dropdown includes section headers (- High Resolutions -) that aren't real sizes. Pick an actual preset.
• Wrong latent shape errors from the sampler - you're probably not using this node. If you feed a Flux.1-shaped latent into a Flux 2 model, the sampler will complain about channel count. This node is the fix.
• OOM at the "High" presets - Flux 2 is heavy. The 1408²+ buckets are for cards with real VRAM; stay in Standard for a 1MP-ish footprint unless you know your hardware.

Flux 2 demands a lot of everything, but the resolution/latent step shouldn't be where you lose. This node makes that one step a non-event.