Empty Flux 2 Latent

Empty Flux 2 Latent is the starting block for any FLUX.2 txt2img workflow - the blank tensor the sampler denoises into a picture. Every text-to-image graph that uses a FLUX.2 checkpoint (the 32B Dev, or the Klein 4B/9B distillations) begins with one of these, because unlike SD 1.5/SDXL-style setups there's no VAE encode step at the start; you generate straight from an empty latent.

Mechanically it's a block of zeros shaped [batch_size, 128, height // 16, width // 16] - 128 channels deep, downsampled 16× per side. The channel depth is the interesting bit: that's four times what an SDXL latent carries, which is how FLUX.2 holds fine texture through edits. The three inputs are what you'd expect - width (default 1024, step 16), height (default 1024, step 16), and batch_size (default 1, up to 4096). Set the pixel size of the image you want, and it hands you the right-sized empty latent.

The thing people get wrong

Don't feed a FLUX.2 model the generic Empty SD3 Latent. FLUX.1 also used a 128-channel, 16× latent, so it's tempting to reuse the old node - but Black Forest Labs retrained the latent space from scratch for FLUX.2, with its own VAE released as a separate artifact. The channel count looks the same and the coordinate system isn't. You'll get output, and it'll be subtly wrong in ways that are infuriating to debug. This node exists precisely because the new family needs its own empty latent; use it.

Resolution is the other lever, and it's model-specific in practice. Klein was built to run on consumer cards and handles 1MP comfortably; Dev wants more headroom and more VRAM. The node allows up to 16384 per side, but remember this is a 128-channel tensor - each doubling of resolution multiplies the latent's footprint fourfold, and the sampler will happily OOM your card before you run out of slider. Start at 1024, verify your VRAM budget before going bigger.

What it wires into

Output is a LATENT, which goes straight into the KSampler's latent input (or into Flux2ImageNode-style tooling that takes a blank latent and handles the conditioning side for you). If you're doing image-to-image or editing instead of text-to-image, you won't use this node at all - you'll encode pixels with the FLUX.2 VAE and feed the resulting latent to the sampler, which is the VAEEncode half of the same family.

It ships with ComfyUI core - no install, no model files of its own; the checkpoint and VAE you pair it with are the downloads that matter. If you're new to FLUX.2, this is the node you'll stop thinking about within a day, and that's the sign it's doing its job.