Meux Smart Empty Latent

ComfyUI's built-in Empty Latent Image has a nasty habit: you type 1000×600, it silently creates a latent that isn't what you asked for, and you don't notice until the sampler or VAE complains. The built-in rounds to safe multiples, sure, but the size in your head isn't the size that comes out, and SDXL-class models increasingly want multiples of 64. MeuxSmartEmptyLatent is the version that tells you the truth: it rounds your target up to a safe value, builds the latent, and hands you the actual generated size on the side.

How it works

You give it a target size and an align (8 or 64). It rounds each dimension up - ceil(target / align) * align - so the generated latent is always at least as big as you asked, never smaller. It then creates a zero latent tensor at the VAE's actual resolution: each side divided by 8 (the standard downsampling for SD-family models), with your batch_size and 4 channels.

Inputs:

• target_width / target_height - what you want, even awkward numbers like 1000×600.
• batch_size - how many empty latents to create in one tensor.
• align - 8 (bare minimum for a latent) or 64 (the default, and the safe choice for SDXL-class models).

Outputs:

• latent - the empty LATENT, ready for the sampler.
• gen_width / gen_height - the actual generated pixel size, rounded up. Wire these into a text display or a downstream resize node so your sizing matches reality instead of your intention.

Installing it

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/fchangjun/Comfyui_MultiSaveImage
cd Comfyui_MultiSaveImage
pip install -r requirements.txt

Or ComfyUI Manager → search "Baidu Meux ComfyTools" → restart.

Why it earns the "smart"

The rounding-up detail is the whole point. A latent that rounds down is a latent that crops your composition; rounding up means the image comes out slightly bigger, which you can then crop or resize exactly. And the gen_width / gen_height outputs close the loop - you never have to compute "wait, what size is this actually?" on a calculator again. If you're doing img2img or outpaint where a downstream node needs the real dimensions, those two INT outputs save you a whole custom node.

Honest caveat: if you always work in clean multiples of 64 anyway, this buys you nothing - the built-in node is fine. But the moment you work with odd target sizes, or you need to know the real dimension downstream, this is the version you want in the graph.