AGSoft Empty Latent QwenImage

QwenImage is picky about resolution in a way that catches people by surprise: its VAE works in 112-pixel blocks, so your dimensions need to be multiples of 112. 1008×1008, 1344×1008, 1792×1008 - those are the sweet spots. Set a "normal" size like 1024×1024 and you get a sampler error that doesn't obviously point at your latent node. AGSoft Empty Latent QwenImage bakes that rule in, so the multiples-of-112 problem never reaches you.

It's the Qwen-flavored member of the pack's Empty Latent family, and it's the one with the strictest geometry - which is exactly why it's the most valuable of the four.

How it works

The node builds a 4-channel empty latent whose pixel dimensions are forced to be multiples of 112. size_mode offers the usual trio:

• Preset - a Qwen-specific menu. The square default is Square - 1008x1008 (1:1) (note: not 1024 - 1024 isn't a multiple of 112), with Qwen-appropriate squares, portraits, and landscapes throughout, like 1120×1120, 1344×1008, 1792×1008 (16:9), 1120×1792, 2240×1120.
• Custom - raw width/height, min 112, stepping by 112.
• Megapixels - megapixels + aspect_ratio, rounded to the 112 grid so the result stays valid for the VAE.

There's no separate divisibility input here because there's nothing to choose: 112 is the law, and the node enforces it. batch_size sets how many latents you get.

Outputs mirror the other Empty Latents: latent, width_px/height_px, and width_latent/height_latent. If you're wiring a resolution label into your workflow, those ints come in handy, and they're guaranteed 112-compatible.

Where it sits in a Qwen pipeline

In a Qwen-Image or Qwen-Edit workflow this replaces the generic Empty Latent entirely: pick a preset, feed latent into your sampler's latent_image, done. The preset menu is also a nice cheat sheet for which resolutions Qwen actually supports - worth keeping open even if you'd normally type sizes by hand.

Installing it

Pack standard: ComfyUI Manager → search comfyui-AGSoft → install, restart. Or:

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

No models to fetch, no dependencies beyond PyTorch - it's a pure geometry node.

Common issues

The failure you'll actually hit isn't this node; it's Qwen-Image-Edit being subtly misaligned even at legal resolutions, which the community has documented across edit workflows (the workaround there is crop-and-stitch inpainting, not a bigger latent). Within this node's own scope, the only real trap is grabbing a preset from the generic AGSoft Empty Latent out of habit - 1024×1024 there is not 1024×1024 here, and the sampler will tell you loudly. Keep the model and the latent node in the same family and you're fine.