Qwen VL Empty Latent

Every text-to-image graph starts with an empty latent - the blank canvas the sampler fills with noise and then denoises into a picture. QwenVLEmptyLatent is that node, sized correctly for the Qwen-Image family. The one detail that matters: it makes a 16-channel latent, not the old 4-channel SD kind, because that's what Qwen's VAE works in. Use the wrong channel count and your sampler errors out or produces mush, so this exists to make "blank canvas for Qwen" a one-click thing.

Why 16 channels? The modern generation of models - Qwen-Image, Flux, SD3, Wan - moved to a richer VAE latent space with 16 channels instead of the 4 that SD 1.5 and SDXL used. More channels means the latent carries more information per pixel, which is part of why these models render fine detail and text so much better. But it also means the empty latent you feed them has to match. An EmptyLatentImage from the SD era gives you 4 channels and simply won't line up with a 20B Qwen-Image transformer.

How it works

You set the resolution and batch size, and it allocates a zeroed 16-channel latent tensor at the right dimensions for the sampler to add noise to. There's no model and no VAE involved - it's just creating an empty tensor of the correct shape. Straightforward, which is exactly what you want from a canvas node.

The inputs and outputs that matter

• width / height - your target resolution. Qwen-Image works well around the 1MP-and-up range; keep dimensions to multiples of the model's expected step and you'll avoid alignment headaches.
• batch_size (default 1) - how many images to generate at once. Bump it to make a grid in one run, mindful of VRAM.

Output is a single LATENT. Wire it into the latent_image input of your KSampler.

How to install it

ComfyUI Manager → ComfyUI-QwenImageWanBridge → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/fblissjr/ComfyUI-QwenImageWanBridge

then restart. Nothing to download for the node - it's pure tensor allocation. Your Qwen-Image model and its VAE are the parts that cost VRAM.

Common issues & troubleshooting

Channel-count or shape errors from the sampler. That's the whole problem this node prevents, so if you're seeing it, make sure this is the empty-latent node in the graph and not a leftover SD-era EmptyLatentImage (4 channels). They look identical on the canvas and produce incompatible latents.

Do I even need this? Honestly, sometimes not. Native ComfyUI ships EmptySD3LatentImage, which also produces a 16-channel latent and works for Qwen-Image; this node is the pack's own equivalent, kept so a Qwen workflow can be self-contained within the pack's node set. The author frames the whole repo as complementary to native nodes rather than a replacement, and this is a clean example - use whichever keeps your graph tidy.

Odd dimensions give artifacts. Stick to sensible, aligned resolutions. Very unusual aspect ratios or sizes that don't divide cleanly can produce edge artifacts regardless of which empty-latent node you used.