Image Latent Creator 🎨

Every text-to-image workflow starts the same way: you need an empty latent in the right shape before the KSampler will do anything. ComfyUI's built-in Empty Latent Image does that, but it leaves you eyeballing width and height. Image Latent Creator instead hands you a dropdown of SDXL's optimal aspect ratios - the nine buckets like 3:2 at 1216×832 that the model was trained to like - so your canvas starts in the sweet spot, not somewhere a model tolerates.

It's part of the Painting Coder Utils pack, and it builds directly on the pack's Image Size Creator 📏: everything that node does, this does, plus it actually creates the latent and takes a batch size.

How it works

Pick mode (Landscape / Portrait / Square), pick a resolution from the SDXL preset list, and the node computes width and height (with an optional scale_factor multiplier). It then allocates a zero-filled latent tensor of the right shape: for an 1024×1024 canvas, that's [batch_size, 4, 128, 128] - the image is downsampled 8× per side because the VAE compresses pixels into the latent space the diffusion model actually works in. That tensor goes out as a LATENT dictionary ready for the KSampler.

The front-end JavaScript filters the resolution dropdown by mode, so in Landscape you only see the landscape ratios and vice versa - a small touch that keeps you from picking 5:12 while trying to build a wide canvas.

The inputs and outputs that matter

• mode - Landscape, Portrait, or Square.
• resolution - the SDXL preset: 1:1 (1024×1024), 9:7 (1152×896), 7:9, 3:2, 2:3, 7:4, 4:7, 12:5 (1536×640), 5:12.
• scale_factor - multiplies the base resolution, from 0.1 to 10. Leave at 1 unless you know why you're scaling.
• batch_size - 1 to 64. Set this if you're generating a batch in one go.
• Outputs: latent (the empty LATENT for your KSampler), plus resolution, width, height, and batch_size as separate outputs you can wire into nodes that need the numbers (like conditioning or metadata).

Where you'll reach for it

As the cleanest start of an SDXL text-to-image or img2img graph, and anywhere you want the canvas to be one of the ratios the model actually trained on rather than a random size you typed in. The batch-size output also makes it handy for batch workflows where a downstream node needs to know how many images it's handling.

One caution: this creates a zero latent, which is exactly what a fresh generation wants. Don't wire it into something expecting an already-sampled latent - that's a blank canvas, not a result.

Install

ComfyUI Manager → search Painting Coder Utils, or:

cd ComfyUI/custom_nodes
git clone https://github.com/jammyfu/ComfyUI_PaintingCoderUtils.git

Restart ComfyUI. No model downloads; the pack's dependencies (numpy, Pillow, torch, gradio) are already part of any working ComfyUI. If you're resurrecting an old workflow and get missing-node errors, the v0.3.0 namespace change is the culprit - use the included fixer at docs/fix/workflow_fixer.html.