Taco Latent Image

The Taco Latent Image node is what happens when someone gets tired of typing 1024x1024 and 1344x768 into Empty Latent Image by hand. Instead of exact pixel dimensions, you pick an aspect ratio from a dropdown, pick a scale, and it hands you a ready-to-sample LATENT. That's the whole job, and honestly it's a good enough job that it's worth knowing even though it's buried in a five-node pack.

It lives under Taco_Nodes in the node menu and is a drop-in replacement for the core Empty Latent Image. Wire its LATENT output into your KSampler's latent input exactly like you would the built-in one.

How it works

The mechanism is about as simple as it gets: it creates a tensor of zeros in latent space, which is exactly what Empty Latent Image does - latents that haven't been through the sampler yet are all zeros, and sampling adds the noise. The dimensions are computed from your two choices:

• aspect_ratio - one of 1:1, 3:4, 4:3, 9:16, 16:9. Each maps to a base resolution: 1:1 → 512×512, 3:4 → 416×576, 4:3 → 576×416, 9:16 → 384×672, 16:9 → 672×384.
• resolution_ratio - small (512), medium (1024), or large (2048). These are multipliers (1×, 2×, 4×) on the base resolution, not absolute sizes. So 1:1 + medium gives you 1024×1024, while 9:16 + medium gives 768×1344. The label says "1024" because that's what the square ratio lands on - a portrait or landscape ratio will exceed it on the long side.
• batch_size - default 1, up to 64. Raise it when you want a batch of identical latents, which is the standard trick for generating variations.

The latent is sized at one-eighth the pixel resolution on each side, since that's the VAE's compression factor - 512×512 becomes a 64×64 latent, 1024 becomes 128×128, and so on.

Which one to reach for

Remember that every model has a native resolution it was trained on. For SD 1.5 that's 512×512; SDXL is 1024×1024. The author's own framing maps nicely onto that: small is "great for small latents and initial images for hires-fix workflows," medium is "good for standard renders and works well for SDXL 1.0," and large comes with the warning "I don't recommend using this unless you know what you are doing." Take that warning seriously - generating straight at 2048×2048 is past what most checkpoints are comfortable with, and it's the classic recipe for duplicated anatomy and tiled patterns. If you want big output, generate at native res and upscale or use a hires-fix pass instead.

Installing it

It ships in ComfyUI-TacoNodes. Easiest route is ComfyUI Manager: search "ComfyUI-TacoNodes" (or "Taco") and install, then restart. Or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/YOUR-WORST-TACO/ComfyUI-TacoNodes.git

Restart ComfyUI and you'll find it under Taco_Nodes. No model downloads, and there's no requirements.txt - the pack only uses PIL, numpy and torch, all of which ship with ComfyUI already. It's a small personal pack (Apache 2.0, single maintainer) that hasn't been touched in a while, so if it ever misbehaves, the GitHub repo is where you'd look for answers.

One honest caveat: this node gives you zero features that Empty Latent Image doesn't - it's a pure convenience wrapper. Its superpower is that it stops you from misremembering exact pixel dimensions mid-workflow, which is exactly the kind of papercut this node was built to fix. If you never fumble those numbers, you don't need it. If you do, it's surprisingly easy to get attached to.