Aspect Ratio Master

Every model has a native resolution, and the newer ones are picky about more than megapixels: Flux wants 16 latent channels, Mochi wants 12, SANA wants 32, LTX Video wants 128 - and if the empty latent you feed a sampler has the wrong channel count, the VAE decode at the end crashes or spits out a grey mess. Aspect Ratio Master is the node that keeps you from having to remember any of that. It picks a resolution, and hands you an empty latent already shaped for your exact architecture. That second part is the whole point.

It's from the 3R3BOS Pack, a small solo pack. This node is genuinely useful, though it has a bit of a stormy past: v1.0.8's changelog is essentially a confession that latent channels were hardcoded to 4 for every architecture, which broke VAE decoding for Flux, Wan, SD3 and friends. That's fixed now - but it's a good reminder to actually update the pack rather than assume you're on the latest.

How it works

Under the hood it's a small lookup table and a bit of arithmetic. Each architecture gets three numbers: a round-to multiple (mod 8/16/32/64 so dims stay VAE-friendly), a VAE compression factor (how much the latent is downsampled per side), and the latent channel count. The quality mode picks a megapixel budget from a preset table, the aspect ratio splits that budget into width and height, and then everything is snapped to the rounding multiple.

The clever output is the empty latent itself:

torch.zeros([batch_size, latent_channels, h // vae_factor, w // vae_factor])

That's the exact shape the sampler expects, generated for you - no separate Empty Latent Image node, no "wait, which divisible-by number does this model want" guessing. A small preview image is rendered too, so you see your dimensions and a visual of the frame before you queue anything.

Inputs and outputs

The ones you'll actually set:

• model_arch - the picker with ~24 entries: Flux, Wan Video (V2.x/V2.5), NVIDIA Cosmos, SANA-1.5, Hunyuan Video/Image, LTX Video, SDXL, Pony, Illustrious/NoobAI, and more. This drives everything, so get it right.
• aspect_ratio - 13 presets from 1:1 to 21:9 and 1:2.
• performance_mode - Low / Middle / High Quality, which just shifts the megapixel budget up and down.

The extras: use_custom_ratio flips on CUSTOM mode with custom_width/custom_height integer fields (auto-snapped to safe multiples), use_batch + batch_size make a multi-frame latent for video or batch work, and custom_scale_factor (0.5–2.0) scales the whole budget up or down.

Outputs: width and height as ints, empty_latent (LATENT) to feed the KSampler, and preview (IMAGE) so you can eyeball the frame.

Installing it

Part of the pack - no model files, no extra Python dependencies. Via Manager, search 3R3BOS Pack, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/3R3BOS/ComfyUI-3R3BOS-Pack

then restart ComfyUI. comfy node install 3r3bos-pack works too.

Where people get burned

Three things. First, update the pack - the channel-count bug I mentioned above caused real crashes, and if you were seeing grey outputs or VAE decode errors on a modern model, that fix is probably yours. Second, trust the architecture presets over custom dims; if you go custom, remember the node locks minimums at 380x680 so you can't accidentally generate a latent so small it's useless. Third, this is a 2026-era calculator for a fast-moving model zoo - the preset table is the author's best guess at architectures that change quarterly. For a wildly new model, the safe habit is still to check the model card before trusting any calculator, this one included.