Select Z-Image Dimensions

Z-Image (Alibaba Tongyi-MAI's 6.15B model that quietly became the local photorealism default) has a resolution story most people haven't memorized yet. SDXL's native set is burned into everyone's brain; Z-Image's recommended dimensions come in three tiers - 1024, 1280, and 1536 - and the 1280 tier especially is where it outruns the base resolution. Select Z-Image Dimensions packages all 33 official presets into one dropdown and emits a ready-made 16-channel latent to go with them.

What it actually does

The dropdown holds 33 presets: 11 at 1024 (~1MP), 11 at 1280 (~1.6MP), and 11 at 1536 (~2.4MP). Each tier covers the full ratio sweep - Square, Landscape/Portrait, Classic 4:3, Photo 3:2, 16:9 Widescreen/Vertical, and 21:9 Ultrawide/Ultra-Tall - and the pack's description notes the set covers both Z-Image Base and the Turbo variant. That tier structure lines up with what the community found: Z-Image's native ceiling is roughly 1MP to 2MP, with 1024×1024 the baseline and the 1280/1536 tiers being where people actually push it. The 1536 tier sits right at the edge of that window, which is exactly where you want to be before artifacts creep in.

Pick a preset and the node emits width, height, and a latent built to match. The latent is 16-channel at 8× spatial compression - the same latent space Flux uses, not SDXL's 4-channel one - so this is the node to reach for instead of EmptyLatentImage when you're on Z-Image. Wire the latent straight into the sampler; no channel mismatch, no hand-computed dimensions.

The inputs and outputs that matter

• dimensions - the dropdown. Three tiers, 11 ratios each; that's the whole node.
• upscale_factor - default 2.0, 1.0–8.0 in 0.25 steps. Just arithmetic.
• batch_size - default 1; raise it for a pre-batched latent.

Outputs: width, height, latent, plus upscale factor, upscaled width, and upscaled height. That last trio is the hires-fix bridge: Z-Image's ceiling is ~2MP, so anything bigger should come from a second pass. Wire upscaled width/upscaled height into a latent upscale node and the 1536 preset becomes a two-step path to genuinely high-res output without generating straight into artifact territory.

Installing it

Part of ComfyUI-LatentPresets, the same small pack that covers SDXL and Flux. No extra Python dependencies, no model downloads - it's pure UI convenience:

cd ComfyUI/custom_nodes
git clone https://github.com/bnwa/ComfyUI-LatentPresets

Then restart ComfyUI. In ComfyUI Manager choose Install via Git URL and paste that address; the pack isn't registered in the Manager's searchable registry yet.

Where people get burned

Two things. First, the version floor: this pack is built on the comfy_api.latest v3 node API and requires ComfyUI 0.3.x or newer. On an older build the node simply doesn't exist in the menu. Update ComfyUI before you suspect the pack.

Second, don't assume the 1536 "2K" tier is a free upgrade just because it's in an official list. Z-Image degrades past ~2MP without a second pass, so treat 1536 as your one-shot ceiling, not a starting point - and remember that on Z-Image Turbo, generation speed and VRAM are cheap, so the 1280 tier is often the smarter daily driver. For a node that's basically a dropdown, that's a lot of bad-typo prevention per install.