Upscale Latent by 1.5x

This is the lean sibling of Upscale Image by 1.5x. Same math, same job - bump a latent up 1.5x for a second, higher-resolution sampling pass - but it skips the VAE step entirely. If you already have a latent in hand (straight out of a first KSampler pass, say), this one takes it directly, no encode/decode detour required. That makes it faster and it sidesteps any VAE-induced softness, since you never leave latent space.

The pack author built this specifically around 16:9 work: their go-to first-pass size is 1280x720, and 1280x720 times 1.5 lands exactly on 1920x1080. That's not a coincidence, it's the reason the node exists - a clean, purpose-built jump from a first pass to a real 1080p output, with no odd rounding.

How it works

You hand it a latent and a batch count. The node scales the latent's spatial dimensions by a fixed 1.5x and duplicates the result amount times across the batch - not to average or blend, but so you get amount independent copies to sample and compare, and keep the one you actually like.

Worth knowing going in: a straight latent-space resize is an interpolation, not new detail. It'll look a touch soft coming out the other side. That's fine - it's exactly what a second sampling pass at low denoise is for, and the pack's fixed 1.5x factor happens to sit right inside the community's usual hires-fix range of 1.5–2x.

Inputs and outputs that matter

• samples (LATENT) - the latent to upscale. This has to actually be a latent; if all you have is a finished image, use the sibling Upscale Image by 1.5x node instead, which does the VAE encode for you.
• amount (INT, 1–10, default 1) - how many copies to stack in the output batch.

The single output is LATENT, 1.5x bigger, batched. Feed it straight into a second sampling pass - this pack's KSampler with VAE is a convenient pairing since it decodes for you in the same node - at a low-to-moderate denoise (0.3–0.5 is the standard range) so the pass adds detail rather than repainting the image.

Installing it

ComfyUI Manager: search mzmaxam in Custom Nodes Manager and install. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/MzMaXaM/ComfyUi-MzMaXaM

Restart ComfyUI. No model downloads, no extra Python dependencies - it's pure node logic against a type ComfyUI already speaks, so the install is genuinely just clone-and-restart.

Where people get tripped up

The most common mix-up is plugging an IMAGE into samples - it wants a LATENT, full stop. If you're coming from a finished picture rather than an in-progress latent, that's what the sibling Upscale Image by 1.5x node is for. Beyond that, the usual batch-size gotcha applies: amount isn't a quality knob, it's a batch multiplier, so pushing it toward 10 costs you 10x the VRAM and sampling time for 10 independent variants, not one better image. And because this node only ever does 1.5x, if your target size doesn't land on a clean multiple of your input, you'll want ComfyUI's stock latent-upscale node with a manual scale factor instead - this one is deliberately fixed and single-purpose.