Upscale Image by 1.5x

The name is a little misleading. This node doesn't hand you back a bigger picture - it takes a finished image, runs it back through your VAE's encoder, and scales the resulting latent up by 1.5x. What comes out the other end is a LATENT, not an IMAGE. Think of it less as "resize this photo" and more as "get me set up for a hi-res-fix second pass, in one node instead of three."

That's exactly why it exists. The pack author builds mostly at 16:9, and 1280x720 times 1.5 lands you exactly on 1920x1080 - a clean jump from a first-pass size to a proper 1080p output. Rather than chaining a VAE Encode node into a separate latent-upscale node, this one does both steps at once.

How it works

Feed it a finished image and the VAE that matches the model that made it. The node encodes the image into latent space, scales that latent by a fixed 1.5x in both dimensions, and - this is the part worth knowing about - duplicates the result across a batch of amount copies. That's not blending; it's amount independent copies of the same upscaled latent, so when you run them through a sampler you get amount different variations to compare and pick your favorite from, rather than one deterministic result.

Inputs and outputs that matter

• pixels (IMAGE) - the source image you want to push into a bigger latent.
• vae (VAE) - the VAE used for the encode. It has to match the family of model that made (and will later decode) the image, or you'll get color shifts and noise instead of a clean upscale.
• amount (INT, 1–10, default 1) - how many copies to stack in the output batch.

The single output is LATENT, sized 1.5x up and batched. Wire it straight into a sampler - this pack's own KSampler with VAE node is a natural pairing, since it hands you the decoded image back in the same step. Whatever sampler you use, keep the denoise low-ish (something in the 0.3–0.5 range is the standard hi-res-fix sweet spot) so the second pass adds detail instead of repainting the composition.

Installing it

Through ComfyUI Manager: search mzmaxam in Custom Nodes Manager and install. Or by hand:

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

Restart ComfyUI afterward. There's nothing to download here - no checkpoints, no requirements.txt fight. It's pure node logic wired against types ComfyUI already understands, so install is genuinely just clone-and-restart.

Where people get tripped up

The biggest one: expecting a bigger picture out of it directly. You won't get one - the output is a latent, and if you want to actually see the upscale you need to run it through a sampler and a VAE Decode afterward. Second, this node only does 1.5x; there's no exposed multiplier. If you need 2x, or a specific target resolution, reach for ComfyUI's stock latent upscale node instead - this one is deliberately single-purpose. Third, watch amount: cranking it toward 10 doesn't give you a better result, it gives you a 10-wide batch through your sampler, which multiplies your VRAM use and run time accordingly. And as with any encode/decode step, keep the VAE matched to the checkpoint family - an SDXL VAE on an SD 1.5 latent (or vice versa) is a classic source of muddy, off-color output that has nothing to do with this node and everything to do with a mismatched VAE.