⭐ Star Model Latent Upscaler

You know the drill: you generate at native resolution, then want it bigger, and the stock way is to throw a KSampler at the latent again. That's a hires-fix, and it's fine, but it's also a round trip through pixel space that the Star Model Latent Upscaler (class Starupscale) skips. This node runs an actual upscale model, resizes to a target size, then encodes straight back into latent space with a VAE you choose - all in one box. If you're doing a second diffusion pass at higher resolution, this is the thing sitting between the two samplers.

It's part of the ComfyUI_StarNodes pack (the ⭐StarNodes collection from Starnodes2024), which you'll find under ⭐StarNodes/Image And Latent after install.

How it works

Three paths, picked automatically from whatever you plug in:

• IMAGE in - it upscales the image directly.
• VAE_INPUT + LATENT_INPUT in - it decodes the latent to pixels, upscales, then re-encodes. This is the hires-fix path.
• Nothing in - it just hands you a configured VAE and a tiny default image, which is a way to reuse the node as a VAE provider.

Whatever path you're on, the flow is: optionally run UPSCALE_MODEL (any ESRGAN-family model in your models/upscale_models folder, or "Default" for plain resize), then resize so the longest side lands on OUTPUT_LONGEST_SIDE, keeping aspect ratio and rounding dimensions to multiples of 16 so the VAE doesn't choke. Finally it encodes to latent with the VAE you picked. You get three outputs: OUTPUT VAE, IMAGE, and LATENT - wire the latent into your second sampler.

The inputs that matter

Honestly only three of them:

• VAE_OUT - which VAE does the encoding. The dropdown lists your installed VAEs and defaults to ae.safetensors, which is the Flux VAE. If you're upscaling SDXL, switch this to your SDXL VAE; using the wrong one here is the classic way to get weird colors on the second pass.
• UPSCALE_MODEL - an ESRGAN upscaler like 4x-UltraSharp if you want the model to add texture, or Default if you just want a clean resize. This is the "more pixels vs more detail" fork; for most hires-fix work a sharp ESRGAN model is the right call.
• OUTPUT_LONGEST_SIDE - the target, in pixels. 2048 is a fine default for SDXL second passes.

VAE_Device (cpu vs cuda) and INTERPOLATION_MODE (bicubic is right 95% of the time) you can leave alone.

Installing it

It ships inside the StarNodes pack, so install once for all of them:

• ComfyUI Manager: search Starnodes and install.
• Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes

Then restart ComfyUI. The pack's requirements.txt pulls opencv-python, psd-tools, scikit-learn and friends automatically - none of which this node needs, so don't panic if the installer churns through a few.

Common issues

The big one is VAE mismatch - Flux VAE on an SDXL upscale gives you that sickly green/brown cast on the second pass. Fix is in the VAE_OUT dropdown, not the model. Second: if your output dimensions aren't coming out exactly as typed, that's the /16 rounding doing its job; it's intentional and correct for latent space. And remember the upscale model only adds plausible detail - it can't invent texture that was never there, so don't expect miracles from a 512px source at 4x.

The pack's other upscaler, Star SD Upscale Refiner, is the all-in-one alternative if you want checkpoint+LoRA+ControlNet Tile in one node. This one's the leaner, more composable pick.