📥 LatentUpscaleModel Download

Quick clarification first, because "upscale" is one of the most overloaded words in this space: this node grabs a latent upscale model, not a pixel one. The distinction matters. A regular ESRGAN-style upscaler (4x-UltraSharp and friends) works on the finished image - more pixels, runs in milliseconds, invents nothing. A latent upscaler works earlier, on the compressed latent representation before the VAE decodes it, and it's part of a different family of models used by newer architectures like the LTX video upscalers. Wrong tool for the wrong job is the single most common upscaling mistake, so know which one you're loading.

This is a download helper, plain and simple. Point it at a URL, it fetches the latent upscale model and hands you a LATENT_UPSCALE_MODEL output that plugs into whatever node consumes one. The value is skipping the manual download-and-place step and being able to pin the model source inside a shared workflow.

How it works

Same machinery as the rest of SDVN's download nodes: it pulls the file from a direct link (Hugging Face or CivitAI direct URLs), using aria2c for a fast, resumable download when it's installed, saves it to the appropriate models folder under the name you specify, and loads it into memory. The output is a ready-to-use LATENT_UPSCALE_MODEL - the loading is done, you just wire it downstream.

The inputs that matter

Two, and they mirror every SDVN downloader:

• Download_url - the direct file link. The actual download URL, not the model's web page.
• Url_name - the filename to save it as (default model.safetensors). Name it something you'll recognize; latent upscalers are model-specific, so ltx-latent-upscaler.safetensors beats model.safetensors when you've got a folder full of them.

The output is LATENT_UPSCALE_MODEL, which you feed into the matching latent-upscale node in your pipeline.

Where it fits

This is a niche node - you reach for it when you're running a stack that specifically wants a latent upscale model, LTX video pipelines being the obvious one, where a latent upscaler bumps resolution before decode as part of the multiscale render. If you're doing ordinary image upscaling (make this SDXL render bigger and sharper), you want a pixel upscale model instead, and SDVN has a separate download node for those. Don't cross the streams.

Common issues

The output won't connect to my upscale node. LATENT_UPSCALE_MODEL is a specific type. It only links to nodes built to consume a latent upscaler - you can't feed it into a standard ImageUpscaleWithModel node, which expects a pixel-space upscale model. Match the node to the model type.

Download fails on Windows or macOS. aria2c isn't installed by default there. The README is explicit: install aria2c manually on those systems or the auto-download nodes won't run.

I loaded it but the result looks wrong. Latent upscalers are tuned to specific model families. Using one built for a different architecture than the one you're generating with produces artifacts or garbage. Confirm the upscaler matches your base model.

Installing SDVN_Comfy_node

ComfyUI Manager: search "SDVN_Comfy_node" → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/StableDiffusionVN/SDVN_Comfy_node

Then pip install -r custom_nodes/SDVN_Comfy_node/requirements.txt from the ComfyUI root and restart. Install aria2c first if you're on Windows or macOS. Find the node under 📂 SDVN / 📥 Download.