FunUI Load Latent Upscaler
The Latent Upscaler Loader for Files ComfyUI's Own Node Won't Touch
- upscaler
A one-input, one-output node that does nothing visible. Two things make it worth its own page: it reads a model folder ComfyUI's own loader also reads but builds the upscaler a different way, and it's the only way to feed FunUI Latent Resample. If you're doing a two-pass video workflow in FunUI, this is the top of that chain.
What it actually does
It loads a .safetensors from ComfyUI/models/latent_upscale_models - the same folder the stock "Load Latent Upscale Model" node uses - and then hands the state dict to whichever installed model module says it can build one. That's the interesting part: the pack is built as modules that announce themselves, so the loader doesn't hardcode "this is an H3 upscaler". It asks around, and the first module that recognises the weights builds the object.
If nobody recognises them, you get a real error rather than a node that "loads" and then does something quietly wrong. And the two most likely failures have specific, useful messages:
- Weights whose keys start with
base.are called out as the newer attention-based H3 upscaler, which FunUI cannot build yet, with the advice to use an older convolution-based H3 upscaler. Useful, because otherwise you'd assume the file was corrupt. - Lightricks' or Hunyuan's upscalers are told to use ComfyUI's own loader instead, since core knows those and this node doesn't add anything.
Inputs and outputs
- upscaler_name - required, a combo of the files in
latent_upscale_models. If the folder is empty or doesn't exist yet, the list comes back empty and the node still registers - you'll see an empty dropdown, which is ComfyUI's way of saying "no files", not a broken install. - upscaler - a
LATENT_UPSCALE_MODEL. Wire it into FunUI Latent Resample'supscalerinput. That's its only sane destination.
Note there's no status output here, unlike most of the pack's loaders. Nothing to read, nothing to check - the success signal is that you didn't get a red node.
Installing it
Same pack as everything else on this list:
cd ComfyUI/custom_nodes
git clone https://github.com/digital-garbage/FunUI
cd FunUI && pip install -r requirements.txt
ComfyUI Manager users: search FunUI. Restart ComfyUI, then put the upscaler file in ComfyUI/models/latent_upscale_models/ - create the folder if it isn't there. (Yes, the pack page says olivv-cs/ComfyUI-FunPack; the project renamed itself to FunUI and lives at digital-garbage/FunUI now.)
There's nothing extra to install for this node specifically. requirements.txt covers comfy-kitchen and gguf; the upscaler weights are yours to download, and they live with the model they belong to, so get the H3 one from wherever you got H3.
Where people get burned
Putting upscalers in models/upscale_models. That folder is for pixel upscalers (ESRGAN, 4x-UltraSharp and friends) used after VAE decode. Latent upscalers are a different animal and a different directory. Cross the wires and you'll get an empty dropdown in one node and a load error in the other.
Grabbing the newest H3 upscaler. Attention-based variants are exactly what this pack can't build yet. If the error mentions base. keys, you have the right diagnosis: find the convolution-based one instead.
Expecting it to work with core's resample nodes. Its output is a FunUI latent-upscale object; the pack's own resample node knows how to run it. A stock node expecting ComfyUI's LATENT_UPSCALE_MODEL dialect may reject it or vice versa - the pack is explicit that core's loader is the right tool for Lightricks/Hunyuan weights.
Thinking it upscales anything on its own. It doesn't touch a latent. It's a file picker with opinions. FunUI Latent Resample is where the work happens, and that's where the "sharpen vs upscale" decision - and the much bigger VRAM bill - actually lives.
Inputs (1)
| Name | Type | Default | Description |
|---|---|---|---|
| upscaler_name | COMBO | 0 options: |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| upscaler | LATENT_UPSCALE_MODEL | — |