Cloud Latent Upscale Model Loader

Cloud Latent Upscale Model Loader is the node that picks which latent upscaler your cloud chain will use. It presents a fixed dropdown of latent upscale models - the Hunyuan Video 15 latent upsamplers (1080p and 720p), the LTX 2 spatial and temporal upscalers, and the LTX 2.3 spatial upscalers - and returns a CLOUD_LATENT_UPSCALE_MODEL handle. The actual model weights live on Comfy Cloud; nothing downloads to your machine.

The "latent" part matters and is worth being honest about. A latent upscaler works in VAE space, on the compressed latent, before the VAE decode turns it back into pixels. That makes it dramatically cheaper and faster than a pixel-space upscaler - which is exactly why video workflows lean on it: upscaling a 14B video model's latent is far more tractable than upscaling every decoded frame. The trade is the ceiling on detail. Latent upscaling adds resolution, but it can't conjure fine detail the latent didn't contain - if you need maximum sharpness, a pixel upscaler after decode is the heavier, higher-ceiling path. For the workflows this pack targets - Hunyuan Video and LTX chains - the latent upscaler is the right tool, and the README's example workflows use it that way.

The inputs and outputs

• model_name - one dropdown, six choices:

  • hunyuanvideo15_latent_upsampler_1080p.safetensors
  • hunyuanvideo15_latent_upsampler_720p.safetensors
  • ltx-2-spatial-upscaler-x2-1.0.safetensors
  • ltx-2-temporal-upscaler-x2-1.0.safetensors
  • ltx-2.3-spatial-upscaler-x2-1.0.safetensors
  • ltx-2.3-spatial-upscaler-x2-1.1.safetensors

  Pick the family that matches your chain (Hunyuan Video for the Hunyuan models, LTX 2/2.3 for the Lightricks ones). "Spatial" upscales within a frame; "temporal" upscales frame count - they do different jobs, so check which your workflow needs.

Output: LATENT_UPSCALE_MODEL (CLOUD_LATENT_UPSCALE_MODEL) - feeds the latent upsampler node in your chain (the pack's Cloud LTXV Latent Upsampler is the sibling that consumes it).

How it works

Standard pack mechanics: it appends a LatentUpscaleModelLoader spec to the accumulating workflow JSON and returns a handle. The model is loaded on Comfy Cloud when the chain runs - you're declaring which model, not shipping any weights. One quirk in the source is worth knowing: the dropdown is hardcoded from the cloud's /api/object_info, with a comment reminding the author to refresh it as new upscalers are added. A brand-new cloud upscaler won't appear here until the pack updates.

How to install

cd ComfyUI/custom_nodes
git clone https://github.com/Dobidop/ComfyUI-CloudAPI-worker

Copy config.json.example to config.json, paste an API key from https://platform.comfy.org/profile/api-keys, restart. Or install via ComfyUI Manager. Dependencies: requests, Pillow, safetensors.

Common issues

• Wrong family in the dropdown - a Hunyuan upscaler in an LTX chain is a shape mismatch waiting to happen. Match family and (spatial vs temporal) to your workflow.
• "The upscaler I want isn't there" - it's hardcoded; a newer cloud model won't appear until the pack's list is refreshed.
• Expecting pixel-space detail - latent upscaling is the cheap path with a detail ceiling. If you need more, plan a post-decode pixel upscaler pass instead.
• PoC expectations - video upscaling is one of the pack's newer, less-battle-tested corners; the author's "very hacked together" disclaimer applies, so start from the example workflows.