Latent Upsampler
Spatially upscales a latent tensor without leaving latent space — typically used between two Generate stages for a fast, high-resolution refinement pass.
Category: ModularDiffusion/Processing
TL;DR
- Upsamples in latent space, so you avoid a costly decode → upscale → re-encode round trip.
providerpicks the upsampler family.- Use between two Generate Media Latents nodes: low-res pass → upsample → high-res refinement pass.
Typical workflow position
Generate Media Latents (low-res) → [Latent Upsampler] → Generate Media Latents (refine) → Decode
Node preview

Inputs
| Name | Type | Required | Notes |
|---|---|---|---|
input_latent |
LatentArtifact |
Yes | Latent to upsample. Must be in the pipeline's canonical latent space (any Generate / Encode output qualifies). |
Outputs
| Name | Type | Notes |
|---|---|---|
output_latent |
LatentArtifact |
Spatially-upscaled latent. |
Parameters
| Name | Type | Notes |
|---|---|---|
provider |
choice | Upsampler family (e.g. LTX2). Switching regenerates the model picker. |
upsampler_model |
HF repo picker | Hugging Face repo ID for the upsampler weights. |
Tips & pitfalls
- Latent-space upsamplers are family-specific. An LTX2 upsampler will not produce sensible output for an SDXL latent. Match the upsampler family to the latent's pipeline.
- You almost always want a refinement Generate after upsampling. The upsampler increases resolution but doesn't denoise — pair it with a short follow-up Generate at low strength.
See also
- Generate Media Latents — pair upstream and downstream.
- Decode Media Latent — final step.