ComfyUI Node
Mpi Latent Upscale
Upscale a latent to a target given in PIXELS, dividing by the model's own VAE stride instead of core's hardcoded 8. Handles the packed audio+video latents (MiniMax H3) that crash core's Upscale Latent, scaling only the video half and passing the audio half through. For a hi-res fix: small stage 1, upscale here, finish stage 2 at the target.
Mpi Latent Upscale
- samples
- LATENT
◄upscale_methodnearest-exact►
◄width1344►
◄height768►
◄stride16►
CategoryMpiNodes/Latent
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| samples | LATENT | — | |
| upscale_method | COMBO | nearest-exact | 5 options: nearest-exact, bilinear, area, bicubic, bislerp |
| width | INT | 134416–16384 | Target width in PIXELS, not latent cells. Unlike core's node this is the number you actually want. |
| height | INT | 76816–16384 | Target height in PIXELS. |
| stride | INT | 161–64 | Pixels per latent cell for THIS model's VAE. 16 for MiniMax H3 and Krea2; 8 for the SD/SDXL family. Core hardcodes 8, which is why its node doubles an H3 target. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| LATENT | LATENT | — |