Tessera upscales video to 2K, 4K or 8K in ComfyUI with MiniMax H3 on one consumer GPU
Tessera is a free, open-source ComfyUI node pack that upscales video with the open MiniMax H3 model, refining it tile by tile and chunk by chunk into 2K, 4K or 8K. Its streamed encode and decode mean the full-resolution clip never sits in memory, so it runs on a single consumer GPU at any clip length.
What it does
Load a clip, set a scale factor and run the included tessera_upscaler.json workflow; the author says the defaults are the validated recipe. Nine nodes do the work: Streamed Encode replaces the upscale and encode chain that runs out of memory at 4x, Refine is the tiled refiner with temporal chunking, hero tiles and a skip for featureless tiles such as sky or smoke (so the model can't hallucinate into them), Tile Plan is a dry run that shows the tile grid and cost before you spend GPU time, and others add a first-frame detail anchor, frame-rate dilation for fast action and matched grain. If the final decode fails, for example on a full disk, the refined latent is saved so it can be decoded again without resampling.
Worth knowing
Install by cloning it into ComfyUI/custom_nodes; there are no extra dependencies. It uses MiniMax H3 ref2va with its video and audio VAEs and Qwen3-VL text encoder, plus Alissonerdx's LMS detail LoRA, with Qwen Image 2.1 and SAM3 optional. Pick high_vram or low_vram tiling to fit your card. The author advises feeding a detailed still as an anchor at 2x and leaving the anchor off at 4x and above. MIT licence; MiniMax H3 keeps its own community licence. It is an entry in Comfy's Open Source Higgsfield challenge.