Introduction

Your H3 clips look lovely at 544p. Now you can take them all the way to 4K without renting a datacenter. Tessera, from developer Hasanain Rashid (hashu786), is a new ComfyUI custom node pack that uses MiniMax H3 as a tiled video refiner. It upscales any clip to 2K, 4K or 8K on one consumer GPU, at any clip length.

It's open source under MIT, uses open weights only, needs no extra Python dependencies, and went live on GitHub on October 30 as an entry in Comfy's #OpenSourceHiggsfield challenge.

What shipped

  • Tool: Tessera, 9 ComfyUI nodes plus an example graph (example_workflows/tessera_upscaler.json)
  • Vendor: independent developer hashu786 (Hasanain Rashid)
  • Modality: video upscaling and refinement (video-to-video)
  • Open vs API: fully open, MIT code on top of open H3 weights. No cloud calls
  • Where to run it today: your local ComfyUI. Clone it into custom_nodes and load the example workflow
  • Concrete fact: streamed VAE encode and decode. The author says the full-resolution pixel clip never exists in memory, and that this is how one card handles 544p → 4K

How it works

Tessera first upscales your clip with a neutral Lanczos resample, then has H3 refine it tile by tile and chunk by chunk:

  • Per-tile conditioning: each tile gets its own crop of the video reference, image reference, guide, and (optionally) a re-encoded text embedding. Every tile anchors to the source, not to its neighbours, so drift doesn't compound across a row.
  • One shared noise field: overlapping seam strips are generated by both neighbouring tiles from the same noise, then crossfaded, so you don't get ghosting.
  • Flat-tile skip: featureless tiles like sky, smoke or seamless backdrops keep the plate instead of being resampled. That way the model can't hallucinate into them, and you get the steps back (flat_skip defaults to 4.0).
  • Crash insurance: if the final decode fails, say from a full disk, the refined latent is saved as tessera_rescue_*.latent. You can recover the render without resampling.

The 9 nodes

  • Tessera Streamed Encode: source to upscaled plate latent plus guide conditioning, in temporal windows. It replaces the upscale and encode chain that runs out of memory at 4x
  • Tessera Refine: the tiled refiner itself, with hero tiles and anti-hallucination tile skip
  • Tessera Streamed Decode: standalone streamed decoder and the recovery path for rescued latents
  • Tessera Tile Plan: a dry run showing tile grid, chunk layout and cost before you spend GPU time
  • Tessera Guide Anchor: splices a detailed still into the guide's fine band
  • Tessera Temporal Dilate / Decimate: per-span frame-rate dilation for fast action, re-timed exactly on the way out
  • Tessera Guide Grain: matched grain for the guide
  • Tessera Extra Options: expert knobs, with validated defaults

Install

cd ComfyUI/custom_nodes
git clone https://github.com/hashu786/ComfyUI-Tessera

On startup the console should print [Tessera] loaded 9 node(s). Models you'll want locally:

  • MiniMax H3 ref2va plus the H3 video and audio VAEs and the Qwen3-VL text encoder (the core)
  • LMS detail LoRA by Alissonerdx (part of the "house recipe" for sharpness)
  • Qwen Image 2.1 (optional, for a first-frame detail anchor) and SAM3 (optional, for a hero-subject mask)

Settings that matter

  • Pick high_vram or low_vram for spatial_tile and temporal_tile. Tiles and windows scale together to fit your budget.
  • At 2x, feed a detailed still into the anchor for real detail relay, and set motion_release_floor to 0.3.
  • At 4x and above, leave the anchor unplugged and let the model render.
  • Run Tile Plan first to preview the grid and cost before you commit a long render.

The defaults come from measured renders. The repo's docs/LAWS.md writes up the failures next to the fixes, including why chunk windows snap to H3's token timing and why empty tiles hallucinate. It's a great read if you love seeing an engineering notebook kept in public.

Why it matters

Upscaling has been the quiet bottleneck of local video. Generators render small and fast, but delivery wants 4K. Tessera is a carefully measured, MIT-licensed bridge that keeps everything on your own GPU, and it pairs nicely with the native H3 latent upscaler and Sol-H3-VAE support that just landed in ComfyUI core.

Original Source

Want fresh H3 clips to feed into it? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=comfyui-tessera-minimax-h3-video-upscaler-4k-8k

Go make something crisp!

—Aurelia ♡