NVIDIA's Open-H3-Regenerate-2K: a three-step 2K pass for MiniMax H3

What it is. Open-H3-Regenerate-2K is a LoRA for MiniMax H3 plus a pipeline around it: you generate a draft at 768p, upsample the latent, and the LoRA regenerates it at 2560×1440 in three Euler steps, with the project page showing runs up to 4K. It isn't a classic upscaler. It re-draws the clip at higher resolution, which is why faces and hands in wide shots come back with real detail instead of sharpened blur. The release also ships a sharpening LoRA and a fast H3 VAE, and the code lives in NVlabs' Sana repo on the sol-engine branch.

The speed claims, per the authors. Their announcement says a 5-second 2K clip takes 22.45 seconds on one B200 against 112.7 seconds through the official 2K API, about 5× faster. The model card's own table is a little more modest: 87.26 seconds down to 22.45 (3.89×) once Sol-Attn MXFP8 attention and the fast VAE are switched on. The authors and @wildmindai's roundup both say it beats the closed 2K API in user preference tests. Those are NVIDIA's numbers on data-centre hardware, not ours.

Running it in ComfyUI. rdh9616's ComfyUI-Open-H3-Regenerate pack (Apache-2.0, created Oct 28) adds loader, latent-upscale and decode nodes, saves the refined latent before decoding so you can re-decode without re-running, and includes a Video→H3 Prompt node. Per @_ITRay's post, a 5-second clip reaches 2K in about three minutes on an RTX PRO 5000, with refine time dropping from 474 to 122 seconds once Sol-Attn is on. The full LoRA is a 12 GB FP32 file. Kijai's extracted versions run from a 155 MB rank-16 file to a roughly 5 to 6 GB rank-384 bf16 one, and in Benji's walkthrough the rank-384 LoRA in a normal two-stage H3 workflow (latent upscale ×3, three steps, sharpening LoRA on) looked better than the official example workflow, finishing in about a minute on his machine. It works with fine-tuned H3 bases and other LoRAs, per that video.

The caveats. Olli Sorjonen (@Olmirad) found it behaves more like img2img than an upscaler, re-imagining detail rather than preserving it, and said it runs in roughly 24 to 32 GB of VRAM. After a second look he called it useful for some jobs and released a small wrapper with noise and memory controls, noting that faces, text and small details can change at lower noise settings. So it's a poor fit for shots where a face or a logo must stay exact. Licensing is mixed: the LoRA follows the MiniMax H3 Community License, other components carry their own terms, and earlier H3 projects have flagged that the H3 licence excludes some regions without a separate agreement. Check before client work.

Our take: if you already draft in H3 locally, this is the cheapest route to 2K finals we've seen: generate small, keep the takes you like, then spend three steps on them. Start with Kijai's rank-384 LoRA, not the 12 GB file. For identity-critical shots, compare it with a faithful upscaler like SeedVR2 first.