Self-hosting
Open-H3-Regenerate-2K: NVIDIA's Self-Hosted Answer to a Paid API
Verdict: Open-H3-Regenerate-2K is NVIDIA's Efficient-Large-Model team (with MIT) doing what it does best: taking a paid API — MiniMax's closed H3-Regenerate-2K upscaler — and rebuilding it as a self-hostable three-step LoRA pass that turns a 768P H3 draft into 2560×1440 in 22.45 seconds on one B200 (down from 87 seconds with the naive stack). The engineering is real: a rank-576 merged LoRA over 312 linears, a distilled fast VAE decoder (1.62B params) that decodes a 2K clip in 1.6s instead of 15s, and a 3.89× total speedup. The strings: the whole thing rides on MiniMax-H3's community license — fine below $20M yearly revenue, but commercial products must display "MiniMax H3" in the UI, territorial carve-outs apply, and restrictions pass downstream. And the obvious catch: the measured hardware is a B200. If you don't already run datacenter GPUs, the paid API you were trying to replace is still the cheaper option on a per-clip basis.
Last checked 2026-10-31. Model weights and the license can change; the numbers below are the ones we could verify today.
What the release actually is
A single Hugging Face repo (Efficient-Large-Model/Open-H3-Regenerate-2K, created 2026-10-26, 136 likes at check time) containing: the merged refinement LoRA (FP32, rank 576 across 312 linear layers, scale 1), the 50 LMS Sharp curve AdaLN branches, a 3D-conv latent upsampler, the fast non-causal 24-layer VAE decoder distilled from the released H3 VAE, and an unmodified mirror of the MiniMax-H3 base components at revision 6818f6c. Inference code lives in the NVlabs/Sana repo (Apache-2.0, 9.7K stars) under the sol-engine branch.
What it does: regenerate a 768P draft (1344×768, 121 frames, ~5 seconds of video) at 2560×1440. The LoRA refines detail; the latent upsampler inflates the latent representation directly instead of repeatedly encoding and decoding; the Sol-Engine stack (sparse attention, MXFP8, CUDA graphs) does the arithmetic. Chinese coverage also credits the release with native 4K (3840×2160) regeneration — we could not find that claim in the repo README we pulled, so we treat it as unverified. Chinese coverage also positions this inside a broader H3 open ecosystem: FastVideo's FastH3 four-step preview, Pika's H3 Flex six-step sampling, ComfyUI and Unsloth adapter and quantized versions.
The speed table, as measured
One B200, 5-second clip, 768P latent to 2560×1440, warm, median of 3 runs — from the model card, not vendor PR:
| Configuration | Total | Speedup |
|---|---|---|
| Standard forward (dense attention, BF16) + released H3 VAE | 87.26 s | 1.00× |
| Sol-Attn, MXFP8, CUDA graphs + released H3 VAE | 37.39 s | 2.33× |
| Sol-Attn, MXFP8, CUDA graphs + fast H3 VAE decoder | 22.45 s | 3.89× |
The distilled VAE decoder alone takes 2K decode from 15 s to 1.6 s; latent upscale plus refine goes from 73.11 s to 20.46 s. Whether 22 seconds per 5-second clip is "fast" depends entirely on whether you're an interactive tool (yes) or a batch pipeline (multiply it out: 2 minutes of source footage ≈ 9 minutes of GPU time on a card you probably rent).
The license you actually inherit
The repo's own license field is other — specifically minimax-h3-community, the MiniMax H3 base model's community agreement. Because the LoRA is a derivative of H3, deploying it means accepting MiniMax's terms, and they are not Apache:
- $20M yearly revenue trigger. Above that, you need prior written authorization from MiniMax (api@minimax.io) before commercial use. Below it, you're fine — this is the most generous of this month's custom licenses (OXR cuts in at $2M total funding or revenue; Hi3D at $1M monthly).
- Attribution duty. Commercial products or services using H3 or its works must prominently display "MiniMax H3" in the user interface. Your white-label video enhancer can't stay white-label.
- Territorial restrictions. The grant applies only within the "Applicable Territory"; excluded territories exist and require separate licensing. Check the agreement's definitions against your deployment regions before you ship.
- Downstream binding. Before providing access to anyone else, you must bind each user to terms at least as protective as the use restrictions — the AUP travels with the weights.
Outputs themselves are not the issue here — the strings govern the model and derived models, and 20 million dollars is a ceiling most video shops will never hit. The point is smaller and simpler: this is an open release in the MiniMax sense, not the OSI sense, and the terms are enforceable by the licensor of the base model your refinement pass depends on.
Who should run this
If you're already generating H3 drafts on your own GPUs and paying MiniMax per clip for the official 2K regeneration API, this release replaces that line item — the LoRA approach keeps your prompts, your drafts and your 2K output on your own hardware. If you're below the $20M line and outside excluded territories, the license is a formality plus a UI badge.
If you don't have B200-class hardware, the decision is less friendly: the speed table has no consumer-GPU row, and nothing in the repo is validated on one. The paid API's price is the number to beat, and renting a B200 by the hour to save it only works at volume. Our honest read: brilliant research artifact and a genuine price ceiling on the closed API — not yet a weekend Docker pull for the 4090 crowd.
Related on this site
- OXR-1.0's license strings — the stricter cousin: $2M threshold and a competing-API ban
- MiniMax M3.1 Flash review — the other end of the MiniMax stack, where the tokens are cheap
Sources: Hugging Face Efficient-Large-Model/Open-H3-Regenerate-2K README (license fields, speed table, file manifest, base revision mirror — pulled 2026-10-31); MiniMax H3 community LICENSE full text via MiniMaxAI/MiniMax-H3 (revenue trigger §IV.1, attribution §IV.2, territorial grant §II, downstream binding §V.2 — pulled 2026-10-31); GitHub API for NVlabs/Sana (Apache-2.0, 9,722 stars, active through 2026-10-30); QbitAI 2026-10-30 feature (4K claim, ecosystem roster, B200 context — transcribed, 4K not found in official README). Hardware cost comparisons and the API-substitution verdict are our own math, dated accordingly.