MiniMax H3 cuts sampling to 4 steps with two open releases

MiniMax H3 samples video and audio in 50 denoising steps. This week two separate open releases cut that to four or eight, which is the difference between minutes per clip on a single GPU and a render that finishes while you wait. The awkward part is choosing between them, because the two releases are different kinds of thing with different installs, different hardware needs and the same license.

FastH3 Omni landed on October 20, 2026, an 8-step distilled student of MiniMax H3 for reference-to-video with audio, built by FastVideo at UC San Diego's Hao AI Lab with Nuva Lab. The same week, the taobao-content-ai account published per-hop PTD LoRAs that take the same task down to 8 or 4 steps. Neither release has published independent quality benchmarks.

What changed, and why 50 steps was the problem

Base MiniMax H3 is a 33B joint audio-video model. Its 50-step schedule produces good motion and natural audio, and it is slow: on FastVideo's own numbers a single 15-second reference clip took a base model 139.69 seconds across eight GB200 GPUs, before any queueing. Few-step distillation trains a smaller student to imitate that long trajectory in a handful of passes, which is what makes local and cheap runs practical.

FastH3 Omni is the reference-conditioned branch of that work. It keeps base H3's text encoder, video VAE and audio VAE and replaces only the diffusion transformer with a student that runs 8 fused evaluations at guidance 1.0, with a video shift of 12.0 and an audio shift of 3.0. The PTD LoRAs take a different route, described below, and reach 4 steps.

What each release actually is

FastH3 Omni is one distilled checkpoint. You download a single diffusion model plus shared components and you are done. It takes a prompt and ordered references: up to 9 images (up to 3 in realtime mode), up to 3 reference videos of 2 to 15 seconds each for 15 seconds total, and up to 3 voice or music references, capped at 12 reference files. At least one image or video is required. It also accepts first and last frame keyframes and supports clip continuation. Output is 5, 10 or 15 seconds at 24 fps (124, 243 or 362 frames) with synchronized 32 kHz stereo audio in one pass. The release includes training code.

The PTD LoRAs are a pair of adapters, not a model. PTD, short for Parametric Trajectory Distillation, is the method in a paper by Lan Feng and co-authors from EPFL, NVIDIA and Stanford, submitted to arXiv on October 8, 2026. It trains one independent low-rank adapter per inference step on the diffusion backbone, so each denoising step selects its own adapter and fuses its own output head. The released Ref2VA files are an 8-step adapter (10.3 GiB) and a 4-step adapter (5.2 GiB), both rank 128 in BF16. A third checkpoint trained with PTD plus RefAsymOPD, aimed at better backgrounds and sharpness at the same step count, followed on October 22. The adapters are the only thing PTD adds, so you still need the full base model underneath.

FastH3 Omni versus the PTD LoRAs

On speed, the two releases are not measured on the same footing. FastVideo reports that FastH3 Omni runs at 5.6 to 11.9 times lower end-to-end latency than MiniMax H3 across 27 matched reference workloads on 8 NVIDIA GB200 GPUs. That is FastVideo's own measurement, comparing complete serving recipes rather than an isolated kernel change, and no outside lab has repeated it. FastVideo also states that its UniServe stack reaches real-time generation at 768p for 5, 10 and 15 second clips on particular tasks. The PTD card publishes no comparable latency table; its claim is the step count, and the paper reports its own blinded human preference shares of 55.1% against PDD and 63.4% against LightX2V Turbo. Those are the authors' measurements, not an outside test.

On install weight, FastH3 Omni is the lighter lift. The PTD adapters require the original, unpruned Ref2VA transformer; pruned, Turbo, PDD and quantized bases are not valid for them, and the full set of weights comes to roughly 140 GB.

How to run FastH3 Omni in ComfyUI

1. Update ComfyUI past its latest tagged release. The newest tag is v0.42.1, published October 19, 2026, and native MiniMax H3 support landed after it, so a tagged build will not load the model. artrealmai reported that ComfyUI confirmed native support about an hour after the release.
2. Download one diffusion model from the FastH3-Omni-Comfy repo and place it in ComfyUI/models/diffusion_models/. The choices are BF16 at 41.2 GiB, INT8 ConvRot at 20.7 GiB (the template default), FP8 at 21.5 GiB from the FP8 repo, or NVFP4 QAD at 12.8 GiB from the NVFP4 QAD repo.
3. Add one Qwen3-VL 32B text encoder (BF16 48.0 GiB, INT8 25.3 GiB, or NVFP4 AWQ 14.6 GiB), one video VAE (FP16 4.85 GiB or INT8 2.62 GiB) and the audio VAE (0.56 GiB), which you need for sound. The NVFP4 encoder requires Blackwell hardware.
4. Load the video_fasth3_ominiref_r2v template, which uses the core MiniMax H3 Reference to Video node. Write the prompt with positional tags (<Picture 1>, <Video 1>, <Audio 1>), leave 8 steps and guidance 1.0 in place, and generate.

How to run the PTD LoRAs in ComfyUI

1. Convert the base yourself. These adapters need the original, unpruned transformer_ref/ from MiniMaxAI/MiniMax-H3, converted with scripts/convert_comfy_base.py. Pruned, Turbo, PDD and quantized bases are explicitly not valid.
2. Install the upstream ComfyUI-PTD on the release/comfyui-inference branch into custom_nodes/ and run pip install -e on it. A bare clone fails at import, because its ptd_comfy module imports a sibling ptd package.
3. Copy the repo's ComfyUI-MiniMaxH3-PTD pack into custom_nodes/, put the adapter checkpoint in ComfyUI/models/ptd/, and start ComfyUI with --disable-dynamic-vram.
4. Build the graph: Load Diffusion Model, Load MiniMax-H3 PTD Adapters, PTD Sampler and PTD Sigmas into SamplerCustomAdvanced, conditioned through MiniMax H3 Reference to Video, then separate the AV latent and decode video and audio. Use the Basic Guider with no CFG, because the student was distilled to run without it. The checkpoint's ptd_num_hops metadata sets the step count, so there is no step slider. Work at 1344 x 768 and 124 frames, the resolution PTD was trained and validated on.
5. Start from the API-format example workflow in the repo's examples/ directory, and copy its reference images from asset/ into ComfyUI/input/ first.

A lighter community path exists. The compact loader, created October 22, accepts pruned and quantized bases and adds int8 versions of the adapters, including a 2.86 GB 4-step Ref2VA file and a 5.68 GB 8-step file. Its author reports that end-to-end renders on an RTX 5090 with pruned int8 bases change the video by about as much as rounding alone does.

What neither release gives you

Neither release has published independent quality benchmarks, and no outside lab has scored either one. FastVideo's own release post describes its matched clips as qualitative comparisons and states plainly that they are not evidence of output parity with base H3. The PTD paper's preference shares are the authors' own blinded votes, not an independent evaluation. If you need a number that says the 4-step output is good enough, there is not one yet; render a test clip on your own references before you commit.

Who should pick which

Pick FastH3 Omni if you want image, video and audio references in one pass and the shortest path to a running graph: one diffusion model, shared components, the official template, and no base-model conversion. On a consumer Blackwell card, start with the 12.8 GiB NVFP4 QAD build; on Ada or older, use FP8 or INT8.

Pick the PTD LoRAs if you want the fastest sampling, 4 steps, and you already have the unpruned H3 base or the disk for roughly 140 GB of weights. On smaller setups, the community int8 conversions are the way in.

Either way, read the license before you ship anything. Both releases carry the MiniMax-H3 Community License Agreement, whose Applicable Territory is worldwide excluding the European Union, the United Kingdom, the Republic of Korea and the United States, and the grant of rights is limited to that territory; MiniMax says people in the excluded territories can contact it about obtaining a license. Commercial products above 20 million US dollars in yearly revenue also need separate written authorization from MiniMax.

If you have a Blackwell card, download the 12.8 GiB NVFP4 QAD FastH3 Omni diffusion model with its shared text encoder and VAEs, update ComfyUI past v0.42.1, and load the video_fasth3_ominiref_r2v template; try that before committing roughly 140 GB of disk to the PTD path.