MiniMax H3 · Acceleration
MiniMax H3 LoRA: the 4-step turbo LoRA, and training your own
Two different things get called a MiniMax H3 LoRA. One is the turbo LoRA: it renders joint video and synchronized stereo audio in 4 sampling steps instead of the usual ~20 — roughly 5× less sampling wall-clock. The other is training your own MiniMax H3 LoRA on a character, product or style. This page covers both, starting with the speed one.
The turbo LoRA’s reputation problem is a single symptom: the audio comes back distorted at 4 steps and everyone assumes the MiniMax H3 LoRA is broken. It isn’t; the sampler is.
This is the first thing to fix, because it is the reason most people abandon the MiniMax H3 LoRA within an hour of installing it.
The 4-step audio artifact
- Symptom
- Video looks fine at 4 steps, but the audio comes back distorted or blown out.
- Cause
- H3 generates a video stream and an audio stream together. ComfyUI's stock samplers advance both on a single shared schedule — harmless at ~20 steps, but at 4 steps it massively over-steps the audio branch.
- Fix
- Advance each stream on its own schedule instead of one shared one. Current ComfyUI does this natively; older builds need the MiniMax H3 Dual-Clock Euler node. Either way, audio stays clean at 4 steps rather than forcing you back to 20.
Audio is the acknowledged weak point of this early checkpoint even with the right sampler. If audio fidelity matters more than speed, generate without the turbo LoRA.
Source: larryvrh/MiniMax-H3-Turbo-Lora discussions · shuaixn/ComfyUI-MiniMaxH3DualClockSampler
Installing the MiniMax H3 LoRA into the official workflow
Adding the MiniMax H3 LoRA to the stock graph takes two changes, and that is the whole integration.
- 01
Install the custom nodes, through ComfyUI-Manager or by cloning the node repo into ComfyUI/custom_nodes.
- 02
Put one .safetensors from the LoRA repo into ComfyUI/models/loras/.
- 03
Open the official MiniMax H3 text-to-video or image-to-video template and insert the MiniMax-H3 Turbo LoRA node between the model loader and the sampler.
- 04
Replace the sampler feeding SamplerCustomAdvanced with the MiniMax-H3 Turbo Sampler (4-step), and set the scheduler to 4 steps on simple.
Nothing else in the official workflow changes, so both text-to-video and image-to-video keep working. A ready-made turbo workflow ships in the LoRA repo if you would rather not edit the graph yourself.
- Node repo
- Larryvrh/ComfyUI-MiniMax-H3-Turbo
- In ComfyUI-Manager
- Search "MiniMax-H3 Turbo" in ComfyUI-Manager
- LoRA folder
- ComfyUI/models/loras/
- Ready-made workflow
- minimax_h3_t2v_turbo.json
Source: larryvrh/MiniMax-H3-Turbo-Lora README
MiniMax H3 LoRA settings that actually matter
- Sampling steps
4- With ckpt850 the publisher states 4 steps is already sharp; it was the earlier checkpoints that needed 6 to 8 to firm up. Any count at or above 4 is valid and more steps still help slightly.
- Scheduler
simple- Keep the scheduler on simple at 4 steps. On a current ComfyUI a stock euler sampler also works, because the audio-video schedule split is handled natively.
- LoRA strength
1.0, usable 0.8–1.2- This is a dial, not a constant — the publisher documents which direction to move it for which artifact. See the strength note below.
- low_vram switch
Off by default- Off applies the LoRA at run time: sharpest result, some extra peak VRAM. On merges it into the weights for the lowest peak VRAM, at the cost of a softer result on quantized and pruned bases.
- Base checkpoint
Any MiniMax H3 base- The publisher states the LoRA works with the full base as well as the pruned and curve variants, and detects a pruned base automatically. You do not need to re-download an unpruned checkpoint for it.
- Sampler node
Dual-clock Euler — or a current ComfyUI- ComfyUI gained native audio-video sampling support (ModelSamplingAV) in the August 6, 2026 build. On anything older you need the MiniMax H3 Dual-Clock Euler node from ComfyUI-MiniMaxH3DualClockSampler, which exists mainly for that backward compatibility.
Strength is a symptom dial, not a quality slider
- Blurry ghosting or smear
- Nudge strength up, around 1.05 to 1.2.
- Over-sharp grain or crunchy artifacts
- Nudge strength down, around 0.8 to 0.95.
Both directions are documented by the publisher, which makes this the one setting worth changing before you start swapping checkpoints.
Source: larryvrh/MiniMax-H3-Turbo-Lora README
Which MiniMax H3 LoRA checkpoint to download
Six MiniMax H3 LoRA files circulate and the names are nearly identical, which is why people end up on a superseded one. The training step count is the distinguishing detail, and it has moved twice already.
| File | Trained | Notes |
|---|
| minimax_h3_turbo_4step_ema_ckpt850.safetensorsstart here | ~850, EMA | Publisher's recommendation and the current final checkpoint of this training round. Start here. |
| minimax_h3_turbo_4step_ckpt850.safetensors | ~850 | Same round without time-averaging — the publisher describes it as even sharper but over-sharpened, and offers it for comparison rather than production. |
| minimax_h3_turbo_4step_ema_ckpt500.safetensors | ~500, EMA | Previous recommendation, superseded by ckpt850. Softer at 4 steps. |
| minimax_h3_turbo_4step_ckpt500.safetensors | ~500 | Older non-EMA, softer. Superseded by ckpt850. |
| minimax_h3_turbo_4step.safetensors | ~200 | First public weights, superseded twice over. |
| minimax_h3_turbo_4step_ema.safetensors | ~200, EMA | Initial EMA release, superseded. |
Every file in the repo is bf16 and roughly 744 MB, applied as a standard low-rank update with alpha equal to rank, so there is no extra scaling factor to set. The six names differ only by training step count and whether the weights are time-averaged.
Source: larryvrh/MiniMax-H3-Turbo-Lora · drbaph repack
The publisher has paused training on this checkpoint
Detail and sharpness at 4 steps are high, but the current weights show plastic-looking skin and over-sharp grain. The publisher has stated training is paused while those are addressed, so expect the recommended checkpoint to move again.
Source: larryvrh/MiniMax-H3-Turbo-Lora README
Sigma shift, and why it is a pair
Every MiniMax H3 LoRA run inherits this pair, so it is worth knowing even if you never touch it. The default is 12 / 3. Two shift values, one for video and one for audio. The model derives both video and audio timesteps from the current video sigma plus these two shifts.
Because the two streams share a sigma but not a schedule, the shift pair is what keeps them aligned. Changing it without understanding that relationship is how people end up with drifting audio.
In the dual-clock sampler these two values are fixed, not tunable — they are taken from the original implementation. If a workflow exposes them as sliders, leaving them alone is the correct move.
Source: ostris/ai-toolkit · packing.py (VIDEO_SIGMA_SHIFT / AUDIO_SIGMA_SHIFT) · shuaixn/ComfyUI-MiniMaxH3DualClockSampler
Acceleration beyond the MiniMax H3 LoRA
A separate approach from step reduction: skip work rather than take fewer steps. These stack with the MiniMax H3 LoRA but come with sharper constraints.
minimax-h3-velocity-cache-v1
Caches target audio and target video rows as one compact tensor to skip repeated work.
Constraint · Locked to 1344×768, 124 frames and shift 12·3. Any other parameter combination is rejected outright — this is a fixed-configuration accelerator, not a general speedup.
ComfyUI-Spectrum-MiniMax-H3
Forecasts post-transformer features with Chebyshev ridge regression so selected transformer evaluations can be skipped entirely.
Constraint · Adaptive scheduling with sampler-aware safeguards and native fallbacks. More flexible than a fixed cache, but it is prediction — expect it to trade some fidelity for the skipped compute.
What running the MiniMax H3 LoRA actually asks of your machine
Fewer steps is not less memory. A MiniMax H3 LoRA cuts sampling time; the base model it attaches to is the same size it always was.
- Base model size
- The base is a ~33B model, and the publisher notes an 80 GB GPU is comfortable at the largest resolutions. The ComfyUI node streams the base and offers the low_vram switch, which is why it still runs on far smaller cards.
- Without a ComfyUI graph
- The repo ships a single self-contained generate.py that loads the base, applies the LoRA, runs the 4-step dual-schedule sampler and muxes an mp4. It still needs a ComfyUI checkout for the model, VAE and text-encoder definitions.
- Trading VRAM for system RAM
- In the standalone script, --offload-adaln moves roughly 13 GB from VRAM to CPU RAM. Lowering resolution or frame count is the other lever.
- Validated sizes
- Width and height are multiples of 32 with a 768px short edge, and frame counts follow the 17k+5 grid at 24fps. The publisher reports validating roughly 124 to 362 frames, which is about 5 to 15 seconds.
Source: larryvrh/MiniMax-H3-Turbo-Lora README
Training your own MiniMax H3 LoRA
The turbo LoRA changes how fast H3 samples. A MiniMax H3 LoRA you train yourself changes what it renders — a character, a product, a look. Ostris’ AI Toolkit added an H3 training extension on August 3, 2026, and the details below come from that extension’s own source rather than from a write-up of it.
- What can be trained
- AI Toolkit's minimax_h3 extension trains text-to-video and first-frame image-to-video LoRAs, with joint audio when the dataset provides it. Image-only datasets train as single latent frames, and sampling with one frame renders a still.
- Which base it loads
- By default it pulls the Comfy-Org repack — the pruned int8-ConvRot transformer and the NVFP4 AWQ Qwen3-VL text encoder, both kept quantized through training. A config switch picks fl2va, fl2va_pruned (the default), ref2va or ref2va_pruned.
- Frame counts snap down, not up
- Training clips are aligned to the previous valid 17k+5 frame count rather than the next one. This is the opposite direction from the ComfyUI workflow, which rounds your requested duration up — worth knowing if you are matching a dataset to a target length.
- No CFG at validation
- H3 is guidance-distilled and the released sampler runs without classifier-free guidance, so previews during training should be generated at guidance 1. Validating at a higher scale tells you nothing about how the LoRA will behave in the real workflow.
- The audio stream trains on its own clock
- The two streams ride different flow shifts — 12 for video, 3 for audio — and the audio sigma is derived from the video sigma at every step, in training exactly as in sampling.
Source: ostris/ai-toolkit · extensions_built_in/diffusion_models/minimax_h3
We have not trained a MiniMax H3 LoRA ourselves yet, so there are no timings or loss curves here — only what the toolkit does and expects. Reference-driven training was not in the extension at the time of writing; text-to-video and first-frame image-to-video were.
Speed you don’t have to configure
A MiniMax H3 LoRA is worth setting up if you generate constantly on your own hardware. H3Video’s hosted text-to-video generator skips the checkpoint hunt, sampler swap and local audio debugging.
Compare hosted pricing →Before downloading anything, confirm your card can hold the base model — the VRAM requirements breakdown explains why the 32B text encoder, not the diffusion weights, is what usually decides whether a MiniMax H3 LoRA run fits. The local setup guide lists every file and folder, and the ComfyUI workflow guide covers the templates a MiniMax H3 LoRA plugs into, including the shipped sampler settings this one replaces.
Frequently asked questions
›How much faster is the MiniMax H3 turbo LoRA?
It renders joint video and synchronized stereo audio in 4 sampling steps instead of the usual ~20, which is roughly a 5× reduction in sampling wall-clock. With the current ckpt850 weights the publisher states 4 steps is already sharp; it was the earlier checkpoints that needed 6 to 8 steps to firm up.
›Can you train your own MiniMax H3 LoRA?
Yes. Ostris' AI Toolkit added a minimax_h3 extension on August 3, 2026 that trains text-to-video and first-frame image-to-video LoRAs, with joint audio when the dataset provides it. It loads the Comfy-Org repack by default and keeps the transformer and text encoder quantized through training, which is what makes it viable on consumer hardware.
›What LoRA strength should I use with MiniMax H3?
Start at 1.0 and treat it as a symptom dial. If the result shows blurry ghosting or smear, raise it towards 1.05–1.2; if it shows over-sharp grain or crunchy artifacts, lower it towards 0.8–0.95. Both directions are documented by the publisher, so it is worth trying before you swap checkpoints.
›Why does audio sound broken at 4 steps?
H3 generates a video stream and an audio stream together, and ComfyUI's stock samplers advance both on one shared schedule. That is harmless at 20 steps but massively over-steps the audio branch at 4, producing distortion. A dual-clock sampler that advances each stream on its own schedule fixes it.
›Which turbo LoRA checkpoint should I use?
minimax_h3_turbo_4step_ema_ckpt850.safetensors — the publisher marks it as the current final checkpoint of this training round. The ckpt500 files that circulated earlier are superseded, and so are the two initial releases. Note that the publisher has paused training over plastic-looking skin and over-sharp grain at 4 steps, so expect this recommendation to move.
›What are the MiniMax H3 sigma shift defaults?
12 and 3 — one for video, one for audio. The model derives both video and audio timesteps from the current video sigma plus these two shifts, which is what keeps the two streams aligned. Change them only if you understand that relationship.
›Can I use the turbo LoRA with a pruned or quantized model?
Yes. The publisher states the LoRA works with the full MiniMax H3 base as well as the pruned and curve variants, and that it detects a pruned base automatically. You do not need to re-download an unpruned checkpoint to use it.