MiniMax-H3 Turbo LoRA

MiniMax H3 is a 33B omni-modal video model that makes synced audio in the same pass as the picture. It's genuinely good - and it's slow, because it wants something like 20 sampling steps per clip. This node is the "turbo" in the equation: a LoRA adapter that collapses that to 4–8 steps so a model that big stops feeling like it costs you a coffee per render. No API, no key, no credit card - it runs the open weights on your own GPU.

It's one half of the two-node pack ComfyUI-MiniMax-H3-Turbo. This one is the MODEL → MODEL loader; its sibling MiniMaxH3TurboSampler provides the sampler that makes 4 steps actually behave. You want both, because they're built to work together.

How it works

Think of it as the H3 equivalent of the LCM-LoRA / SDXL Turbo lineage - a distilled adapter that patches a base model so it can take big, few denoising jumps instead of many small ones. The details are where it gets clever, because H3 isn't a normal image model:

• Two application modes via the low_vram switch. Off (default) applies the LoRA at run time (a "bypass"), keeping the update in activation space - sharpest, at the cost of some peak VRAM. On merges the low-rank update into the weights - lowest VRAM, but the tiny delta can get rounded away when folded into a quantized (int8/fp8/pruned) base, so it comes out softer.
• One LoRA file, every base. The node detects whether you loaded the full or the pruned/curve H3 checkpoint. On a pruned base, the LoRA's time-conditioning can't be a plain weight patch, so the node re-injects it at run time from a small silu(t_emb) grid that ships with the pack. You don't have to hunt for a "pruned version" of the LoRA - it just works.
• It streams the base model, so it runs on far smaller GPUs than a 33B model's file size suggests. An 80 GB card is only needed for the largest resolutions in bypass mode.

The inputs that matter

Four inputs, and honestly you touch two of them:

• lora_name - the file picker; put the turbo .safetensors in ComfyUI/models/loras/ and it appears here. Grab minimax_h3_turbo_v4_step600_ema.safetensors for almost everything; the older v1 (~850) is only friendlier for the narrow case of 4 steps and heavy, fast motion.
• strength - leave it at 1.0. The LoRA is tuned for 1.0 across the whole 4–8 step range. Only nudge it per-clip: blurry ghosting/smear → up to ~1.05–1.2, over-sharp grain → down to ~0.8–0.95.
• low_vram - leave off unless you OOM, which is exactly what its tooltip says.

The single output is a patched MODEL. Wire it between your model loader and SamplerCustomAdvanced - Load Diffusion Model → MiniMaxH3TurboLoRA → SamplerCustomAdvanced - and leave the rest of the official H3 graph (t2v or i2v) untouched.

Install

Via ComfyUI Manager, search "MiniMax-H3 Turbo" and hit install. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo

Restart ComfyUI, then drop the LoRA .safetensors into ComfyUI/models/loras/. You also need the base H3 model, VAEs, and text encoder from the official release (see the MiniMax-H3 ComfyUI tutorial). The pack itself is pure Python with no extra dependencies - it rides on ComfyUI's own torch and samplers. Keep it updated via Manager or git pull: it evolves alongside the weights.

Common issues

• Out of memory - flip low_vram on, or drop the resolution/frame count. The node streams the model, so OOM usually means you're at the edge, not that it's hopeless.
• Motion smear / trailing ghosting at exactly 4 steps with big fast motion - that's a known v4 quirk, and 6–8 steps mostly removes it. v4 also tolerates higher step counts better than v1 did.
• Stepping past 8 steps stops helping and starts over-sharpening. The sweet spot is 4–8, period.

One heads-up that isn't the pack's fault: the base MiniMax H3 weights carry a community license that excludes the US, EU, UK, and South Korea. The pack and LoRA are Apache-2.0; the base model you're patching is the part with the territory string attached.