Four Turbo checkpoint lines now compete to run MiniMax H3 in 4 to 8 sampling steps instead of about 20. This page tells you which one fits your task, how many steps to actually use, and where the quality goes when you push it too far.
Turbo selector
Choose what you are generating, what you are optimizing for, and what card you have. Every recommendation below comes from the checkpoint authors' own model cards, not from our benchmarking.
Recommended setup
by larryvrh · Apache-2.0
The author's current recommendation, and a new training recipe rather than another step count. v4 fixed the over-sharpened, plastic look of the earlier ~850 line and clearly improved faces, fingers, and fine texture on static and small-motion shots.
Open the model cardWhat to expect
The range the v4 author points at: noticeably better than 4 steps, with the micro-detail on faces and fingers that the earlier line lost.
Known trade-off
If your shot is dominated by fast motion, the v1 ckpt-850 line still handles it better at low step counts.
Steps
The step count matters more than which Turbo checkpoint you picked. Both Turbo authors put the quality sweet spot at 6 to 8, not at the 4 the LoRAs are named for.
4 steps
The advertised floor
What the Turbo LoRAs are named for. Fine on static shots, slow pans, and talking-head framing. Large or fast motion is where it breaks down first.
6 steps
First honest setting
The v4 author describes 6 to 8 as noticeably better than 4. Micro-detail on faces and hands returns, and most smear on moderate motion clears up.
8 steps
Top of the useful range
The other end of the author-recommended window. Past this, the same author reports no further gain — over-sharpening appears instead of detail.
~20 steps
Base model
Undistilled MiniMax H3. Still the reference for fast motion and for final work where a retry costs more than the extra minutes.
Side by side
Two independent distillation lines, plus a reference-to-video preview. All Apache-2.0, all community releases rather than official MiniMax models.
| Checkpoint | Tasks | Resolution | Steps | Pick it when |
|---|---|---|---|---|
| larryvrh Turbo LoRA v4 (step 600, EMA)larryvrh | T2V · I2V · FL2V · Ref2V | Mixed | 4–8 | Default for most work at 6–8 steps |
| LightX2V Turbo 4-step v1.0 (768p)lightx2v / ModelTC | T2VA · FL2VA | 768p (1344×768) | 4 | Maximum speed with native 768p output |
| LightX2V Turbo 8-step v1.0 (544p)lightx2v / ModelTC | T2VA · FL2VA | 544p mixed | 8 (or 4) | ModelTC's own workflow default |
| larryvrh Turbo LoRA v1 line (ckpt 850, EMA)larryvrh | T2V · I2V · FL2V | Mixed | 4 | Heavy fast motion locked at 4 steps |
| LightX2V Ref2VA Turbo 4-step v0.1lightx2v / ModelTC | Ref2VA only | 544p mixed | 4 | Reference-to-video (only option) |
The highest-resolution distillation of the group: trained at 1344×768 and guidance-free. Pick this when you want 768p out of a 4-step run rather than upscaling from 544p.
minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors
The default in ModelTC's own T2VA and I2VA example workflows. Distilled at 8 NFE, and the same checkpoint can be pushed down to 4 when you want the speed back.
minimax_h3_fl2v_turbo_8step_v1.0_bf16.safetensors
The author's current recommendation, and a new training recipe rather than another step count. v4 fixed the over-sharpened, plastic look of the earlier ~850 line and clearly improved faces, fingers, and fine texture on static and small-motion shots.
minimax_h3_turbo_v4_step600_ema.safetensors
Settings
Most of these are the same across all three checkpoint lines. The one worth reading twice is strength, where the model card and community practice genuinely disagree.
LoRA strength
1.0 (author) / 0.6–0.8 (community)
The v4 model card tunes for 1.0 across the step range. Community testing converged on 0.6–0.8 with 6–10 steps, and Kijai's conversions suggest lowering strength if outputs come out noisy. Start at 1.0, drop it if you see noise.
Scheduler
simple
The author's stated default. Kijai's conversion notes also show working examples on er_sde and sa_solver at 4 steps with 0.75 strength.
CFG / guidance
1.0
These distillations are guidance-free. Raising CFG does not sharpen the output; it fights the distillation.
Steps
4–8
Four is the floor, 6–8 is where both authors say quality lands, and past 8 the v4 author reports over-sharpening rather than gains.
EMA vs non-EMA
Prefer EMA
The larryvrh repo publishes both. EMA files are the intended ones; the non-EMA copies exist for comparison.
Install path
ComfyUI/models/loras/
MiniMax H3 has native ComfyUI support, so no custom node is required. Insert the LoRA between the model loader and the sampler.
Speed, honestly
The roughly 5× figure describes sampling steps alone — about 20 down to 4. Model loading, text encoding, VAE decode, and audio decode do not shrink with it, so a 5× sampling cut never produces a 5× shorter run.
The number that decides whether Turbo is worth it is time per usable clip. A 4-step run you throw away because the motion smeared costs more than a 6-step run you keep.
One community-reported machine
RTX 5080, 15 seconds at roughly 0.5 MP.
A single user's report on one machine and configuration, not a controlled benchmark. Treat it as the shape of the speedup, not as numbers you should expect to reproduce.
Low VRAM
Roughly in order of how much they cost you in quality. Work down this list rather than jumping straight to the most aggressive quantization.
drbaph · Abiray
Smaller downloads of the same Turbo weights, restructured for ComfyUI. The lowest-friction step down from full bf16.
RepositoryKijai
The conversions most ComfyUI workflows are built against. The repo notes to try a lower strength if you get noisy outputs.
Repositoryrzgar · t8star
Quantized builds of the LightX2V 4-step FL2V line. Experimental, and worth A/B testing against bf16 before you trust a batch to them.
Repositoryinfosave
The most aggressive tier. Reach for these only when nothing else fits, and check audio quality specifically — quantization tends to hurt sync before it hurts pixels.
RepositoryTroubleshooting
Almost every Turbo complaint traces back to one of six causes, and most of them are fixed by changing steps or strength rather than changing checkpoint.
Cause. The documented failure mode of 4 steps on large, fast motion. v4 is more prone to it than the older v1 line here.
Fix. Raise to 6–8 steps first. If you are pinned to 4 steps on action-heavy shots, switch to the larryvrh ckpt-850 checkpoint.
Cause. Either the older v1 (~850) line, which had this baked in, or a step count past 8 on v4.
Fix. Move to v4 step-600 EMA and keep steps within 6–8. Adding steps past 8 makes this worse, not better.
Cause. LoRA strength too high for your checkpoint conversion, particularly on the ComfyUI conversions.
Fix. Drop strength from 1.0 toward 0.6–0.8. This is the one setting where community practice diverges from the model card.
Cause. Audio is still being actively improved on the larryvrh line, and aggressive quantization degrades audio before it degrades video.
Fix. Test at bf16 before blaming the LoRA. If you are on a Q4/Q2 build, step back up a quantization tier and compare.
Cause. Reference adherence is the weakest part of every 4-step distillation, and Ref2VA Turbo is still a v0.1 preview.
Fix. Add steps before changing checkpoint. For final work, compare against a base-model run rather than shipping Turbo output directly.
Cause. The model is not fitting in VRAM, so ComfyUI is offloading over PCIe.
Fix. Stay at 544p rather than 768p, enable offload and block swap, and consider a pruned or INT8 build before dropping quality settings.
How it got here
This whole ecosystem is younger than most people assume, which is worth knowing before you treat any one checkpoint as settled.
August 5, 2026
The first larryvrh 4-step preview appears — a single experimental LoRA, and the start of the whole line.
August 7, 2026
LightX2V/ModelTC publish their first H3 Turbo distillation, a 4-step FL2V v0.1.
August 8, 2026
larryvrh v4 lands with a new training recipe. Community testing gets a working 2-step pass with clean audio, though video comes out visibly softer.
August 11, 2026
LightX2V Turbo v1.0: an 8-step checkpoint plus a 4-step variant trained at 768p, both with ComfyUI-ready bf16 weights.
Now
Pruned, INT8, and CMF-quantized conversions keep arriving. The open question has shifted from whether Turbo works to which Turbo to use.
How to use it
ModelTC ships example workflows for T2VA, I2VA/FL2VA, and Ref2VA in its repository. Start from those rather than rebuilding the graph by hand.
Official workflows and specsReference-to-video has only one distilled checkpoint; the FL2V line covers everything else. Your VRAM tier decides whether offload and 544p are mandatory.
H3 has native ComfyUI support, so no custom node is needed. Put the checkpoint in ComfyUI/models/loras/ and insert the LoRA between the model loader and the sampler.
Strength 1.0, scheduler simple, CFG 1.0 because these LoRAs are guidance-free. If frames come out noisy, drop strength toward 0.6–0.8 before changing anything else.
Run a single clip at the recommended steps. If motion smears, add steps before switching checkpoints. Judge by time per usable clip, not per run. While the base model is still downloading, render the same prompt on the hosted undistilled H3 first — it gives you a reference to hold the Turbo clip against.
Render the reference clip on hosted H3Questions
It is not an official MiniMax model. H3 Turbo refers to a family of community LoRAs that distill MiniMax H3 so it generates joint video and synchronized stereo audio in roughly 4 to 8 sampling steps instead of about 20. The main lines are LightX2V/ModelTC and larryvrh, with ComfyUI conversions from Kijai and others.
For reference-to-video, the LightX2V Ref2VA checkpoint is the only one distilled for that task. For everything else, larryvrh v4 step-600 EMA is the author-recommended pick at 6 to 8 steps, while the LightX2V 768p 4-step checkpoint is the better choice when you want maximum speed at 1344×768.
The roughly 5× figure describes sampling steps alone, dropping from about 20 to 4. Your total time also includes model loading, text encoding, VAE decode, and audio decode, none of which shrink by 5×. Judge the setup by time per usable clip, including the runs you throw away.
Use 4 when speed matters more than fast-motion fidelity. Use 6 to 8 for most work — both Turbo authors identify this as the range where quality lands, and the v4 author reports no additional benefit past 8 steps, where over-sharpening starts instead.
That is the documented failure mode of the step floor on large, fast motion. Raise to 6 or 8 steps first. If the shot is still dominated by fast motion, the older larryvrh ckpt-850 line handles that case better at low step counts than v4 does.
No. Turbo LoRAs are local ComfyUI and Diffusers checkpoints, and no hosted API exposes them yet. The generator on this site runs the standard MiniMax H3 model. This page is a setup reference for running Turbo on your own machine.
Both the LightX2V/ModelTC and larryvrh Turbo checkpoints are published under Apache-2.0. Check the individual model card before commercial use, since the base model and the LoRA carry separate terms.
No. MiniMax H3 has native ComfyUI support, so the Turbo LoRA loads through the standard LoRA loader. Put the file in ComfyUI/models/loras/ and insert it between the model loader and the sampler.
This is the one setting where practice diverges from the model card. The v4 card is tuned for strength 1.0 across the step range, while community testing converged on 0.6 to 0.8 with 6 to 10 steps, and Kijai’s ComfyUI conversions suggest lowering strength if outputs look noisy. Start at 1.0 and drop it if you see grain.
Community testing got a working 2-step pass with clean audio after v4 landed, but the video comes out visibly softer. It is a real capability rather than a practical default — 4 steps is the floor worth planning around.
Start from a pruned or INT8 conversion rather than full bf16, stay at 544p instead of the 768p checkpoint, and enable offload and block swap. Quantized builds down to Q4 and Q2 exist, but check audio sync specifically, since quantization tends to degrade audio before it degrades the picture.
Yes, and they stack. SageAttention speeds up attention while the Turbo LoRA cuts sampling steps, so they address different parts of the run. One community report on an RTX 5080 shows the pair landing well ahead of either one alone.
Keep reading
The long-form version of the step trade-off, including what "5× faster" leaves out and where Turbo still loses to base.
Where to download the weights, what the licence actually permits, and the ComfyUI setup the Turbo LoRA plugs into.
Fewer steps punish vague prompts. The structure that gets the model to commit to one interpretation.
This page keeps saying the same thing: fast motion and final delivery belong to the undistilled model. Locally that is the ~20-step run Turbo was meant to spare you. The generator on this site runs base MiniMax H3 in the cloud — so use Turbo to find the shot, and render the one you keep here without tying up your card.
Checkpoint names, step recommendations, and settings on this page are taken from the LightX2V/ModelTC and larryvrh model cards and repositories, verified August 15, 2026. MiniMax H3 Turbo LoRAs are community releases, not official MiniMax products, and this site is not affiliated with MiniMax.