⚡ MINIMAX H3 TURBO LoRA
4-STEP AUDIO + VIDEO ACCELERATION
EXTREME SPEED // NATIVE AUDIO // COMFYUI READY
MiniMax H3 Turbo LoRA is an experimental acceleration adapter designed to drastically reduce MiniMax H3 video generation time while preserving its native synchronized stereo audio capabilities.
It reduces sampling from the usual higher step count down to only 4 steps, providing extremely fast previews and significantly shorter generation times.
The LoRA supports both:
OFFICIAL EKKIVOK WORKFLOW :https://civarchive.com/models/2838258/minimax-h3-4-steps-turbo-video-aio-workflow
TEXT TO VIDEO
IMAGE TO VIDEO
FIRST / LAST FRAME VIDEO⚠️ EARLY EXPERIMENTAL RELEASE
This LoRA is currently an experimental and under-trained prototype.
The 4-step mode provides impressive speed, but it may produce:
Softer image quality
Reduced fine details
Unstable textures
Motion ghosting
Motion trails
Lower overall quality compared with the full MiniMax H3 model
For improved visual quality, using 6 to 8 steps may produce cleaner results while still remaining much faster than standard generation.
🚀 RECOMMENDED SETTINGS
LoRA Strength : 1.0
Steps : 6–8 recommended
Fast Preview : 4 steps
Scheduler : simple
CFG : 1.0 or workflow default
Frame Rate : 24 FPS
Resolution : multiples of 32
Short Edge : approximately 768 pxRecommended LoRA file
minimax_h3_turbo_4step_ckpt500.safetensorsThe non-EMA ckpt500 version is currently the recommended checkpoint.
The EMA version may produce softer results and can be more prone to ghosting or visible motion trails.
🔊 NATIVE SYNCHRONIZED AUDIO
MiniMax H3 generates video and audio together:
VIDEO STREAM
+
32 kHz STEREO AUDIOThe audio is generated natively and synchronized directly with the video.
A dedicated Turbo sampler is required because MiniMax H3 uses different sampling trajectories for video and audio:
Video Flow Shift : 12
Audio Flow Shift : 3Using a standard ComfyUI sampler at only 4 steps may incorrectly sample the audio component, resulting in distorted, saturated, or unusable sound.
🧩 REQUIRED COMFYUI CUSTOM NODE
Install the following custom node through ComfyUI Manager:
ComfyUI-MiniMax-H3-TurboOfficial repository:
https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo
The node pack includes:
MiniMax-H3 Turbo LoRA
MiniMax-H3 Turbo SamplerThe custom node also supports pruned MiniMax H3 checkpoints by automatically applying the required AdaLN LoRA layers during generation.
🧠 BASE MODEL COMPATIBILITY
The Turbo LoRA is compatible with several MiniMax H3 model formats:
BF16
INT8 ConvRot
Pruned INT8
Pruned FP8The dedicated Turbo node automatically detects the model structure and applies the LoRA appropriately.
📂 INSTALLATION
Place the LoRA file inside:
ComfyUI/models/loras/Recommended workflow structure:
Load Diffusion Model
↓
MiniMax-H3 Turbo LoRA
↓
MiniMax-H3 Turbo Sampler
↓
Video and Audio OutputRecommended sampling configuration:
Scheduler : simple
Steps : 4–8Do not use a standard sampler for the 4-step audio workflow.
👤 ORIGINAL CREATOR
Larryvrh
This LoRA was originally created and trained by Larryvrh.
Original Hugging Face repository
https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora
Official ComfyUI custom node
https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo
All credit for the LoRA training, development, research, and dedicated Turbo sampler belongs to Larryvrh.
📤 CIVITAI UPLOAD INFORMATION
Original Creator : Larryvrh
Civitai Uploader : EKKIVOKThis Civitai page is a community re-upload intended to make the model easier to discover, download, and use with ComfyUI.
I am not the original creator of this LoRA.
Please support and follow the original creator through the official repositories linked above.
⚡ SPEED AND QUALITY GUIDE
4 STEPS = MAXIMUM SPEED / PREVIEW QUALITY
6 STEPS = BALANCED TURBO MODE
8 STEPS = BEST CURRENT TURBO QUALITYThe Turbo LoRA is especially useful for:
Rapid prompt testing
Motion previews
Seed comparison
Fast workflow testing
Video drafts
Quickly testing image-to-video inputs
Reducing MiniMax H3 sampling time
For maximum final quality, the standard MiniMax H3 model without the Turbo LoRA and with a higher step count may still produce better results.
CREATED BY LARRYVRH
UPLOADED TO CIVITAI BY EKKIVOK
MINIMAX H3 // TURBO MODE // FUTURE IN MOTION
Description
ALPHA VERSION.
FAQ
Comments (16)
Could this be used as a form for preview generation? And disabled for a "final" (same seed etc)?
Yes. but wait i made a complete incredible workflow that is about to come ;)
Sounds great, thanks @EKKIVOK
I just read /realized you can also ofc just reduce the image size, say render video at 0.2 Mpix first to check how it is, then increase for a final one.
And maybe combine with this and other speedup tricks, idk.
I think these are loras from https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora which is still in active training. Please correct me if I am wrong.
didn't notice there was already the same model uploaded. (they are the same)
So there is "no much better"...
@meeatsmeat yeah it's wrote on the thumbnail & desciption.....
@EKKIVOK i'm tests both. And your version is suck.
It works great with 8 steps, but I'd love it if, in the future, it didn't require custom nodes to work and could be used directly in SwarmUI. Thanks for the lora. I thought it would take us weeks to see something like this.
Love from earth! Thank your for the hard word!
I get way worse motion blur when using it at the suggested 4 steps
Scumbags repackaging other peoples' work. GTFO.
Wait? I thought he is the owner of this lora.
tried at 4, 6 and 8 steps. The higher I go the blurry it gets. I don't know why but I get very bad quality audio with this lora at any step. Unsure whether using it along with following nodes plays an impact: RIFE frame interpolation, sage attn, sol-attn, easycache.
disabled this lora and had no sound issues.