I tested MiniMax H3 Turbo LoRA in ComfyUI at only 4 steps, and my fastest RTX 5090 generation finished in about 19 seconds.
The speed is excellent, but 4-step generation can make faces and skin look too sharp or processed. I tested different LoRA strengths, SageAttention, EasyCache, and Low VRAM mode to find a better balance between speed and quality.
Which Turbo LoRA should you use?
My original tests used the CKPT 850 EMA Turbo LoRA.
However, the Turbo node developer now recommends the newer v4 step600 EMA checkpoint for most generations.
The older CKPT 850 version can still be worth testing for strong motion when you want to stay at exactly 4 steps.
So, if you are starting today, I recommend testing the newer v4 checkpoint first.
Install the Turbo nodes
Open ComfyUI Manager and install:
ComfyUI-MiniMax-H3-Turbo
After restarting ComfyUI, you should see:
MiniMax-H3 Turbo LoRAMiniMax-H3 Turbo Sampler
Place the Turbo LoRA file inside:
ComfyUI/models/loras/
Then connect the Turbo LoRA node to your normal MiniMax H3 workflow and replace the normal sampler with the Turbo sampler.
Use:
- Scheduler: Simple
- Steps: 4
- LoRA strength: 1.0 as the starting point
My 4-step test
For my first test, I used:
- CKPT 850 EMA
- 4 steps
- Strength 1.0
- SageAttention enabled
- EasyCache disabled
- Torch Compile disabled
- Low VRAM disabled
- RTX 5090
The generation took about 48 seconds.
The motion worked, but the face and skin looked too sharp.
Lowering LoRA strength improved the image
I then tested different strength values.
At 0.9, the result looked slightly better, but the colors were still distorted.
At 0.8, the result improved again.
At 0.7, the face looked more natural in my CKPT 850 test.
However, I would still start at strength 1.0 with the newer Turbo LoRA and lower it only if your image looks too sharp.
EasyCache reduced my generation time
Next, I enabled EasyCache while keeping SageAttention active.
My generation time dropped from about:
48 seconds → 34 seconds
The image was still slightly sharp, but the overall result looked good.
Low VRAM gave me a 19-second generation
The Turbo LoRA node also has a Low VRAM option.
When I enabled it, one RTX 5090 generation finished in only:
19 seconds
However, Low VRAM also changed the image quality in some of my tests.
I would keep it disabled if your GPU has enough VRAM. Enable it mainly when you are getting out-of-memory errors.
Clear faces work better
I also tested a wider image with a full-body character and several objects.
The movement and background looked good, but the face became slightly distorted.
When I used another source image with a larger and clearer face, the result improved significantly.
For 4-step generation, use a source image with a clear face whenever character identity matters.
My recommended settings
For fast previews, start with:
| Setting | Value |
|---|---|
| Turbo LoRA | v4 step600 EMA |
| Steps | 4 |
| Scheduler | Simple |
| Strength | 1.0 |
| SageAttention | On |
| EasyCache | Test separately |
| Low VRAM | Off unless needed |
If the result looks too sharp, try reducing the LoRA strength.
If motion or facial quality is weak, increase the generation to 6–8 steps instead of forcing everything into four steps.
Final result
MiniMax H3 Turbo LoRA can make ComfyUI generation much faster.
In my testing:
- First Turbo test: about 48 seconds
- SageAttention + EasyCache: about 34 seconds
- Low VRAM test: about 19 seconds
Four steps are excellent for fast previews and testing prompts. For final videos, I would also compare 6–8 steps because they can give better faces, texture, and motion.
The best setting is not always the fastest one. Use the lowest step count that still gives you the quality you need.
