I tested EVERY sampler × scheduler combo for MiniMax-H3 + LightX2V FL2V Turbo 4-step 1.0 so you DON'T need to
What I tested
The setup uses MiniMax-H3 with LightX2V's fl2v_turbo_4step LoRA:
I tested the full sampler × scheduler combination matrix.
Interestingly, there are quite a few combinations that produce good results besides the ones that have been discussed so far.
I also included combinations that normally benefit from higher step counts. They're not necessarily optimized for 4 steps, but I included them for completeness.
There's also a rating system at the top of the site. Since different sampler × scheduler combinations can produce surprisingly different results, I split the ratings into four separate dimensions:
Graphic Quality — overall image quality and whether the rendering looks correct
Motion Quality — motion, action sequencing, and whether the physics feel natural and coherent
Sound Effect Quality — whether voices and sound effects are generated correctly and match the scene
Music Quality — whether the background music is appropriate and correctly generated
Each category is rated from 1 to 5 stars.
If you find a combination that performs particularly well in any of these areas, please give it a rating. This should make it easier for everyone to identify which sampler × scheduler combinations are worth trying for different use cases.
I was short on time, so I used 480p instead of 768p for the comparison. The differences between the combinations are still quite apparent.
For consistency, all tests use the Baker example from Case 1 of the official MiniMax prompting guide.
Comparison:
Which one is the base (no lora) . It's hard to compare without it
this is so helpful, thank you.
this mode looks best
What is your own verdict?