Instructions to use larryvrh/MiniMax-H3-Turbo-Lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Inference
- Notebooks
- Google Colab
- Kaggle
4 steps
Your Turbo LoRA is the fastest and has the best quality I’ve seen so far at 6 steps. Other Turbo LoRAs turn into a buggy mess
Do you think you could eventually get it to run at 4 steps while keeping the same quality as 6 steps? IDK if that’s even possible, but either way, great job!
I'm still working on 4 steps version, but just no good outcome yet, I'm not sure even is it achievable eiter.
got it
Can you make one for ref2va too?
ref2va please
I'm still working on 4 steps version, but just no good outcome yet, I'm not sure even is it achievable eiter.
Definitely achievable, @silveroxides on huggingface made dareties (DARE + TIES) implementation with both yours and LightX2V, giving outstanding video and audio quality results at 4 steps. Try it out - it's currently the best Turbo LoRA.
I'm still working on 4 steps version, but just no good outcome yet, I'm not sure even is it achievable eiter.
Definitely achievable, @silveroxides on huggingface made dareties (DARE + TIES) implementation with both yours and LightX2V, giving outstanding video and audio quality results at 4 steps. Try it out - it's currently the best Turbo LoRA.
What video and audio shift, sampler and scheduler should be used for this LoRA?
safe to go 6 and 3, euler and simple is good across all turbo LoRAs.
If you walk to experiment, you can try er_sde and lcm instead of euler.
Same with video_shift, if you change 6 to 8 for better prompt adherence. See samples here:
https://huggingface.co/lightx2v/Minimax-h3-Turbo-SLA/discussions/1
CTRL+F the page and type "Video 5" to find the video with shift 8 and see for yourself my sample.
I feel like I'm taking crazy pills. The only Turbo loras that have looked any good are the .1 lightx and Larrys 600 ema. All of the others i've tried introduce significant artifacting and smudging