A smart merge of lightx2v & larryvrh's turbo loras plus JonXL's H3 photorealism lora. Apologies as I do not have a merge formula to give out for others to tweak and improve upon, a happy accident during testing pushed things in the right direction but unfortunately I no longer have the workflows needed to reproduce the merge.
During development this lora & variations of it have been great for my use case, silly little clips with natural language prompts on relatively low-spec hardware (16GB RTX A4000 Ampere) with comfyUI. I've been getting ~330s per 10s of output @ 0.5mp using the SLA attention node. But importantly video and sound quality have been adequate.
I am time-poor and have not been able to test this extensively but this seems to be an improvement over any other singular trained turbo lora.
No support is offered or implied.
Tested with the following parameters Sampler: Euler Scheduler: Simple Steps: 6 Shift: Not set
See h3_fast.json for the workflow used during testing.
Description
a few people have requested the old version so here it is.
FAQ
Comments (18)
is this different from the version that was taken down (which gave very nice results, by the way...in the wan2gp discord, testers were putting it over larry 600 turbo).
@hard_gravy_2 Thank you! Great work, BTW, this has become my new default accelerator lora.
It's different, I'll re-upload the one taken down as v0.1
what do you mean with putting it over the larry? @ltcdrbroccoli
@paolobassd sorry, I was speaking too idiomatically. In the wan2gp discord, a bunch of users downloaded it and tried it out. We all voted for the results of this lora over the larry 600 4 step turbo lora, which was the previous leader.
Have you tried the new lightx 768p v1.2 ?
What's the improvements over the old?
Most turbo LoRAs produce blurring with fast action/fighting scenes—curious to see how this one handles it.
just checked clarity is good .Don't see that blurry thing. But it's too strong. Need to reduce strength between 0.6 - 0.8
@anadarco12227 thanks for the feedback. Unfortunately I haven't had time to test out a wider range of video types and tune the merge accordingly. maybe over the next week or so.
@hard_gravy_2 but this is far better than other turbo loras which i have used.
I am still using version 0.1 , & I really like its results, not sure if v1 is a bigger improvement over it or not,
@hazzoom82659 I'm glad people are finding either one useful. I've tried finding a sweet spot between the two but I don't think it's going to work that way.
I've used this accelerator off and on for different kinds of clips, but right now I'm using it with rev2video and voice cloning from a starting frame, making videos for an H3 lora dataset. I'm using this model at 0.70 strength and the results I'm getting are fantastic. Very natural voice inflection and motion, face and hand gestures and mannerisms... it's an improvement over the base model even though I'm only using 8 steps. Nice stuff!
Any recommended settings for Video/Audio shift values?
I think it depends on the workflow. These speed loras are usually trained with video shift 12 and audio shift 3. but my workflow need video shift 8 and audio shift 3
One of the best turbo loras I've used so far for the MiniMax H3.
very good work! One of the best one that i've tested! Btw just curious does it work with ref2va as well?