A smart merge of various turbo loras plus JonXL's photorealism lora, adding the latter somehow helped to improve the video and audio quality.
Has been great for my use case, silly little clips with natural language prompts on relatively low-spec hardware (16GB RTX A4000 Ampere). I've been getting ~300s per 10s of output @ 0.5mp using SLA
No support is offered or will be given.
Description
Initial Upload
FAQ
Comments (8)
Nice work, how does it handle liquids? A big problem with all the turbo loras can be how they make liquids change/disappear mid flight or parts of the backgrounds alter/change through the scene.
Could you maybe add a simple liquid video? Might make people switch to your lora!
Yeah, I’ve noticed that too. It’s not just turbo LoRAs—regular LoRAs can also have a pretty serious impact on fluid effects.
great work, but could you at least mention the Strength, sampler, schedular, audio/video shift?
Workflow is attached to the sample vids.
sampler: euler scheduler: simple, shift was not set and idk what the default values are
lot of fl2v turbo loras, shame no any ref2v.
Grab yourself a hybrid model based on the fl2va model, and you'll get better results in reference mode just by swapping the model out. I like the 20-49 checkpoint for reference, myself- https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models
Use that in your reference workflow with a fl2va turbo model, and you're good. There's also some "delta-fused" models out there which are basically a different flavor of the same thing.
Either way, Minimax themselves admitted the open weight ref model is broken, so people figured out it's basically the fl2va with some modified layers that let it accept reference inputs. So they fixed it just by modifying fl2va.
@acedelgado143 To ask, do you specifically have to use MiniMax H3's FL2V node? Or can the REF2V provide the same results?
This works suspiciously well. Nice job!