As someone informed me, the turbo LoRa is a bit huge. It's 600mb and some of those layers don't do anything but affect flux guidance and make the outputs shiny and plastic. So why not go with what I've learned and just excise anything that is related to double_blocks.
The result is now 300mb and I seem to get less plastic/body horrors. I'm genning on 1.9 flux CFG, worked on openflux to reduce the steps some. Can't get anything at 4 steps but the original doesn't either.
Description
This is the bytedance hyper lora done up the same way. Its still pretty big and has to be used at .125. I think the newer ones are better, but if you like hyper. Here is it at 600mb vs 1.xGB
FAQ
Comments (3)
I don't have lots of LoRA experience, but this one is great for portrait photorealism. And you can use it with any flux checkpoint! It literally helps with every aspect of generating realistic portraits with Flux. They look better, it speeds up render time, the morphology is better and it adheres to the prompt better. The only downside is I don't think it would be great for complex scenes?
Here's my advice:
Figure out the minimum number of steps your particular checkpoint needs to not produce screwy morphology too often. For example RealDream needs about 12 steps. This obviously isn't enough to make the image look good, but at least a person will usually have the right shape.
Reduce the steps by 20% or so (so 10 in this example) because this LoRA does help get you better morphology in fewer steps, but don't rely on that too too much.
Once you have those settings dialed in, enable this LoRA, set the strength_model from .25 to 1.0. I like .75 at 10 steps in this scenario.
It will now render your image faster and with far more detail and beauty and realism than your model could likely do at 80 steps, especially if your prompt specifies those things.
One last tweek - reduce your FluxGuidance because this LoRA will increase prompt adhesion when it comes to details and lighting. The more you reduce it, the less plasticy things will look. I like 2.4 for RealDream.
I don't recommend using strength_model >1.0 if you want the best color tones, but if you are shooting for something like 6 step renders 2.0 could work. I don't think this is necessarily a good LoRA for that, as it's guidance isn't great even with t5xxl_fp16 and a clip_l finetune, but you can get a decent percentage of great looking simple images even at 6 steps.
Another use for this Lora which might be useful if you want to check a bunch of seeds for portrait layout / facial type is to set the strength_model to like 3, and do a bunch of 4 step renders of a bunch of seeds real quick to find one you like.
"And you can use it with any flux checkpoint!" That is not entirely true... these do not work with pruned Chroma checkpoint in Wan2GP... Not sure about in comfyui, I suspect it would run into issues there as well.
It would be so awesome if you made these for the pruned versions of chroma; https://civitai.com/models/1956762/chroma1-hd-gguf-and-fp8?modelVersionId=2314264
I have heard people got bytedance hyper lora to work with chroma in comfyui, but it doesn't work in wan2gp... maybe it's because I guess I'm using the fp8 version of chroma, though I suspect that wan2gp's more strict compatibility checks might prevent flux.1 dev models from working with chroma, that in fact do work, albeit potentially differently than they do for the f1d model.