This LoRA slider is trained on the concept of derelict versus pristine.
As usual with slider LoRAs, no keyword needed. You don't need to specify "decay" or anything else.
The recommended weights are -1.0 to 3.0.
I've tried to apply the effect on a specific item in the image but it seems that it's not possible, or maybe only with strong wording of the "immaculate" objects and low slider value.
Hope this helps someone create cool images.
Description
FAQ
Comments (7)
The BIG QUESTION is can these slider methods be deployed with i2i. Obviously the examples here show 'consistency' issues, which with t2i aren't really a thing (or concern). But with the comparison images, one imagines the same idea as a 'style' applied to an existing image- which would be both very cool and very useful.
guess that depends on your denoise setting.. id suggest trying really low values at high step count in several passes to maintain consistency
I have no idea how i2i models works.
The sliders LoRAs constist of telling the model to add weight to a concept.
Say you have a prompt like "a derelict vase", z-image is able to generate that. The slider only (in theory) increases or decreases the "derelict" keyword weights.
As you said there is a consistency issue with t2i seed-locking, because increasing the weights of a concept is a bit like adding a keyword to the prompt.
It would be like "a derelict and decaying vase" if you increase the slider weight, except you don't actually have a new word in the prompt so there is less shift.
So based on that, I don't see why i2i would have issues with it. Adding keywords to an i2i prompt will still yield consistent results, so applying a slider LoRA should be the same.
That said, I started training LoRAs 5 days ago so don't blame me if I'm wrong :)
Side note: I tried the sliders with z-image + controlnet depth and line and they do not overwrite the controlnet but seem less powerful. Shininess slider worked well.
@l226 I was recently told how edit models work basically (eg qwen edit): with regular models, you have your text encoder converting the prompt to tokens, which the gen model then translates to an image. With edit models, the VLM tokenises the image, thereby in a way recreating that image as accurately as possible in a language that the image gen model can understand; once you have that description, you can modify it with the prompt by simply tokenising that regularly and combining that with the VLM description (how it does that exactly, I do not know). So a slider should work with that just as effectively, assuming it works on the global level, both the prompt and the base, which would be sensible given loras can and do modify weights at all layers (if designed & trained to do so).
But, zimage edit has yet to be released.
The other standard i2i technique is inpainting - what it does is, rather than start from a blank slate (pure noise) and diffuse / denoise over some steps, is start from the raw input image, then reduce the amount of noise that's added (via the strength parameter). This can be used on any model, but it reduces the potential variation from input and level of comprehension it has of the details. Sliders will probably work, but getting a natural looking effect is difficult since it's a blunt tool (like malek says, slow cooking is one way).
Thanks @firemanbrakeneck, that makes sense! Can't wait to test it when z-image edit is released.
Also one possible issue is that I'm training for z-image turbo with ostris' training adapter, so I'm not sure how well these will translate to the non distilled model.
@l226 In my past experience, sliders will work on just about anything - even across heavy finetunes like asp / pony / illu, so I'm not too worried. What's really worrying me is compatibility for regular loras - by the time base releases, we may have thousands of loras (we've broken 1k last week without the help of major batch trainers), which may all be rendered ineffective.
This will be fun, thanks again.