Just a test Lora using an older Stable Diffusion 1.5 Lora and refining the outputs using Anima or Z-image with img2img. Stable Diffusion 1.5 as a drafter model and Z-image or Anima as the refinement model, interesting results overall.
Anima's and Z-image clip models allows for better understanding and allows for accurate images but they lack something. Something about rolling random chance with this older models that makes image generation much more varied and worth spamming hundreds of generations just to see something new. Anima and Z-image can refine the images and fix obvious mistake given the right denoise and steps.
Description
different loras, training images refined by Z-image
FAQ
Comments (4)
This is a very interesting endeavour. I'd be very interested to know more comprehensively about your findings sometime, if you were inclined to write them up.
I would start by saying that both anima and z-image produce really nice images so long as you feed them the prompt and add positive tags to steer the overall quality otherwise anima does know what to do and leaves the images looking the same and z-img hallucinates and adds weird details everywhere.
It was weird seeing a side by side comparison of an image created by SD15 and the refined image, the quality jump is as high as one might expect from the progress made but you can see that SD15 creates minute details z-img and anima cannot or do not if not prompted to do so.
I will be releasing more Lora's in the coming days.
I'm not exactly sure I follow, are you saying that you generated with SD1.5 and then img2img fixed with Anima? This LoRA is the collective result of those images?
Yes I did generate the images with SD1.5 and then either used Anima or Z-Image depending on the look I was going for. Z-Image is better overall for a sharper crisp look while anima can add tons of variety with artist knowledge and many style lora being made available.