djz Season 2 !
produces images of Poster Geo
THESE MODELS ARE DESIGNED FOR MERGING WITH OTHER MODELS, STACKING MANY LORAS ON TOP. thx ;)
Description
FAQ
Comments (4)
hey @driftjohnson I would like to know how you create these models, I ask because you upload a lot and that's great! I would like to know if you create them from scratch, that is, do you only use the SD base model but do not mix it with any other model from the community? Also, if this model is specific for this, how flexible is it? I will try one of your models
@jesuelmdesign I create them from scratch using various other tools, each image is edited and graded with basic photography techniques like color grading, xposure, contrast, vibrance, saturation, luminance noise, colour noise reduction, grain, sharpen and parametric tone (dark/light/high/shadow/midtone). Sometimes objects will be removed from images used to train these models. Once Mastered all SD2.1-768 models are trained on the default pruned model from Stability.
All these CKPT "Tight Models" are designed for mixing with each other or with "Large Models" found here on Civit from other popular Authors. I recommend testing merging in pairs on sum weight at 0.5 or by injection to a target model (in the same way an inpainting model is created.
Some of these models are quite rigid and this is done to prevent "dilution" that happens in the 0.5 sum weighted merge process. Also stacking several LORA on top is now furthrer enhancing this approach too.
Thanks for asking!
@driftjohnson I find the way you create your models very interesting, I ask because I have been doing an exercise to see how much the models differ from each other, what I do is repeat the same prompt in different models under the same conditions! what I have found is that a great majority of the models of a (specific category), tend to give very equal results! So I have discarded many models that are actually merged from other models, and now I try to find models that are different enough in their base dataset, to have a gallery of more interesting resources. https://jesuel-m.imgbb.com/ https://jesuel-m.imgbb.com/
@jesuelmdesign yes, i'm focused on creating entirely original Diffusion models, They are designed for merging and this is why they have such a "tight" scope in terms of the subject matter. I believe that leads to less "collisions" when merging or injecting, this can be very dynamic and explorative in terms of experimental & abstract projects. When you factor in the merging of each of my models in pairs, plus the additional multiplier of the LORA stacks - this is infinite art. Of course you can do all of this with a single prompt and seed, so "the sky is the limit" is an understatement. The unpredictable nature of mixing/blending concepts combined with clever prompts can also yield very unique results.
I tend to look at the djz series as Paint Colours. Artists can then mix and paint with concepts as they might with colours.
Also many of my models can be used without the trigger word, you might be surprised that the influence is remaining :)