Model trained with 100 images ( 50 + 50 fliped ) using the Seek Art Mega model as a base model before training and later used to "fix" the damage caused by dreamboothing through multiple merges: https://civarchive.com/models/1315/seekart-mega
Workflow and more samples with according tested artists: https://www.artstation.com/artwork/2qPoAx
Trained with some empty backgrounds with different solid colors to easily produce removable backgrounds;
Stacked training with stacked merges using the base model to avoid loss of other artstyles;
Most of the trained images contains tattoos, jewelry, stamps, and detailed equipment of scifi fantasy fashion;
Trained with micro crops of large compositions and according original compositions to easyly insert the characters within scenarios;
Description
The raw file already merged with SPYBG Toolkit ( just for ease of use if you don't know how to merge models ), then, after that, another dreambooth training session to restore the lost artstyle.
That model is highly biased towards 768x768 resolutions, anything below that resolution scale will only create ugly outputs, and strongly biased towards realistic / photografic characters ( loosing artstyles ), however:
Fucking good to generate character reference sheets and character turnarounds ( without even asking while prompting );
High quality hands as long as you don't prompt to pose them ( just let the model decide how the hands should be placed );
Tiling / motif probability ir nearly zero ( duplicated features );
Reduced amount of intricate details ( unless you prompt for );
Largely resilient to alternating keywords without loosing coherence;
Trained tokens inherited from original model;
Since most of the artsyles bias towards the same look, it's easy to blend lots of artists without causing noisy features or incoherence;
Tends to "obey" your prompts without a huge CFG scale;
FAQ
Comments (6)
All three versions are great models, but ... can you prune/reduce the file size to 2Gb also for the other two bigger versions ?
Is there a reason you don't make these LORA files? I am not criticizing just trying to learn.
Lora tends to "converge" outputs by hacking the layer outputs, thus, it's not like "blending", because even if you try to add triggering keywords many other aspects may deviate from your original prompt output, Lora is like "a post processing image effect"... also, dreambooth "damages" the model mostly destroying every artstyle that diverges from your training data ... the best alternative ( and tedious because of the amount of necessary training data and precisely detailed caption files ) is "fine tunning" the missing artstyles on stable diffusion ( instead of using unknow triggering keywords, you literally use an already know keyword by stable diffusion to FIX the artist broken artstyle) .... that model is fixing many broken artistyles and artists from Magna Carta, The War of Genesis, Blade and Soul, Aion online and Black Desert....
dreambooth = damaging the model like a sculptor to make it output the desired art
Lora = hacking the outputs to make it look like the desired art
Hypernetwork = much better than Lora because it's not exactly a hacking and it's like a "patch" or additional data input on your model, however, no one ever did a thing to evolve that technology and added "regularization images" that could nearly make that technology shine like LORA
Embbedings = It's like storing the best prompt you could ever write to reach the desired art and then save that behind a token
Aesthetic gradients = when you don't know how to write a prompt because it's hard to describe with words, then, use aesthetic gradients to "enhance" your prompt ( an image tells more than words )
Wow thanks a lot for indepth answer. I gained more insight on the subject. It aslo explains why extracted loras doesn't do as well as the models they are extracted from.
Hi was going to download a version to try out. Noticed that the trigger word for the seekartmega version was different.
thanks !!! it's CCDHT artstyle !!! I will fix right now !!!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.