Triggers are OPTIONAL: voronoi, basalt columns, pixelized, voxels
This LoHa model was trained on a mix of 3D Concept arts, real life basalt column coast images and many SDXL generated conceptual low-poly scenes with Voronoi-shaped prisms protruding from ground (low-light mostly).
Depending on weight, it produces the micro and macro Voronoi patterns and 3D shapes. Also, it can render miniature square voxels. You can expect more contrast, crispy image with shiny wet stones, sharp thin edges. Background and nature fixtures won't be affected by this LoHa most of the times. It tries to inject big Voronoi objects into big smooth surfaces and then turn smooth landscape of subject surface into grainy micro-voxel texture.
Trees and bushes in background, fog and mist looks flat. Mountains and buildings will look like constructed with basalt columns.
IMPORTANT NOTE:
I use elemental "Lora Loader (Block Weight)" node in ComfyUI with block weights of "1,0.73,0.43,0.13,0.03,0.063,0.83,0.93,0.999,1,0.97,0.67,0.15,0.05,0.125,0.65,0.85" with Model Strength of 0.3...0.5 and CLIP strength of 0.3...0.75
This LoRa is quite strong and also it affects CLIP and Model in different ways. It is more fun to play with two strengths and weight of blocks as it gives interesting and surprising results.
Didn't test this Lora in A1111 WebUI. No time for it... sorry.
Description
FAQ
Comments (9)
The model metadata indicates that the training dataset consistently used the spelling "low-poli" for the word "low-poly". It looks like more than a half of the images were associated with the caption containing "low-poli". It might be better to use "low-poli" rather than "low-poly" in your prompts (though I have not tested in detail). This is a note for the users of this model and not a criticism to the creator.
That is true.. By some reason auto captioning tool decided to use "poli" instead of "poly". It was to late to fix as training process was in progress when this issue was detected. And it might be actually a positive side effect - we can use non-existing word as trigger.
Another note. I don't see any of the words "basalt", "columns", or "pixelized" in the training dataset as far as the model's metadata indicates. I believe they are not the "trigger words" in ordinary sense. Perhaps they are intended as hints for the base checkpoint model. "voronoi" and the word "voxel" and its derivatives (such as voxelize or voxel-based) were actually used during the training (though with lower frequencies than "low-poli".) I hope this information help prompting better.
no idea why but using the pony version to generate cubes, yields nsfw images lol
edit: scratch it, genny is not working well with lycoris at the moment
Burned through quite a bit of blue buzz before giving up on this on the generator. I guess your edit comment makes sense - sounds like if I went to the XL version it wouldn't do great either.
The pony version really doesn't work very well...
which model are you trying to use with? More details please...
I see many recent samples that show up the intended style.
@homoludens The base model? When I was trying with pony, I used the basic pony model that it defaults to. I even tried adding one of the sdxl loras to the mix when I was trying with pony and it still didn't work out great.
It still kind of works, but trying to do anything with a character lora washes out a lot of the effect. Comparing it to the SDXL versions, the pony version is a lot more difficult to get a similar effect...
@eldersedai505 This LORA most likely was trained using some CinEro realistic Pony model. Not an original anime-ish Pony v6. Probably this might be the root cause. Can you try some realistic Pony?
Or look what other users use as the base...