AlekseyCalvin/Somov_Carnivalesques_V1_FluxLoRA_BySilverAgePoets

Somov's C𝕒𝓇𝓃ⅈ𝓋𝔞𝓁esque 𝓛𝒾𝒶𝒾𝓼𝓸𝖓𝓼

HISTORIC 𝒜𝓡𝒯𝚒𝕤𝕥 STYLE ℒ𝓸ℜ𝐴𝘴

||| By SilverAgePoets.com |||

Note: Variant I checkpoints of this model, herein, were fine-tuned at 64dim128alpha over DNA_1_618's Genova Apex De-distilled + LongCLIP-GmP-ViT-L-14 by zer0int.
Variant II checkpoints, found HERE, were fine-tuned at 64dim64alpha and over AfroMan4Peace's Colossus2.1_Dedistilled + LongCLIP-GmP-ViT-L-14 by zer0int.

A Low-Rank Adapter (LoRA) for FLUX-family models...
Fine-tuned on around 130 artworks inspired by Pierrot's lace tears, Columbina's satin gasps, Harlequin's velvet prances, by the toy moonlight of dreamed-up 18th centuries, and every achingly real rendezvous with fleeting masks of closeness...
And these artworks, of course, were crafted once upon a world by:
KONSTANTIN SOMOV (1869-1939), though representing but one facet, a brightest of dreams and among the deepest, of his manifold eclectic masteries.

Prompt with:

Scribble Somov style art, artwork by Konstantin Somov, detailed carnivalesque illustration & implore past-masked futures to guide you astray into endless carnival gardens; hedgerows tinseled with rainbows and moonglow, dresses & frock coats scuttle on echoes, rushing to catch for the billionth time early dawn fireworks of yearning igniting... Perhaps in these gardens you'd find at long last some mysteries yet to be dreamed...

Hyperparameters:

Fine-tuned on around 130 images, picked from a broader collection of 300+ Somov art files consolidated by us, and manually pre-processed/restored by us.
Upscaling was used sparingly, applied to only a small portion of the pieces; namely, only to those images which could not be found anywhere online in high quality or/and resolution.
Of the upscaled images, we added to the set only those that, after the upscale, did not appear to suffer extreme degradation texture/detail-wise, but seem unambiguously vivified in their roles as conflatable artistic translations apropos respective sources.
Somov the key Hyperparameters: 64dim/128alpha, Adamw8bit, lr.0005, ema:NONE, linear timesteps:FALSE, FluxCFG, resol.768, 1xbatch, on a single L4 via Google Colab Pro. Alas, due to the size of the dataset, we had to limit the resolution to 768. Even 1024 would cause out-of-memory errors, sometimes well into the training, at the tail end of the first epoch.

Use it with the 🧨 diffusers library

from diffusers import AutoPipelineForText2Image
import torch

pipeline = AutoPipelineForText2Image.from_pretrained('black-forest-labs/FLUX.1-dev', torch_dtype=torch.float16).to('cuda')
pipeline.load_lora_weights('AlekseyCalvin/Somov_Carnivalesques_V1_FluxLoRA_BySilverAgePoets')
image = pipeline('your prompt').images[0]