A realistic LoRA for creating atmospheric post-soviet urban environments.
Works with FLUX.1-dev and QWEN-IMAGE.
This LoRA captures the haunting aesthetic of post-Soviet and Eastern Bloc urban landscapes, specializing in brutalist architecture, melancholic atmospheres, and the unique visual language of concrete panel housing estates (panelkas). Trained on a comprehensive dataset of 300+ meticulously captioned images, this model excels at generating atmospheric scenes that evoke the distinctive character of Soviet-era modernist architecture and urban environments.
Technical Notes:
Different captioning approaches were used for FLUX and QWEN training data to optimize each model's specific architecture and tokenization patterns
Subjectively, QWEN on BF16 produces more striking and impressive results with enhanced atmospheric depth and visual impact
Examples below and in the gallery include detailed settings and prompts for each version (FLUX.1-dev and QWEN-IMAGE) to help you achieve optimal results
Description
Qwen-Image LoRA:
Model Information:
* Base Model: Qwen-Image
* Network Type: LoRA
* Rank: 32
* Alpha: 32
* Trigger Word:Sovurban style:
Training Configuration:
* Steps: 3000
* Epochs: ~10
* Batch Size: 1
* Learning Rate: 1.5e-4
* Optimizer: AdamW8bit
* Caption Dropout Rate: 5%
* Seed: 42
Usage Recommendations:
* Content Type: Style / Concept
* Training Focus: Urban, Soviet, Post-Soviet architecture and atmosphere