smkrv/panelka-flux

Generates moody Eastern Bloc architectural photography with panelkas, brutalist buildings, and post-Soviet urban environments

Public
71 runs
Playground API Examples README Versions

Panelka (Cinematic Eastern Bloc | Concrete Melancholy) - LoRA for FLUX.1-dev

Model Description

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.

Key Features & Capabilities

Architectural Styles:

  • Brutalist concrete apartment blocks with characteristic vertical and horizontal lines
  • Soviet-era panel housing (panelkas) in various states of decay and preservation
  • Modernist public spaces, metro stations, and observation towers
  • Industrial structures, cooling towers, and factory complexes
  • Distinctive architectural details: balconies, window patterns, weathered facades

Atmospheric Qualities:

  • Overcast, moody lighting with blue-gray tones
  • Dramatic lightning storms over urban landscapes
  • Golden hour and sunset scenes with warm-cool color contrasts
  • Foggy, misty conditions enhancing melancholic mood
  • High-contrast nighttime urban scenes with artificial lighting

Environmental Elements:

  • Abandoned and decaying structures with peeling paint and rust
  • Overgrown vegetation reclaiming industrial spaces
  • Vintage public transportation (trams, trains, buses) in various conditions
  • Seasonal variations: snow-covered buildings, autumn foliage, summer greenery
  • Urban wildlife (cats, dogs) in stark architectural settings

Unique Visual Signatures:

  • Documentary-style framing with straight-on, medium close-up compositions
  • Strong geometric patterns and repetitive architectural elements
  • Reflections in puddles and water surfaces
  • Overhead power lines and utility infrastructure as compositional elements
  • Human presence adding scale and narrative context (optional)

Training Specifications

  • Base Model: FLUX.1-dev
  • Training Steps: 3000
  • LoRA Rank: 16
  • Learning Rate: 0.00015
  • Resolution: 768x768 and 1024x1024 (square format)
  • Optimizer: AdamW 8-bit
  • Batch Size: 1
  • Caption Dropout: 5%
  • Gradient Checkpointing: Enabled
  • Latent Caching: Enabled for optimal training efficiency

Usage Instructions

Trigger Word: SOVURBNLVE

Recommended Settings:

  • LoRA Weight: 1.0 for strong stylistic effect, 0.4-0.6 for subtle influence
  • CFG Scale: 3-7.0
  • Steps: 28-50 (FLUX.1-dev defaults work well)
  • Resolution: Works best at 768x768, 1024x1024, or similar square/near-square ratios

Prompt Structure:

Start your prompt with the trigger word followed by descriptive elements. The model responds well to:

  • Architectural descriptions (materials, patterns, shapes)
  • Lighting conditions (overcast, dramatic, nighttime, sunset)
  • Camera angles (straight-on, low angle, elevated)
  • Atmospheric qualities (moody, melancholic, stark)
  • Environmental details (abandoned, snow-covered, overgrown)

Style Characteristics

This LoRA produces images with:

  • Color Palette: Dominated by cool grays, blues, and muted earth tones, with occasional warm accents (sunset oranges, autumn yellows, rust reds)
  • Composition: Documentary-style framing emphasizing geometric patterns and architectural scale
  • Texture: High detail in weathered surfaces, concrete grain, peeling paint, and rust
  • Atmosphere: Nostalgic, melancholic, cinematic quality reminiscent of Eastern European art cinema
  • Temporal Quality: Timeless aesthetic that can evoke both historical and contemporary periods

Best Use Cases

  • Cinematic concept art for films/games set in post-Soviet environments
  • Architectural visualization with brutalist/modernist aesthetics
  • Atmospheric urban landscape generation
  • Historical documentation-style imagery
  • Album covers, book illustrations, and editorial content
  • Urban exploration (urbex) inspired artwork
  • Melancholic narrative scenes with environmental storytelling

Technical Notes

  • Trained with comprehensive, detailed captions describing camera angles, lighting, composition, and subject matter
  • Excellent understanding of architectural terminology and photographic language
  • Optimized for square format outputs (768x768, 1024x1024) but adapts to various aspect ratios
  • Responds well to detailed prompts specifying technical photographic details
  • Can be combined with other LoRAs for hybrid styles, though works best at higher weights

Limitations

  • Primarily optimized for architectural and urban scenes; may not perform as well with close-up portraits
  • Best results with prompts that include photographic/technical descriptions
  • Strong stylistic bias toward Eastern European/Soviet aesthetics may affect other architectural styles
  • Works optimally with FLUX.1-dev; compatibility with other base models not guaranteed

Credits

Trained using ai-toolkit for FLUX.1-dev fine-tuning: https://replicate.com/ostris/flux-dev-lora-trainer/train. Dataset comprises original curated collection of Eastern Bloc architectural photography with detailed manual captioning for optimal training quality.

License: Please respect the original FLUX.1-dev license terms when using this LoRA.