1. Project Concept and Training Dataset Formulation
Idea and Artistic Objective The Soviet avant-garde and constructivism of the 1920s (the VKhUTEMAS school, the work of the Stenberg brothers, Alexander Rodchenko, El Lissitzky, Viktor Deni) revolutionized the global visual language. The rejection of bourgeois easel painting in favor of utilitarian graphic design established a rigid symbolic system: diagonal dynamism in perspectives, cinematic photo-montage, the integration of massive, carved typography into the flat plane, and a limited, laconic color palette (the contrast between deep black, signal red, and ochre-paper background).The project’s goal is to explore the plasticity of the generative diffusion model Stable Diffusion XL by fine-tuning a LoRA adapter to extract the visual invariant of the constructivist poster style from the 1920s and to transfer this formal logic onto contemporary and science fiction narratives (cybernetics, deep space exploration, urban utopia).
The experiment’s challenge was not to imitate individual subjects, but to reconstruct the thinking of the constructivist artists: generating posters where textual-graphic blocks and characters are subservient to a unified diagonal architecture.
Dataset Legal Compliance (Copyright Compliance)
In strict adherence to academic integrity rules and intellectual property law, the dataset was compiled exclusively from works in the public domain (Public Domain / PD-Old-70 / PD-RusEmpire) due to the expiration of exclusive copyrights (over 70 years since the authors' deaths or publication before 1943), as well as materials from open museum collections (Duke University Open Archive, Wikimedia Commons, DPLA).Dataset Composition and Data Normalization For training, a monolithic corpus of 11 authentic posters from the 1924–1929 period was selected, excluding stylistic noise from later Socialist Realism.
2. Learning Methodology and Technical Pipeline
The training is implemented on the Google Colab platform utilizing an Nvidia Tesla T4 GPU (16 GB VRAM). Architectural Solution: DreamBooth + LoRA Training the full SDXL model (2.6 billion parameters in the UNet block) on a single user GPU is technically infeasible due to memory limitations. Therefore, a combination of these methods was employed: DreamBooth: links an abstract visual style to a unique token identifier, sks, preventing the dilution of the base model’s pre-trained knowledge. LoRA (Low-Rank Adaptation): freezes the original Cross-Attention weights of the model and injects trainable low-rank matrices (r = 4). This reduces the volume of parameters needing updates by orders of magnitude and creates a lightweight weights file (.safetensors, approximately 50 MB).Selected Hyperparameters Base Model: stabilityai/stable-diffusion-xl-base-1.0 Instance Prompt: a poster in sks constructivism style Resolution: 1024×1024 px Batch Size / Gradient Accumulation: 1×4 (effective batch size = 4) Learning Rate: 1×10^(-4) (Constant schedule) Training Steps: 300 steps Mixed Precision: FP16 Memory Optimizations: 8-bit Adam (bitsandbytes), Gradient Checkpointing3. Resulting Visual Series
Below are key generations from the trained neural network, demonstrating the transposition of the constructivist canon onto new semantic contexts.
[RESULT PAGE 01: «LUNAR AVANT-GARDE»] File: soviet_gen_01.png (1024×1024 px, PNG) Prompt: a poster in SKS constructivism style, an astronaut standing on the moon holding a red flag, bold typography, geometric shapes, 1920s Soviet avant-garde lithography, 8k Negative prompt: photo, 3d render, modern vector, digital gloss, blurry, bad anatomy, deformed face, smooth airbrush, oversaturated neon Parameters: Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7.5, Seed: 100
[RESULTS SHEET 02: «CYBERNETIC LABOR»] File: soviet_gen_02.png (1024×1024 px, PNG) Prompt: a poster in SKS constructivism style, a giant industrial cybernetic robot assembling an engine, dynamic diagonal composition, red, black, and beige palette, agitprop art, textured vintage paper Negative prompt: photo, 3d render, modern vector, digital gloss, blurry, bad anatomy, deformed face, smooth airbrush, oversaturated neon Parameters: Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7.5, Seed: 101
[RESULTING SHEET 03: «THE HIGHWAY OF TOMORROW»] File: soviet_gen_03.png (1024×1024 px, PNG) Prompt: a poster in SK S constructivism style, futuristic train moving through a modern architectural bridge, perspective lines, constructivist photomontage, textured vintage paper Negative prompt: photo, 3d render, modern vector, digital gloss, blurry, bad anatomy, deformed face, smooth airbrush, oversaturated neon Parameters: Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7.5, Seed: 102
[RESULT SHEET 04: «ENERGY OF SPORT»] File: soviet_gen_04.png (1024×1024 px, PNG) Prompt: a poster in SKS constructivism style, young female athlete running forward, ray of light, heroic minimalism, bold block text, lithographic print Negative prompt: photo, 3d render, modern vector, digital gloss, blurry, bad anatomy, deformed face, smooth airbrush, oversaturated neon Parameters: Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7.5, Seed: 103
