<h3 id="overview-nlhhgci10">Overview</h3><p><em>This F1D LoRA strives to reproduce the artistic style of </em><a target="_blank" rel="ugc" href="https://www.instagram.com/kilianeng/"><em>Kilian Eng</em></a><em>, a Swedish visual artist known for his intricate, highly detailed, retro-futuristic sci-fi illustrations. Working under the name 'DW Design', he creates art prints and posters for major clients like HBO, Wired, Warner Bros, and The New York Times. His work blends dreamlike surrealism with physical realism, drawing inspiration from artists like Syd Mead and Moebius. Eng's focus on light and color brings a fantastical, alternative reality to life in his pieces. He regularly releases personal and commissioned works, including prints for video games like Castlevania and Shadow of the Colossus.</em></p><p></p><h3 id="usage-dg6rbqxe9">Usage</h3><p>Gallery images were generated with the following settings:</p><pre><code>Model: flux1-dev (fp8e4m3fn)
Text Encoder: t5pxxl_fp16
Sampler: euler
Scheduler: 24 steps (normal)
Flux Guidance: 4
LoRA Strength: 1</code></pre><p></p><h3 id="training-w9eh9zbzt">Training</h3><p>The LoRA was fine-tuned on an RTX 3090 using the <a target="_blank" rel="ugc" href="https://github.com/ostris/ai-toolkit">AI Toolkit</a>. A total of 970 images of various resolutions (ranging from 544x544 to 1440x1803) and mostly high quality were used, captioned with <a target="_blank" rel="ugc" href="https://huggingface.co/openbmb/MiniCPM-V-2_6">MiniCPM-V 2.6</a> using the prompt "Make a detailed description of this image" with the automatic insertion of "In style of Kilian Eng." at the beginning of each caption.</p><p>Training parameters were:</p><pre><code>Rank/alpha: 8/8
Optimizer: lion
Steps: 5800
Batch size: 1
Learning rate: 8e-5
Learning rate scheduler: linear
Betas: (0.95, 0.98)
Weight decay: 0.01
Resolution: (512, 768, 1024)</code></pre>

Files with SHA256: 35744c06a460a9f63edc2dc384b59376e1d35973e4db3cb3e7193d57c24d25e8