<h3 id="overview-9nfkkt5gt">Overview</h3><p><a rel="ugc" href="https://lilidesbellons.com/"><em>Lili des Bellons</em></a><em> is (most likely) an alias of French illustrator Laurent Meriaux.</em></p><p><em>Drawing inspiration from pop culture, medieval art, African art, Japanese animation, and the iconography of European poster artists, Lili works in series of 10 to 15 images, exploring and developing personal worlds. The viewer becomes a witness to vibrant, somewhat unsettling, and futuristic illustrations, where strange characters come to life, interacting with their peculiar universes. Lili creates dynamic narratives through art, inviting the viewer to journey through strange yet delightfully colorful worlds. </em></p><p></p><h3 id="usage-i92lwo4k9">Usage</h3><p>For testing the following parameters were used:</p><pre><code>Model: flux1-dev (fp8_e4m3fn)
Text Encoder: t5pxxl_fp16/clip_l
Sampler: euler
Scheduler: normal
Steps: 24
Flux Guidance: 4
LoRA Strength: 1</code></pre><p></p><h3 id="training-xzrc1ma5i">Training</h3><p>Several versions of LoRA with various combinations of parameters were trained, but capturing the surrealistic style of the illustrator proved challenging. Most early LoRAs tended to produce images where the line between surrealism and poor execution was too thin.</p><p>This LoRA is the most successful of all, though it still does not fully succeed in reproducing the compositional structure of the artist's creations. However, it does properly convey the palette, linework, and color transitions. While it does not always produce aesthetically pleasing results, I have decided that it is good enough to be published.</p><p></p><p>The LoRA was created using the <a rel="ugc" href="https://github.com/ostris/ai-toolkit">AI Toolkit</a>. A total of 148 images were used, all captioned with the same phrase: "Image in style of Lili des Bellons." Training config includes:</p><pre><code>optimizer: dadaptation-adam
lr: 1
rank: 8
alpha: 8</code></pre><p>All other parameters were left default.</p>

Files with SHA256: db08b0c82cee62cd64bf24efe2353a8dde6e9021279077c9cfd29e4a541680d1