<h3 id="overview-ioefk0use">Overview</h3><p><a target="_blank" rel="ugc" href="https://helianthus-annuus777.tumblr.com/"><em>Akiya Kageichi</em></a><em>, who also goes by "Golden Gravel", is a Japanese artist known for his intricate, dreamlike illustrations that merge mysticism and darkness. His work often presents complex, surreal worlds filled with shadowy figures like eerie jesters, lazy monarchs, and elusive warriors. These characters inhabit visually dense compositions where forms shift and overlap, creating an almost chaotic yet controlled flow of imagery. Kageichi frequently incorporates celestial and astrological symbols to invoke a sense of otherworldly influence. This symbolism, along with his unique blending of forms, gives his art an immersive, layered quality, suggesting hidden dimensions governed by magic and dark forces. The morphing elements in his illustrations invite viewers to explore a mysterious, interconnected world that seems to be shaped by unseen powers.</em></p><p></p><p><span style="color:rgb(37, 38, 43)">T</span>his was, to date, the most challenging F1 style LoRA I done. But if you look at the original artist’s creations, you must acknowledge that this style is very hard to reproduce. Eventually, after a bunch of unsuccessful attempts I managed to get LoRA that was close (yet far away) in vision to original artist's creations, was somewhat consistent, and introduced fewer abominations. To some extent, I might even say "abominations" are part of the artist's style, but he makes them look genuine. So, sometimes this model tends to produce "weird" results, but when it gets it right, it produces absolutely sick stuff.</p><p>You can probably use it with lower strength (0.8), in this case it will serve more like visual enhancement filter, producing less weird and disfigured shapes, but I did not test this extensively.</p><p></p><h3 id="usage-ln8ywsk7i">Usage</h3><p>Showcase images were generated in ComfyUI, here are the generation params I used:</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-88h9sxvq7">Training</h3><p>For training I used 178 images. The captions were simply "Image in style of Akiya Kageichi." (well, it was nearly impossible for me to even comprehend what was depicted in his creations, let alone caption them properly, so I just took the easy route 🤭).</p><p>LoRA was made with an <a target="_blank" rel="ugc" href="https://github.com/ostris/ai-toolkit">AI Toolkit</a> on an RTX 3090. Config parameters were all default, except for:</p><pre><code>Rank: 8
Alpha: 8
Steps: 3600
Optimizer: dadaptation-adam
Learning rate: 1</code></pre>

Files with SHA256: 784442eed83d1f3b52565742e377587d72ab0b53ebe42fa29dab8c96ba6adf77