This LoRA is trained on a dataset of images from "Ну Погоди!" (Just you wait!) soviet era animated cartoon series. This is my first (published) LoRA attempt and I welcome your feedback.<h3 id="v1.2.1-qwen-2.1">v1.2.1 [Qwen 2.1]</h3>CFG: 2.0 - 4.0 (2.5 - 3.5 Best)Trigger word:<pre><code>soyuzmultfilmstyle</code></pre>Suggested strength range:<ul><li>Characters featured: 0.75 - 1.25 (0.90 - 1.10 Best)</li><li>No characters: 0.5 - 1.5 (0.75 - 1.25 Best)</li></ul><h3 id="v1.2.1-krea-2">v1.2.1 [Krea 2]</h3>This version simplified the trigger word structure. I would still recommend using the "helper phases" as described in v1.2 if you want to nail the look.Trigger word:<pre><code>soyuzmultfilmstyle</code></pre>Suggested strength range:<ul><li>Characters featured: 0.75 - 1.25 (0.90 - 1.10 Best)</li><li>No characters: 0.5 - 1.5 (0.75 - 1.25 Best)</li></ul><hr /><h3 id="v1.2-krea-2">v1.2 [Krea 2]</h3>This version simplified the trigger word structure. I would still recommend using the "helper phases" as described in v1.2 if you want to nail the look.Trigger word:<pre><code>soyuzmultfilmstyle</code></pre>Suggested strength range:<ul><li>Characters featured: 0.75 - 1.25 (0.90 - 1.10 Best)</li><li>No characters: 0.5 - 1.5 (0.75 - 1.25 Best)</li></ul>A new approach to image captioning was used, focusing on visual analysis, as outlined by @mariano_art in his <a target="_blank" rel="ugc" href="https://civarchive.com/articles/31233/how-i-train-artist-loras-not-style-filters-but-visual-logic">How I Train Artist loras: Not style Filters, but Visual Logic</a> article. Not sure if I nailed it, so please let me know your thoughts.Trigger word: "soyuzmultfilmstyle" or:<pre><code>soyuzmultfilmstyle, [YOUR PROMPT HERE.] Cel animation style, thick black ink outlines on characters and foreground props, borderless background scenery, flat unshaded colors, hand-painted background texture.</code></pre>Most of the time using just "soyuzmultfilmstyle" in the beginning of your prompt will produce really good results. Use full trigger word for more predictable results.Please note "thick black ink outlines on characters and foreground props." If your image features neither the characters nor the foreground props, please remove the entire part of the sentence for best results (generated images will likely feature characters or image will be framed to feature a prominent foreground).Suggested strength range: 0.75 - 1.25 (0.90 - 1.10 Best)Helper phases (see prompts in sample images for more examples):<ul><li>If your items/furniture/props are too complex, prepend:<ul><li>"minimalist"</li><li>"simple"</li><li>"flat"</li></ul></li><li>If the background is too detailed:<ul><li>"soft, hand-painted [background]."</li></ul></li><li>If you want a specific, simple 2D-look:<ul><li>"simple geometric forms"</li><li>"blocky shapes"</li></ul></li><li>Neutralize digital gradients:<ul><li>"featureless"</li><li>"solid unshaded [color]"</li></ul></li></ul>Character prompting/outfits:<ul><li>Wolf:<ul><li>"anthropomorphic grey wolf"</li><li>"pink collared shirt, yellow ascot tie, black bell-bottom pants, and a dark green flat cap"</li></ul></li><li>Hare:<ul><li>"an anthropomorphic [grey/brown] hare"</li><li>"bright green deep V-neck shirt over a white turtleneck shirt, and dark green shorts"</li></ul></li></ul><hr /><h3 id="v1.1-krea-2">v1.1 [Krea 2]</h3>Trigger word:<pre><code>s0yu3mu1tfi1m_style. high contrast shadows, flat color fills, vintage 1970s animation aesthetic. a retro cartoon illustration of</code></pre>Suggested strength range: 0.75 - 1.25 (0.90 - 1.10 Best)<hr /><h3 id="v1.0-krea-2">v1.0 [Krea 2]</h3>Trigger word:<pre><code>s0yu3mu1tfi1m_style</code></pre>Suggested strength range: 0.90 - 1.30 (1.00 - 1.10 Best)

Soyuzmultfilm Style [Krea 2, Qwen 2.1]

<ul><li>Adjusted the dataset captions to simplify trigger word</li></ul>

Krea2_SoyuzmultfilmStyle_v1.2.1.safetensors

Files with SHA256: f5a39562a6bef5d8ad22b5e46d181dc1f46cd25e1ed27ba60efea8696cccbafa