<h2 id="hq-page-layout-or-style-lora-base:-anima">📖 HQ Page Layout | Style LoRA — Base: ANIMA</h2>CONSIDER DOWNLOADING AND TESTING THE 1800 STEPS CHECKPOINT TOO — IT OFFERS A FAIRLY DIFFERENT RESULT FROM THE 800, AND YOU MIGHT END UP PREFERRING IT.🙏 Before anything else: this is, by far, the LoRA that took me the longest to make so far — over 9 hours of training, at least 8 hours testing prompts, plus 3 days just hunting down text-free comic/HQ images and pages for the dataset, on top of another 4 hours cleaning up and adjusting the dataset files themselves. If you have a moment, please drop a Like and especially post your results in the comments (with the prompt/settings you used). I really want to improve this LoRA further, and I need to see more diverse results from different people to know where it's falling short. 🔥⚠️ This is a complex style training. Other steps beyond the two released here may perform better in certain specific cases. I'm not posting every checkpoint at once since it's a lot of files — if you'd like to test a specific step individually, let me know in the comments and I can upload it on request.⚠️ Trained with a mix of safe and NSFW comics/HQ pages. Contains no female content — it can be used for that purpose, but the focus is exclusively masculine. Only adult male characters and adult male comic pages were used in training.📦 Trained for 2800 steps total. Releasing the 800 and 1800 checkpoints.<hr /><h2 id="about-this-model">📊 About this model</h2>The vanilla ANIMA base already produces decent comic pages with the right prompt, but it usually ends up looking like a plain collage of images rather than an actual comic. What this LoRA does is bring real comic-page feel to that layout — it helps adjust and separate the panels properly, and can also recreate that natural "overlapping panel" effect comics typically have.This LoRA is a bit tricky to use — it needs some prompt practice. I strongly recommend reading through and studying how I structured my own test prompts in the posted preview images. No negative prompt was used, and generation settings were the same standard ones used across all my other releases, except for image size (see the Resolutions section below).<hr /><h2 id="checkpoints">📊 Checkpoints</h2>Step 800Leans toward more pastel tones, pulling slightly pink — released because I liked this particular resultStep 1800 Firmer, more solid result overallBoth checkpoints can still make mistakes. I recommend testing both and adjusting LoRA strength to see which fits your project best.<hr /><h2 id="generation-settings-used-for-preview-images">⚙️ Generation settings used for preview images</h2>All preview images were generated using:<ul><li>Base model: Anima 1.0</li><li>Acceleration LoRA: strength 1.0</li><li>CLIP: anima default</li><li>Steps: 12</li><li>CFG: 1.0</li><li>Sampler: <code>er_sde - beta</code></li><li>Image size: see the resolution guide below (varies by panel count)</li></ul><hr /><h2 id="panel-count-and-recommended-resolutions">📐 Panel count &amp; recommended resolutions</h2>These resolutions aren't mandatory, but they reflect the average resolutions used during training — sticking close to them makes the LoRA's job easier.1 panel<ul><li>Recommended ratio: 3:4 (standard portrait)</li><li>Ideal resolution: 768x1024 or 896x1152</li></ul>2 panels<ul><li>Recommended ratio: 1:2 (A4) or 2:3 (elongated portrait)</li><li>Ideal resolution: 724x1024 or 832x1216</li></ul>3 panels<ul><li>Recommended ratio: 3:4 (standard portrait)</li><li>Ideal resolution: 768x1024 or 896x1152</li></ul>4 panels<ul><li>Recommended ratio: 2:3 (elongated portrait)</li><li>Ideal resolution: 832x1216</li><li>Test note: the 1x2x1 layout performed best at 2:3 ratio, resolution 768x1152, in my tests</li></ul>5 panels<ul><li>Recommended ratio: 2:3 (elongated portrait) or 1:2 (A4)</li><li>Ideal resolution: 832x1216 or 724x1024</li></ul>⚠️ I don't recommend generating pages with more than 5 panels. Past 6, the model tends to lose coherence — it wasn't trained on that much density. A future V2 may address this.⚠️ Keep in mind the model can occasionally generate more panels than requested, depending on the length and amount of context/detail in your prompt.<hr /><h2 id="recommended-panel-layouts">🧩 Recommended panel layouts</h2>Always think about scene composition beforehand, and stick to a 2:3 or 3:4 ratio as a starting point. The following layouts work well:<ul><li><code>2 / 1</code></li><li><code>1 / 2</code></li><li><code>1 / 2 / 1</code></li><li><code>2 / 1 / 1</code> (recommend testing and adjusting the prompt)</li><li>Vertical 3-column: <code>1 / 1 / 1</code> followed by a tall single panel on the left or right (works well — try putting the smaller panels in close-up and the tall one in full body; requires some practice)</li><li><code>3 / 1</code> (not tested much)</li><li><code>2 / 2 / 1</code> (works, but needs testing with LoRA strength)</li></ul><hr /><h2 id="base-prompt-structure">✍️ Base prompt structure</h2><pre><code>[Activation Tag], [Style/Lighting Tags]. This is a [N]-panel comic page. First panel ([Spatial Position]): [Character physical description and actions]. Second panel ([Spatial Position]): [Character physical description and actions]...</code></pre><h3 id="spatial-position-tags">📍 Spatial position tags</h3>Always describe each panel's position in parentheses right after identifying it — this guides the model to assemble the correct layout:<ul><li><code>(top)</code> / <code>(bottom)</code> — full-width horizontal layouts</li><li><code>(top-left)</code> / <code>(top-right)</code> / <code>(bottom-left)</code> / <code>(bottom-right)</code> — 2x2 grids</li><li><code>(middle)</code> / <code>(middle-left)</code> / <code>(middle-right)</code> — center panels in 3-or 5-panel layouts</li><li><code>(main central)</code> / <code>(large foreground)</code> — highlight/focal panels</li></ul><hr /><h2 id="key-tags">🏷️ Key tags</h2>Core activation tags:<ul><li><code>hq_page_layout</code> — main tag to activate the comic-page layout</li><li><code>comic book page</code> — sets the comic/HQ aesthetic</li><li><code>multi-panel</code> — instructs the model to build a grid with multiple panels on the same page</li></ul>Style &amp; finish tags:<ul><li><code>cel shading</code> — flat-color shading style, common in comics/anime</li><li><code>digital coloring</code> — clean digital painting</li><li><code>ink outlines</code> — bold ink/nanquim linework</li></ul>Color palette &amp; lighting tags:<ul><li><code>warm colors</code> — warm palette (red, orange, yellow) — present in ~50% of training files</li><li><code>vibrant colors</code> — bright, intense colors</li><li><code>cool colors</code> — cold palette (blue, purple, cool tones)</li><li><code>soft lighting</code> — soft lighting</li></ul><hr /><h2 id="using-it-with-character-loras">🤝 Using it with character LoRAs</h2><ul><li>When combining with a character LoRA, don't place that character's tags at the very start of the prompt — use them only inside the specific panel where that character appears. If they show up in another panel, just repeat their tags there. This matters most when using 2 characters via LoRA.</li><li>If you're only using a single character, put their tags once at the start, then just call them by name in the other panels.</li><li>NSFW content: place the NSFW tag right after the LoRA's activation tag.</li><li>Even though this LoRA has NSFW content in its training, I still recommend using an additional LoRA to help with NSFW generation — it tends to make mistakes in that area on its own.</li><li>Eye and hand errors are also common.</li><li>I don't recommend using it to generate scenes with more than 3 characters via LoRA in the same page — it handles 1 or 2 well, or generic characters without a LoRA.</li></ul><hr /><h2 id="lora-strength">⚖️ LoRA strength</h2>Weight really matters here. Testing was done at strength 0.8 to 1.0, but you can go lower if needed for your specific case.<hr /><h2 id="one-more-thing">👍 One more thing</h2>I'm just getting started with this — if you have a moment, drop a Like and share your generations with the settings you used. It really helps me improve the next versions. Have fun! 🔥

Files with SHA256: 81c3eb0c8c9e5fe7336a4379924dfe09874c956b5509ce15da7dc1d052e60b92