Last updated: July 15, 2026OnikaTM is just a merge — no big story behind it. I mixed a bunch of Illustrious-based checkpoints, adjusted ratios, tested, redid it, tested again, until I landed on something I actually liked looking at. That's pretty much it.I'm not going to pretend this is some groundbreaking model or oversell it — there are tons of solid Illustrious merges out there already. This is just my version, shared in case it fits your style too.Recommended generation settings<ul><li>Sampler: DPM++ 2M / DPM++ 2M SDE / DPM++ 2M SDE Heun / DPM++ SDE / Euler a / Euler / LMS / Heun / DPM2 / Restart / DDIM / DDIM CFG / PLMS / UniPC / LCM</li><li>Schedule type: Karras or Align Your Steps or Beta</li><li>Steps: 35</li><li>Resolution: 1040×1520 (same 2:3-ish ratio as 832×1216, just scaled up — gave noticeably richer detail in my own tests, more than Hires. fix did on its own)</li><li>CFG Scale: 4.5–6.5 — both work well, just test what fits your prompt best</li></ul>Style-specific sampler combos (found through testing):<ul><li>For a retro anime drawing look: <code>LCM</code> + <code>Karras</code></li><li>For a more modern anime look: <code>DPM++ 2M</code> + <code>Karras</code>, <code>DPM++ 2M SDE</code> + <code>Karras</code>, or <code>Euler a</code> + <code>Karras</code></li></ul>Prompting tips (learned from a lot of testing)<ul><li>Quality tags matter, a lot. Starting your prompt with</li></ul><pre><code> masterpiece, best quality, amazing quality, very aesthetic, absurdres, highres, extremely detailed, intricate details, year 2025, newest</code></pre>makes a noticeable difference — it pushes the merge toward its best-looking outputs.<ul><li>Pick one theme and commit to it. This merge shines when the whole prompt reinforces a single coherent idea (one color palette, one outfit style, one mood) rather than mixing a bunch of unrelated elements in one shot. A focused prompt beats a crowded one.</li><li>Nuance on the above: stacking mood/aesthetic tags is fine (e.g. <code>cyberpunk lighting</code> + a single robotic accent, or <code>punk texture</code> + <code>neon palette</code>) — it's specifically stacking multiple clothing materials in one outfit (<code>latex</code> + <code>denim</code> + <code>chrome</code> together) that confuses the model. One material family per outfit, but mood layering is more forgiving than I first assumed.</li><li>Keep composition simple when you want max quality. <code>close-up</code> / <code>face focus</code> gives cleaner, more reliable results than full-body + complex background + busy outfit all at once — there's just less for the model to get wrong.</li><li>Camera angles: <code>dutch angle</code>, <code>high angle shot</code>, <code>low-angle shot</code>, <code>three-quarter view</code> all work well and change the vibe a lot for the same character. The merge also handles extreme versions better than expected:</li></ul><pre><code> extreme low-angle shot, extreme top-down shot, extreme dutch angle, fisheye lens, worm's eye view</code></pre>— as long as the outfit/material stays simple, these push composition further without breaking anatomy much.<ul><li>Body type &amp; age descriptors render cleanly. For an adult, established look (rather than default youthful output):</li></ul><pre><code> slim waist, generous curves, soft chubby face, mature adult woman, woman in her 30s</code></pre>combine well and stay coherent even under extreme angles.<ul><li>Facial expressions vary reliably. Tags like <code>smug</code>, <code>defiant</code>, <code>serene</code>, <code>playful</code>, <code>intense</code> work well even under close-up/face focus, and help break repetitiveness across a themed batch without hurting quality.</li><li>Materials &amp; outfits: <code>denim</code>, <code>latex</code>, <code>chrome</code>/metallic textures, <code>leather</code>, <code>silk</code> all render well — but avoid stacking conflicting materials in the same prompt. One per generation.</li><li>Texture tags like <code>skin texture</code>, <code>fabric texture</code> help a lot if your results feel too smooth or plastic-looking.</li><li>Lighting: <code>neon lighting</code>, <code>film grain</code>, <code>specular highlights</code>, <code>iridescent</code>, paired with a coherent night-time palette (<code>magenta</code>, <code>violet</code>, <code>deep blue</code>), is consistently one of the merge's strongest zones stylistically.</li><li>Negative prompt: keep it short and weighted on the model's actual weak points rather than a giant generic list. Hands are the main one:</li></ul><pre><code> (bad hands:1.3), (extra fingers:1.3), (six fingers:1.4), fused fingers, missing fingers, deformed hands, mutated hands</code></pre>makes a real difference.Example prompt<pre><code>masterpiece, best quality, amazing quality, very aesthetic, absurdres, highres, extremely detailed, intricate details, year 2025, newest, 1woman, solo, close-up, face focus, looking at viewer, [hair color], glossy hair, [eye color], detailed face, [outfit], skin texture, pores, neon lighting, specular highlights, film grain, depth of field, detailed background</code></pre>Example negative prompt<pre><code>worst quality, low quality, blurry, (bad hands:1.3), (extra fingers:1.3), (six fingers:1.4), fused fingers, missing fingers, deformed hands, mutated hands, bad anatomy, bad proportions, artist name, watermark, signature, text</code></pre>Give it a spin and see what you get. Feedback and generated images are always welcome — genuinely curious what other people get out of it.

OnikaTM

Genesis — the first release.This is the starting point: the merge that finally gave me results I was happy with after a lot of trial and error. I've updated the recommended settings and prompting tips above based on testing since the initial upload — should give you a much better starting point than what I had at launch.If you get interesting results (or run into anything weird), feel free to drop a comment — it'll help shape whatever comes next.

Files with SHA256: caf997a0df2f45d61a0f4f76cf42e878db4c68b5076b62c18d36dded7148f3af