<p><strong>Turn Krea 2's photorealism into clean webtoon / manhwa panels with one slider.</strong> sushi_tv is a style LoRA: keep your prompt, raise the strength, and the same seed goes from photo to a full webtoon rendering - panel borders, caption boxes, glossy cel shading. It works on people, couples, groups, animals, creatures - SFW and NSFW alike.</p>
<h3>How to use</h3>
<ul>
<li><strong>Trigger</strong>: put <code>sushi_tv</code> at the start of your prompt.</li>
<li><strong>Strength is the style dial</strong>: 0 = the photo baseline; around 1.0-1.4 you get a stylised, illustrated-photo hybrid; <strong>the full webtoon flip kicks in around 1.6</strong>, depending on the checkpoint - late checkpoints flip earlier (run 160 step 2750 is fully converted at 1.0), early ones want 1.8-2.25. Useful range tops out around 2.25-2.5.</li>
<li>Load it <strong>model-only</strong> (LoraLoaderModelOnly) - no clip patch needed.</li>
<li>Recipe used for every showcase image: Krea 2 Turbo (int8), 8-10 steps, cfg 1, er_sde sampler + simple scheduler.</li>
<li>On wide 16:9 canvases past the flip point it loves to produce full manga PAGES (multiple panels, speech bubbles with made-up lettering). Lean into it, or stay on 9:16 for single panels.</li>
</ul>
<h3>Pick your checkpoint</h3>
<ul>
<li><strong>run 156 - step 1500</strong>: the very first pass - softest, painterly hybrids at 1.2-2.0.</li>
<li><strong>run 158 - step 1000</strong>: subtlest of run 158, keeps the most photo DNA.</li>
<li><strong>run 158 - step 1500</strong>: the balanced favourite of our tests.</li>
<li><strong>run 158 - step 1750</strong>: our reference pick - its gallery shows same-seed strength triptychs (0 / 1.4 / 2.0).</li>
<li><strong>run 158 - step 2500</strong>: strongest style commitment of run 158.</li>
<li><strong>run 160 - steps 1000 / 1500 / 2500 / 2750</strong>: the v2 retrain, richer shading - the later the step, the earlier the flip (2750 converts fully at strength 1.0).</li>
</ul>
<h3>Training</h3>
<p>Trained on <strong>Krea 2 Raw</strong> (each run shares one dataset; run 160 is a v2 retrain). Dataset prepped, captioned and trained end to end with <a target="_blank" href="https://github.com/perfectgf/lora-dataset-studio" rel="ugc">LoRA Dataset Studio</a>, our free open-source dataset-to-LoRA app.</p>

Krea Webtoon Converter (NSFW)

Checkpoint from run 156, the very first training pass. Softer, more painterly linework than run 158 at equal strength - both showcase images run at 2.0.

Files with SHA256: 17530c0cc0adbcd0cb567e090be5dc0637fad39da82bc641ec52b6889a238d1f