Turn Krea 2's photorealism into clean webtoon / manhwa panels with one slider. 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.

How to use

• Trigger: put sushi_tv at the start of your prompt.
• Strength is the style dial: 0 = the photo baseline; around 1.0-1.4 you get a stylised, illustrated-photo hybrid; the full webtoon flip kicks in around 1.6, 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.
• Load it model-only (LoraLoaderModelOnly) - no clip patch needed.
• Recipe used for every showcase image: Krea 2 Turbo (int8), 8-10 steps, cfg 1, er_sde sampler + simple scheduler.
• 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.

Pick your checkpoint

• run 156 - step 1500: the very first pass - softest, painterly hybrids at 1.2-2.0.
• run 158 - step 1000: subtlest of run 158, keeps the most photo DNA.
• run 158 - step 1500: the balanced favourite of our tests.
• run 158 - step 1750: our reference pick - its gallery shows same-seed strength triptychs (0 / 1.4 / 2.0).
• run 158 - step 2500: strongest style commitment of run 158.
• run 160 - steps 1000 / 1500 / 2500 / 2750: the v2 retrain, richer shading - the later the step, the earlier the flip (2750 converts fully at strength 1.0).

Training

Trained on Krea 2 Raw (each run shares one dataset; run 160 is a v2 retrain). Dataset prepped, captioned and trained end to end with LoRA Dataset Studio, our free open-source dataset-to-LoRA app.