<h3 id="moss">MOSS</h3>Starting to experiment with moss training<h3 id="krea-2">KREA 2</h3>Showcase images :<ul><li>Turbo model : 10 steps / CFG 1 &amp; Euler simple</li><li>944x1408</li><li>x2 Hires (10 steps, 0.35 with 4x_NMKD_Siax)</li><li>Lora strength 1.3</li></ul><h3 id="ideogram-4">IDEOGRAM 4</h3>Json caption style with :<ul><li>art_style : Naoki Urasawa style</li><li>medium : manga</li><li>aesthetics : black_and_white</li></ul><a target="_blank" rel="ugc" href="https://pastebin.com/KJ4fG954">Example caption</a><h3 id="klein-9b">KLEIN 9B</h3>v1:Showcase images :<ul><li>Distilled model (same prompt images posted with the base model below)</li><li>768x1152</li><li>Euler / normal / 4 steps / CFG 1</li><li>Hires steps : 4 / NMKD_SIAX @0.35 / x2</li><li>Lora strength of 0.8-1</li></ul><h3 id="zit">ZIT</h3>v1:I'm relatively new to Z-Image turbo model, so still experimenting, but I had good results with these settings :<ul><li>1056*1584 // 10 Steps // CFG 1</li><li>DDIM &amp; Beta</li><li>Lora strength of 0.9</li></ul>I've noticed that using previous prompts (like the ones from Qwen v2) doesn't work that well, so I've used a vision model (Qwen3-VL-8B) with <a target="_blank" rel="ugc" href="https://pastebin.com/EcuRaPaP">this prompt</a> to get updated caption which was giving way better results.<h3 id="qwen">Qwen</h3>v2:Same as for my <a target="_blank" rel="ugc" href="https://civarchive.com/models/523485/ghibli-style-kikis-delivery-service?modelVersionId=2540186">ghibli lora</a>, I've retrained with everything I've learned so far, and this version is way better : closer to the target style and more flexible.v1:Trained with Ostris Toolkit: Dataset is 44 images (mainly from "Monster", "20th Century Boys" and "Pluto" by Naoki Urasawa) with no trigger word (but sometimes it helps to add "black and white manga style image of"). All the image have been generated in 4 steps in addition to the <a target="_blank" rel="ugc" href="https://huggingface.co/lightx2v/Qwen-Image-Lightning">lightning lora by lightx2v</a> The prompt adherence of Qwen is really impressive, but I feel like the likeness of the style is not as good as with Flux (even tho is still good), but I'm still experimenting so it's just an early opinion.<h3 id="flux"> Flux</h3>Trained on images from the manga "Monster" and "20th Century boys" by Naoki Urasawa.Start the prompt with :<h2 id="a-naourasawa-style-image-of">A naoUrasawa style image of</h2>Using <a target="_blank" rel="ugc" href="https://openmodeldb.info/models/2x-Manga-Ora">this upscaler</a> seems to improve the style towards a true manga scan most of the time (and get rid of colors). I've used it with hires fix and denoising strength of 0.5.

Manga style (Naoki Urasawa)

naoki_urasawa_moss.safetensors

Moss v1

Anima v1

Krea2 v1

Ideogram4 v1

Klein 9B v1

ZIT v1

Qwen v2

Qwen v1

Flux v1

Files with SHA256: edd11ccdf6baf9c948622d4cb7e8aa114613fcbcb1431e3bb9b09f4c996c05d7