Main prompt tags:<pre><code>
seiichikinoAP, 1boy, glasses, blue polo shirt, anime coloring </code></pre>Style control:<ul><li>for the original anime TV show style add following positive prompt: anime coloring</li><li>to avoid the style from the anime show, remove anime coloring from positive prompt and put it into negative prompt</li></ul> Recommended samplers: - Euler Ancestral Scheduler: Simple CFG: 4.0 Steps: 30 Lora strength: 1.0 Prompt guide (green - positive, red - negative): - anime coloring - most important tag to keep the original style - chubby/fat - if your model doesn't generate him fat enough - glasses, blue polo shirt, pants - base clothes - white shirt, red polo shirt, light purple polo shirt, grey newsboy cap, white and green striped shirt, blue undershirt, brown coat, cowboy hat, bandana around neck - additional canon clothes - stubble, messy hair - messy version - 6 fingers, six fingers - I think Anima preview sometimes has issues with too many fingers, not sure if this actually helps The LoRas should work with Anima checkpoints. Tested working well with the following:<ul><li>AnimaYume v0.1 [Preview1]</li><li>AnimaYume v0.25 [Preview2]</li><li>AnimaYume v0.4 [Preview3]</li></ul> Last image in the gallery can be imported into ComfyUI with a basic workflow: <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/fdb51547-09f9-478e-8da1-27b857c5d672/original=true/fdb51547-09f9-478e-8da1-27b857c5d672.jpeg" /> No custom nodes required.

Seiichi Kinoshita - Anima Preview1/2/3 (SHIROBAKO)

Re-trained the same dataset for AnimaPreview2.My thoughts on Preview2 model: - better prompt adherence - better overall quality when used with character Loras - don't need to use fine-tunes such as animeYume - better text rendering - I was able to get correct text generated much quicker Overall really happy with Preview2, its a direct upgrade over Preview1 and hopefully fine-tunes will push the quality even further.

SeiichiKino_AnimaPreview2.safetensors

AnimaPreview2

AnimaPreview3

AnimaPreview1

[File reupload cuz Civit is being buggy as hell]Re-trained the same dataset for AnimaPreview2.My thoughts on Preview2 model: - better prompt adherence - better overall quality when used with character Loras - don't need to use fine-tunes such as animeYume - better text rendering - I was able to get correct text generated much quicker Overall really happy with Preview2, its a direct upgrade over Preview1 and hopefully fine-tunes will push the quality even further.

Files with SHA256: dae43bf1670eaa933adc8924196df0f7365e71af10c3582f73d2e698afe7c70f