yodayo-ai

/

kivotos-xl-2.0

like 100

So far feels like a considerable downgrade from 3.1, poor consistency, less prompt following. Tried several previous prompts, it's either something off with my build and this one needs some tricky setup or model just came out poorly. Uhh. I'm really sorry about that.

Alright, sorry, the last one was rude. Let me explain with a clear head. 🐧

• Firstly, dynamic pose tag is not ideal to start with, as there are only 610 images with this tag in Danbooru.
• Also, Animagine XL V3 does not fully cover Danbooru, so the results may be unexpected.

I've already written this section that explains how to prompt better based on how we train the models. This prompt format is incorrect:

masterpiece,best quality,very aesthetic,absurdres,animal ears,
rio (blue archive),1girl,white background,
full body,ke-ta,as109,dynamic pose,

To prompt the model correctly, please follow this format:

1girl, rio \(blue archive\), blue archive, ke-ta, as109, white background, full body, animal ears, masterpiece, best quality, very aesthetic, absurdres

• Start with gender tags, followed by character tags, series tags, artist tags, general tags, and finally, special tags (quality tags, aesthetic tags, meta tags, etc.).
• But, even with the correct format, this prompt may not yield the best results. You need better context; at least let the model know what and where.
• Unless you have DanTagGen or DartV2 to upsample the incomplete prompt, full body has not been a good tag since Animagine XL V1.
• I recommend using a cowboy shot or visit this image composition wiki.
• I haven't tested prompts with multiple artists, so the results may be unpredictable.

Finally, this is at least a good prompt with better context, still needing tweaks here and there:

1girl, rio \(blue archive\), blue archive, ke-ta, as109, animal ears, cowboy shot, indoors, masterpiece, best quality, very aesthetic, absurdres

{
  "prompt": "1girl, rio \(blue archive\), blue archive, ke-ta, as109, animal ears, cowboy shot, indoors, masterpiece, best quality, very aesthetic, absurdres",
  "negative_prompt": "nsfw, (low quality, worst quality:1.2), 3d, watermark, signature, ugly, poorly drawn",
  "resolution": "896 x 1152",
  "guidance_scale": 7,
  "num_inference_steps": 28,
  "seed": 1662602690,
  "sampler": "Euler a",
  "add_quality_tags": true,
  "quality_tags": "Standard",
  "use_upscaler": null,
  "Model": {
    "Model": "Kivotos XL 2.0",
    "Model hash": "e3c47aedb0"
  }
}`

Let me know if you have any questions or need further assistance.