In the Dark - LORA (Blades in the Dark Style)版本beta2 (ID: 85958)

The world ended, a long time ago. The sun cracked, the gates of death broke, and the earth became a cursed world. Still, humanity survives. Stuck in the crowded slums of the lightning-caged cities, your only hope for a decent life is to seize it for yourself. By blade or bullet; by a charming word or occult incantation: you will take this city or die trying, your broken body lying in the dark.

Update 23/6/13:

A smaller update. I fixed captioning issues and stretched out the dataset a bit. The biggest changes were from hyper-parameter optimization, so the model is both smaller and more true to the original style at high weights (with the right checkpoint of course).

Update 23/5/30:
I fixed a lot of issues, especially with the captioning. The model can now reproduce the original style at higher weights, and is overall more flexible. I will need to do more testing and figure out some way to expand the dataset before publishing a final version, but this update should be a major improvement.

Notes:

• The trigger world bitdstyle is now impactful.

• The recommended weights are lower. 0.6 - 0.8 should work well, although higher weights may give you interesting effects.

• The monochrome effect has been somewhat disentangled from the general style, but you can guarantee it by adding "monochrome" to the prompt.

Original Post 23/5/28:
Hey folks,

This model is based on the art of the Blades in the Dark tabletop roleplaying game. This is an alpha release on a limited dataset, but I wanted to try applying the style to anime models. Right now, the LORA feels a bit overtrained in some areas and undertrained in others, but I feel that it offers a cool monochrome style. I decided to release what I have so far and then work on an improved version down the line.

This was trained on animefull, so it should have an impact on derived CLIP SKIP 2 models and merges. I tested with and was pretty happy with the results using Counterfeit V3. To get the monochrome effect, I recommend a weight of 0.8 - 1.0. There's no explicit trigger word, but using "silhouette" should get you closer to the style.

Thanks,
Machinique

描述:

Fixed a lot of issues when it comes to captioning. Fixed the trigger word, which is now quite impactful. The model has very different behavior than the alpha version, but it's more accurate to the original style and should be more flexible/responsive overall. The dastaset was increased the from 24 images to 36. The monochrome feature has been somewhat dissociated from the style, but you can guarantee it by adding "monochrome" to the prompt.

New recommended weights are 0.6 to 0.8, although you might be able to get some interesting silhouette effects at higher weights.

训练词语: bitdstyle,monochrome