Didn't like other controlnets since they gave off distortions and couldnt copy pose, so i did my own.LLite model trained with various images scrapped from twitter and previous models datasets.Previews generated with pixel perfect copy resolution with "My prompt is more important option".To generate cond images script was used:<pre><code>from PIL import Image
import os
imgs = [img for img in os.listdir(os.getcwd()) if (os.path.isfile(os.path.join(os.getcwd(),img)) &amp; (os.path.splitext(img)[1] in ['.jpg','.png','.jpeg']))]

from controlnet_aux.processor import Processor
os.makedirs(str("roast"), exist_ok=True)
for img in imgs:
 temp = Image.open(img).convert('RGB')
 width, height = temp.size
 processor_id = 'lineart_realistic'
 processor = Processor(processor_id)
 temp = processor(temp, to_pil=True)
 temp.save('./roast/' + os.path.splitext(img)[0] + '.jpg', 'jpeg', quality=95)</code></pre>To make controlnet_aux package generate pixel perfect cond images file controlnet_aux\lineart\__init.py was modified at 138 line like following:<pre><code>...
 device = next(iter(self.model.parameters())).device
 if not isinstance(input_image, np.ndarray):
 input_image = np.array(input_image, dtype=np.uint8)
 h1, w1, _ = input_image.shape / starting here
 detect_resolution = min(h1, w1)
 image_resolution = min(h1, w1) / ending here
 input_image = HWC3(input_image)
 input_image = resize_image(input_image, detect_resolution)
...</code></pre>

Anima ControlNet Lite Custom

trained on anima-base-v1, LR = 0.000073, 7 epochs

Files with SHA256: a720985c000f4263f63519126f651499a590cfcd3f2dae87bb46cac148b57b96