Style Description<ul><li>highly-saturated postcard-like colors, flat high-key lighting, strong rim-lighting, 40s and 50s lifestyle</li></ul>Recommended Usage<ul><li>Your prompt must include "tchnclr style"</li><li>Use CFG of 7 or 8 for best results</li><li>The model was trained with and excels at closeup portraits of men and women</li><li>Try including "glitter" in your prompt!</li><li>Putting "b&amp;w" as a negative prompt will help ensure color image</li></ul>Known Limitations<ul><li>It strongly tries to insert 40s and 50s hairstyles, clothing, and scenery</li><li>As you can see from the examples, you can insert some modernity and blend with other styles. But if your prompt insists on all modern elements, the technicolor effect may disappear.</li><li>The model tends to turn men into women. It also likes to add hats!</li></ul>Training Process20 images from movies filmed in technicolor, 200 photo-like classifiers, 6000 steps, using the Dreambooth Extension for Automatic1111.

Technicolor-50s-Diffusion

New in version 2<ul><li>higher quality output</li><li>better match to the technicolor style</li><li>greater flexibility for diverse prompts</li><li>not strictly 1950's imagery</li></ul>Recommended Usage<ul><li>Add "tchnclr" and/or "1950s" to enhance the technicolor effect</li><li>The model was trained with and excels at closeup portraits of men and women</li><li>The model makes beautiful hair, and describing the hair style helps</li><li>Generating at 1024x with hires fix produces much better results than 512</li><li>Use CFG of 4-7 for best results. Yes, 4!</li></ul>Known Limitations<ul><li>Faces tend to look similar. To avoid this, prompt a celebrity at low weight or use an alternating prompt, e.g "[woman|celebrity-name]"</li><li>The more modernity that the prompt contains modern elements, the weaker the technicolor effect</li><li>Blending with other artists and art styles reduces the technicolor effect but is worth trying</li></ul>Training Process<ul><li>d8ahazard's Dreambooth Extension for Automatic1111 webUI</li><li>40 high-quality training images from movies filmed in technicolor, manually captioned</li><li>a version with 600 photo-like classifiers merged with a version with no prior preservation to balance style and flexibility</li><li>12,000 steps with a 2e-7 learning rate</li></ul>

Files with SHA256: 8835e696976ad35370c7682193ebfcc056728a19776d6dd1496ce6675d62c775