search

eps696 / aphantasia

CLIP + FFT/DWT/RGB = text to image/video
MIT License
775 stars 102 forks source link

Draw only one object? #19

Closed 8KV closed 3 years ago

8KV commented 3 years ago

.

eps696 commented 3 years ago

alas, that's a systemic feature, the downside of high resolution rendering. since the optimization with CLIP is limited to the model resolution (224 or 288 px), the image is being drawn by separate pieces (patches) all over the frame, and the patches know nothing about the distant others, hence the repetitions occur. you may try changing the parameters - progressive_grow, no_text, enhance (and/or noise on continuous mode Illustra script) all somehow affect the result; but there's no guaranteed way at the moment.