Thats true! Since nat lang text encoders are more complex , the negatives of the stuff being encoded is rarely the opposite.
As in , the model itself (Chroma on perchance) isnt trained to comprehend negative vectors. Tho lodestones (creator of Chroma) never specified but I assume not.
Negatives are an off shoot in training that sorta worked in CLIP based models SD1.5 SDXL (Pony / illustrious) that carried over into the nat lang mod releases out of habit in the community.
It worked in CLIP models because the CLIP encoder is simple. Write ‘ice cream’ in CLIP and the text encoding vector will point roughly same direction no matter where in the 75 token batch ‘ice cream’ .
Compare that to the many number of different answers you can get from chatGPT or GROK containing the word ‘ice cream’ and you can see how the 512 token batch encoding of the T5 in Chroma , or Qwen encoder in Klein / Z image varies drastically depending how common words are arranged in the text.
Gone, because negative prompt is not working
Thats true! Since nat lang text encoders are more complex , the negatives of the stuff being encoded is rarely the opposite.
As in , the model itself (Chroma on perchance) isnt trained to comprehend negative vectors. Tho lodestones (creator of Chroma) never specified but I assume not.
Negatives are an off shoot in training that sorta worked in CLIP based models SD1.5 SDXL (Pony / illustrious) that carried over into the nat lang mod releases out of habit in the community.
It worked in CLIP models because the CLIP encoder is simple. Write ‘ice cream’ in CLIP and the text encoding vector will point roughly same direction no matter where in the 75 token batch ‘ice cream’ .
Compare that to the many number of different answers you can get from chatGPT or GROK containing the word ‘ice cream’ and you can see how the 512 token batch encoding of the T5 in Chroma , or Qwen encoder in Klein / Z image varies drastically depending how common words are arranged in the text.
Quite sad… It’s being harder than ever to create something that doesn’t look either a loli or a bimbo, and both are often naked or almost…