> context length?

[hlchen23]

in the case of paper(cook egg), the model generate interleaved image-text which contains 5 images (~ 5K tokens), but chameleon is trained with 4K context. So I think it can not generate such good outputs, and I do experiments and find it cannot generate 5 images.

[EthanC111]

Hi @hlchen23, the context length for Chameleon/Anole is 4096 tokens. Extra post-training can increase the context length of the model.

[EthanC111]

Although the nominal context length is 4096, we discovered that the model can occasionally generate longer sequences (as demonstrated in the cooking egg example). However, the official context length remains 4096.