a question about inference, no "output_img_token" (playground.py)

[FromA2Z]

Regarding the inference effect, in my experiment, there was only [IMG2] in the generated text, but not [IMG0]. I wonder if this is normal. Because [IMG0] means "output_img_id", without it you cannot enter the image generation branch. Looking forward to your answer, thankyou。

[KzZheng]

Yes, the model will generate an image only when [IMG0] is predicted. This error may be caused by the newest transformers library (They update the tokenizer), Maybe you can try transformers=4.31.0. Thanks!

[FromA2Z]

Yes, the model will generate an image only when [IMG0] is predicted. This error may be caused by the newest transformers library (They update the tokenizer), Maybe you can try transformers=4.31.0. Thanks!

Thank you for your answer. It was indeed a version problem. I changed the version of the transformer and the problem was solved.