Why the result image is all black, when cfg-scale == 0 and temperature == 0 in Class-conditionalImageGeneration

[870572761]

Hi.
When I test Class-conditionalImageGeneration function in LlamaGen's GPT-L model.
When I use cfg-scale == 0 temperature == 0 in the model.

When I set cfg-scale 0, when temperature == 0

When I set cfg-scale 2, when temperature == 0

When I set cfg-scale 3, when temperature == 0

When I set cfg-scale 4 when temperature == 0

When I set cfg-scale 10 when temperature == 0

So when cfg-scale to 10 could make 360 class otter(Also maybe sth wrong in #76 )
Also I find when I set cfg-scale 0, 2, 3, 4, 10 when temperature == 1




Question 1: Why the result image is all black, when cfg-scale == 0 and temperature == 0 in Class-conditionalImageGeneration? Could I understand that the model depend heavily on the temperature random? Is it only when setting temperature != 0 could working?(Maybe > 1.0)
Also I notice your Experiments that model performance very bad when cfg-scale == 0. But I do not find temperature Experiments

[870572761]

It is my hyperparameter

python3 autoregressive/sample/sample_c2i.py \
  --vq-ckpt ./pretrained_models/vq_ds16_c2i.pt \
  --gpt-ckpt ./pretrained_models/c2i_L_384.pt \
  --gpt-model GPT-L \
  --image-size 384