example_inference.py script yields weird results

[pbruneau]

I just tried to run the example script after having installed dalle2-laion (I just changed setup.py so that dalle2-pytorch==1.1.0 is installed, to make downloaded models happy with the pytorch version in place),

but the results I get seem a bit odd - e.g. here is my stack trace for running the script:

# python3 example_inference.py dream
Enter your prompts one by one. Enter an empty prompt to finish.
Prompt 1: a cat playing piano
Prompt 2: 
How many samples would you like to generate for each prompt? [1]: 4
Downloading https://huggingface.co/laion/DALLE2-PyTorch/resolve/main/decoder/v1.0.2/latest.pth to models/new_decoder.pth
Downloading https://huggingface.co/Veldrovive/upsamplers/resolve/main/working/latest.pth to models/second_decoder.pth
WARNING: This decoder was trained on an old version of Dalle2. This may result in the model failing to load or it may lead to producing garbage results.
Downloading https://huggingface.co/Veldrovive/upsamplers/raw/main/working/decoder_config.json to models/second_decoder_config.json
Downloading https://huggingface.co/laion/DALLE2-PyTorch/resolve/main/prior/latest.pth to models/prior.pth
WARNING: This prior was trained on an old version of Dalle2. This may result in the model failing to load or it may produce garbage results.
Downloading https://huggingface.co/laion/DALLE2-PyTorch/raw/main/prior/prior_config.json to models/prior_config.json

And here is an example of results I'm getting: https://ibb.co/SmpTFHs

Is this normal/expected? Am I missing something?

[Veldrovive]

Yep, this just looks like what you can get right now. Both the initial decoder and the upsampler are undertrained and undersized. You'll see better results if you come back in a month or so.