This model is a fine-tuned version of facebook/opt-350m on the pszemraj/text2image-prompts-multi dataset. It achieves the following results on the evaluation set:

Loss: 1.6669
eval steps per second: 16.21
perplexity: 5.29

Example

The above example was created with DALL-E 2 but will of course work with any text2image model.

Intended uses & limitations

The model will generate augmentations that are biased towards the training data, i.e. what people already asked for in the SD/midjourney discords, etc. Creating a larger dataset was an attempt at mitigating this through more data from different datasets.

Training and evaluation data

See the pszemraj/text2image-prompts-multi dataset card for details. The dataset is a compilation of several text-to-image prompt datasets on huggingface :)

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0002
train_batch_size: 8
eval_batch_size: 4
seed: 42
distributed_type: multi-GPU
num_devices: 2
gradient_accumulation_steps: 16
total_train_batch_size: 256
total_eval_batch_size: 8
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_ratio: 0.04
num_epochs: 4.0

Training results

Training Loss	Epoch	Step	Validation Loss
2.1677	1.0	990	2.0888
1.856	2.0	1980	1.8215
1.6864	3.0	2970	1.6935
1.6228	4.0	3960	1.6670

Framework versions

Transformers 4.25.0.dev0
Pytorch 1.13.0+cu117
Datasets 2.6.1
Tokenizers 0.13.1