DALL·E 3 Image Generation Test

---

• Open AI Image Generation Docs
• Kaggle Version of this notebook

General Notes

---

• Use Google Colab with this notebook for best results.
• An Open AI API key is required to use this notebook.

Image Settings

---

• Image sizes: 1024x1024, 1024x1792 or 1792x1024 pixels.
• Quality Settings: "standard" for basic detail, "hd" for enhanced detail.
• Square, standard quality images are the fastest / cheapest to generate.

Some Take Aways

---

• You are only able to render one image at a time with Dalle-3
• There is no way to seed this model at the moment or feed a previous image into it. That means that fine-tuning the model is not possible like Dalle-2
• Each generation of an image creates a new prompt on the backend for moderation. Even if you don't change any parameters, the model will create a newly generated token string and then generate the image from that.

Conclusion

---

• After some testing, the model won't accept a string of tokens, a list of numbers, or an object of tokens. It throws the security warning each time.
• With the modeeration aspect of the process, the act of rewriting of the prompt with each injection of the prompt doesn't allow for repeatablility. From my research, there is no real way to get relaible and repeatable results from the model.
• The hope is they open this API like they did with DALL·E 2 to allow for a more iterative approach without altering the state of the prompt with each request.