Spaces:
Running
Can Boris see what is being generated behind the curtain? Can/Should we share our text prompt learnings?
Thank you to Boris and HF for making this model so very accessible. I'm having a lot of fun playing with this.
I wonder if Boris or HF can see what is being generated behind the scenes?
I'd love to see a button that allows me to share what is generated with the community of users. I'd like to develop more understanding of how text prompt engineering can provide greater control over the output. I wonder if there is a way to share our prompts and 'crowdsource' a collective understanding of how Dalle-mini responds to text prompts? Of course, moderating garbage or low value text prompts would not be a trivial process. Also text prompt engineering skills are what differentiates one user from another in an otherwise zero barrier to entry and so some users may not be willing to share that information.
At the moment I'm using 50+ word prompts and manually testing different combinations. ~30 words of scene description and 5+ artist's names. Changing the position of the artist's name, e.g. moving to the front of the prompt, changes the output style significantly. Some artist's names, eg Zdzis艂aw Beksi艅ski, seem to 'overwhelm' the model and totally swamp the output in one specific style. I'm guessing this is due to some kind of weighting bias inside the model.
I find that this method of providing a large number of scene descriptors gives me a pleasing number of different output scenes still within a kind of single cohesive style e.g. the stylistic love child of HR Giger and Zdzis艂aw Beksi艅ski and Salvador Dali
I find that I can get stunning results in this model with very long prompts, combining many scenes, concepts, styles, etc.
The most powerful keyword I've found so far is 'detail'. Adding the word 'detail' focuses the attention of the model on whatever you want. e.g., 'detail face' means you want to focus on drawing a face properly.
Folks should realize that Dall does understand abstract concepts to some degree, so you can describe the type of art you want to see, and then focus the attention of the model on this description using the keyword 'detail'. For example, if you want 'beautiful art', simply say so! And add the word 'detail' to get the model to focus on that and enhance!
I use dalle for art prompts. As an artist doing a PhD, I have a LOT of ideas, but need to get them out of my galaxy brain so to speak. Sketching is time laborious, so this tool helps me generate images of what I am envisioning in my brain - which I can then use to help me create my photography/sound/installation/textile and performance works. The tool is forming part of my research practice methodology and I'll be citing the hell out of it :)
I have found using different word synonyms, a combination of words meaning same thing, word choice order, commas, and words that mean something visually similar can refine choices. I.e. I used the words cape, skirt, curtain and veil to generate similar conceptual frameworks for image creation.