How do prompt modifiers impact #dalle2 image quality? I've run ~600 prompts with random modifiers and ranked the results 1-5. Here are my findings! 🧵
Unsurprisingly, "Trending on artstation" seems to have the biggest impact. On average I've rated those 2.9 (compared to 2.0 for prompts with no mods)
* "rendered in octane": 2.4
* "vfx": 2.3
* "elegant": 2.2
* "digital art", "control the soul", "rendering", "wondrous", : 2.1
Mods that appear to make the results worse include:
* "contest winner", "wormhole", "magnificent", "detailed" : 1.7
* "8k", "stunning": 1.8
* "fantastic", "#wow": 1.9
Now of course this is all subjective and in many cases may not be statistically significant, YMMV. Did I miss any good modifiers that you love to use? Let me know!
• • •
Missing some Tweet in this thread? You can try to
force a refresh
@mattyglesias As one of the authors of CHIME I'm biased. We've done some internal comparison and found that the IHME model is a serious outlier. CHIME predicts the highest numbers. Due to the simplicity of the SIR modeling approach. Over short periods all models are equivalent. 1/n
@mattyglesias As @zeynep says "Don’t Believe the COVID-19 Models, That’s not what they’re for." CHIME is built for short-term (2-3 week) planning. On these short timelines, all existing models have similar results. However I would not trust the IHME model 2/n
@mattyglesias@zeynep "All models are wrong but some are useful." IHME model uses arguably flawed death data out of Wuhan and Italy and a non-standard epidemiological approach. Assumes the best case scenario (Wuhan style lockdown). CHIME assumes simple/standard SIR model which overestimates 3/n