Magic Prompt Kiss
I have always talked about the dreamability of Midjourney, because of prompt expansion under the hood. Someone implemented GPT-2 expansion as part of the Dynamic Prompts extension in Automatic1111.
Gustavosta’s MagicPrompt model is Trained on 80,000 prompts from Lexica.art, and will do prompt expansions using very simple prompts. All of these images startw ith something intentionally simple:
two muscular men, gay couple
But the extension added other keywords: gay couple, shorts, kissing, wet dripping, volumetric lighting, godrays, vivid, trending on artstation, anime art style
This changes for every generation, and thus you get a wide range of results based on very simple prompts.
I have often talked about how Midjourney does a lot of things behind the scenes — and this is exactly what I mean. Midjourney additionally uses a few tricks that I can see, which are hard to articulate, so I won’t elaborate here. But in a gist, I could see that it is doing some type of clustering besides GPT expansion.
These are technical tests, so besides Lightroom color correction, no additional fixes have been applied. They’re also rendered at extremely low settings, mostly so I can see what the extension does.
- Stable Diffusion txt2img
- Dynamic Prompt Magic Prompt
- Virile Fusion v3 Beta 1
- Euler 10 steps
- Hires 20 steps 1.5x
- ADetailer 10 steps
- Gigapixel AI 4x
- Adobe Lightroom
Relevant Resources:
- SD Dynamic Prompt extension
- MagicPrompt - Stable Diffusion. This is a model from the MagicPrompt series of models, which are GPT-2 models intended to generate prompt texts for imaging AIs, in this case: Stable Diffusion.
