ok, after lots of testing, here is the best way I've found to create consistent characters in Midjourney.
It starts with a model, and it seems to work best with a head and shoulders portrait photo like this one. Prompts in ALT if you wanna test this out yourself
𧡠Thread
It starts with a model, and it seems to work best with a head and shoulders portrait photo like this one. Prompts in ALT if you wanna test this out yourself
𧡠Thread
Once you have your model, make sure you have the --seed for the image you're going to use. You can either add --seed 12345 (or any number you'd like) at the end of your prompt while iterating, or you can react to the image u want to use with the :envelope: emoji to find the seed
Next, copy the image URL and start a new prompt. In your prompt, include the URL, and the same prompt + seed you used to generate the image. Like this
/imagine [URL] 1960s street style photo of a young woman, latina, dior, silk, diamonds, dress [...] --ar 16:9 --seed 1622
/imagine [URL] 1960s street style photo of a young woman, latina, dior, silk, diamonds, dress [...] --ar 16:9 --seed 1622
You should already start seeing some interesting variations.
Now, try changing up some of the variables. Here i added "sitting" to the prompt.
Now, try changing up some of the variables. Here i added "sitting" to the prompt.
Here I changed the location to "Central Park"
Here we went to the beach
And let's finish with a nice sunset at the beach
Feel free to remix this and reply with some of your best side by sides. Have fun, and lmk if you have any questions.
I'll be going into this in more depth on Youtube, and pushing the limits of this technique as much as I can.
I'll be going into this in more depth on Youtube, and pushing the limits of this technique as much as I can.
Some notes:
-I've found models that have a more unique/distinct look are more likely to result in consistent results.
-keeping the variables similar to the prompt used to generate the original image helps
-you don't have to use the same seed, but i find it helps
-I've found models that have a more unique/distinct look are more likely to result in consistent results.
-keeping the variables similar to the prompt used to generate the original image helps
-you don't have to use the same seed, but i find it helps
more notes:
-try changing clothing, camera perspective, lighting, or location
-try using variations with remix mode turned OFF. I find remix mode results in more drastic changes, while regular variations are great for creating unique changes in things like clothing
-try changing clothing, camera perspective, lighting, or location
-try using variations with remix mode turned OFF. I find remix mode results in more drastic changes, while regular variations are great for creating unique changes in things like clothing
I've ran a lot of tests, like many hundreds of tests, and found lots that worked, and even more that didn't.
I'm working on a series for this based on the results from the poll. Full process breakdown will be up soon on yoututbe (link in bio)
I'm working on a series for this based on the results from the poll. Full process breakdown will be up soon on yoututbe (link in bio)
Here is a side-by-side with different clothing. Using additive prompting + this technique you only need to change a couple of words in your prompt to get some really great results.
this is black silk, diamonds / green silk, pearls
this is black silk, diamonds / green silk, pearls
you can also try this with your own portrait, or a stock image like these.
you might not get as much consistency, but with the right prompt & enough variations, you'll get some good ones. Keeping it simple is π
Here are 2 examples using portraits I pulled from upsplash.
you might not get as much consistency, but with the right prompt & enough variations, you'll get some good ones. Keeping it simple is π
Here are 2 examples using portraits I pulled from upsplash.
Just made a new model. DM me if you wanna have a go at training a model in stable diffusion with these and i'll send you the images.
@TomLikesRobots i have an idea for an experiment I could use your help on if you're down.
@TomLikesRobots i have an idea for an experiment I could use your help on if you're down.
I'm currently in the process of casting some new models. Here are my favs from a new series I'm working on called Sequin & Chandeliers
BTW, this entire workflow is mobile.
Anyone w/ a phone & a discord account can create compositions like these.
Think about how powerful that is. Thatβs why I focus on midjourney rather than Stable Diffusion
I hope my work can inspire others to let their imaginations run wild.
Anyone w/ a phone & a discord account can create compositions like these.
Think about how powerful that is. Thatβs why I focus on midjourney rather than Stable Diffusion
I hope my work can inspire others to let their imaginations run wild.
Loading suggestions...