Creating Variations
Lesson Goals
- Create AI variations of any photo
- Create a higher resolution image
- Learn about ControlNet Modes
Step 1: Render an image
First provide a base image. You can upload an image or render an original one, like this:
/render a close-up photo of men's shoes
Optional: Reply to the image with /showprompt to see what seed, guidance, model, and sampler were used. These are selected at random unless specified.
Step 2: Try the “more” command
Right click on the image and choose reply, as if you were about to talk to it. Then type the more command
/more
On a technical level, what “more” does it look at the parameters (in the optional showprompt step mentioned above) and increase them slightly to give you more of the same prompt. You can also use the /more command to upscale the image.
The More command is deceptively simple but incredibly powerful. In fact, we recorded an entire hour of video about it!
Consider these optional parameters for /more
Strength | Determines how much the image should change. The max value is 1. The smallest value is 0.01. So for example, /strength:0.2 will give us an image that is roughly only 20% different than the original, where as /strength:1 will give us something completely different. |
Guidance | Guidance aka CFG is how strictly to follow your prompt. The default range is 6-10 unless specified. Very low guidance can look blurry or surreal, and very high guidance can look sharp but may introduce glitches if the AI doesn’t have enough reference material to fill in the image. If you get blue boxes or artifacts, try lowering the guidance. |
Size | You can use the /more command as a handy upscaler, but you’ll want to reduce the image count or the request will take too long.
Putting it all together: /more /size:1400×1400 /images:1 /strength:0.2 /guidance:7 |
But what if we wanted to introduce a more radical change, and prompt for some changes?
Step 3: Try the “remix” command
Remix is similar to more and uses the same parameters, but it requires a second prompt. It is also known as IMG2IMG in other Stable Diffusion front-ends. You can also pass a different base model (art style) to the remix command to do a style transfer. This has many interesting use cases. For example, you can convert a realistic image into an illustration, and vice versa.
Remix is reply command. You’ll first need to right click on the image to select it, then Reply to it as if you’re having a conversation. Then type:
/remix a pair of clown shoes <sdxl>
In the example above, we’re turning a normal set of shoes created in Stable Diffusion 1.5 into clown shoes in the art style of SDXL
Remix is a destructive command, meaning it will change every pixel. To preserve pixels (such as not changing the face), consider drawing a mask with Inpaint instead.
Controlnet
Once you’ve mastered More and Remix, we encourage you to continue mastering IMG2IMG commands using ControlNet. ControlNet is interesting because you can save a preset image input, and then choose whether you want to trace its edges, steal its pose, copy it’s depth map, swap just the face, and more.
Check out the full tutorial on ControlNet
Glossary
Render – The basic image creation command. Pair it with models to change the art style
More – Creates more images without changing the prompt. It does this by increasing the seed value and changing the guidance and sampler slightly
Remix – Also known as Img2Img. You are uploading or creating an image first, and then “remixing” it into something new. None of the original pixels are preserved.
Guidance – A number between 0.50 and 20 that tells the AI how strictly to follow the prompt. A low guidance is considered under 7, and high above 12. If you are getting artifacts, lowering guidance helps.
Seed – A random number assigned to a generation, for the purpose of repeating the result. A seed by itself doesn’t mean a specific image, it is more like an anchor. When the exact same prompts, the exact same guidance, samplers, and other settings are used, the same image can be generated by another person on a different computer.
ControlNet – A suite of tools that create image variations, by using an image as an input. You can input a photograph, a pose, a contour, and create variations.