A compact and streamlined workflow for Flux2 Klein that allows for the use of up to 10 input images.
The first input tends to hold the most strength in the output, and controls the output aspect ratio and resolution. So it's best to consider the first image as the image you're editing with the rest being conditioning inputs, although they can overpower the first image entirely, depending on a variety of factors.
Basically, this configuration can be used to create results you might not otherwise be able to get or get as easily in one go, but also may not give you the result you're looking for at all. It can act somewhat like Instant-ID or Face-ID controlnets or like a concept lora when it comes to people and objects, but it's probably more useful when the inputs are used as a reference for different elements. Even different orders of the same inputs can affect the result, so it's more tinkering and randomness than an exact process, due to the increased variables.
So who is this workflow for? People who want to experiment with randomness, or people who are willing to put in the effort to be very intentful with every aspect of their generations to get the results they want. For most tasks though, 2 image inputs will be more controllable and useful.
The multiple "Megapixels" settings are for each input image. The behavior is that the input image will be scaled to 1 megapixel and then processed, which also affects output quality and size, so it's like an input/output resolution slider. I generally keep them all the same number, but different megapixel numbers for inputs can have an effect. "Global Megapixels" is just for if every input is disabled and you want to do TXT2IMG only.
Description
• Just like the Multi Input workflow, except the amount of inputs are doubled to what is supposed to be the maximum amount of inputs allowed with Flux2 Klein
FAQ
Comments (3)
Wow, this is really convenient — it’s simple, but still has everything you need. I’ve tried it a few times and it’s a great workflow file.
One question though: would it be possible to add inpainting using a mask node, even if it’s only for outpainting or just for Input 1 (Image 1)? Everything else is excellent, but having outpainting at least would make it perfect.
I'm not sure how to implement inpainting because of the way nodes are setup. When I tested it with masking, it simply still used the reference image + the mask, which basically resulted in bleed between the mask and the reference image. So you can see the mask, but it looks more like smudges over the reference image, rather than replacing the masked area. If I disable input 1 as a reference and do masking on it, that works, but not well. Even though it might seem interesting, inpainting is really not designed for this.
It also makes less sense with the Klein model, considering one can simply prompt something to be replaced seamlessly. It would probably be best to do inpainting in another more singularly focused workflow, as it won't really work like you might hope.
One thing you can do if you want to try to mask multiple things into an image at once, is to simply create a collage by pasting images directly into a main image, then prompt for something like realism or prompt someway to "fix" the image. The same method worked with Flux Kontext, where you could paste an image onto an existing image and prompt it to "make more realistic" to blend them seamlessly.
So you don't necessarily need masking at all, and if you do, it's probably best done as a pre-process before using it as an input to this type of workflow. You could even for example, just fill in parts of an input image with white and then prompt it to fill in those areas. So you can sort of do what you might want, just in more basic ways.
I don't even think 10 inputs would work like some people might hope, but it's still useful to have slots for the inputs so that people can use only a couple images at a time and toggle between them on the fly. Using multiple inputs of the same image can also strengthen an image in the workflow. For example, if you use a base image and then use the same reference image 3-4 times of what you want to add, it can blend a person into a scene while essentially keeping the full reference image look/pose, but adjusts it to look like the lighting/picture style of the base image.
An absolutely God Tier workflow.
amazing results, thank you