This is an Advanced Img2Img workflow for ComfyUI based on google's rectified flow inversion. It allows you to change your image, transfer style, inpaint, face swap etc... only by using prompts. No manual inpainting and controlnets.
To simplify: imagine basic img2img workflow, where you send input image -> make random noise from it -> send it to sampler and get output image, but in that case you can control what kind of noise you are getting from input image, depending on your settings. For instance, you can change either, part of image(inpainting) or transfer image style(ip-adapter).
So it looks like this: input image -> unsample image to noise (with additional "textual attention mask" a.k.a prompt input for img2noise unsampler. It should work good enough even without prompt, but you can specify what part of image you want to get "masked" for denoising) -> noise from unsampler goes to the sampler and you get output image edited with unsampled noise.
In other words it's almost like basic img2img wflow, but with additional "unsampling" step that provides more controled noise. [input image->unsampler(with optional prompt)->sampler(with standart prompt)->output image]
I recommend to use Hyper and Turbo merged models or loras of Flux, so instead of 28*2 steps you can do 12*2 with similar quiality.
GitHub - logtd/ComfyUI-Fluxtapoz: Nodes for image juxtaposition for Flux in ComfyUI
Some stuff from youtube:
(1) Turbo-Powered Unsampler & Inpainting with Flux.1 in SECONDS with ComfyUI! - YouTube
(1) Google RF Inversion: Image Editing with Prompting! - YouTube
More instructions provided in the workflow itself.
Description
FAQ
Comments (8)
This sounds amazing, and the paper suggests another set of powerful and useful methods. However, I'd suggest 'selling' the workflow here with more practical detail, to 'encourage' a greater number of people to give it ago.
The whole workflow concept strikes me as excessively 'technical' for the majority of potential users- which is part of the excitement of the entire punk-AI movement- the degree to which it challenges people.
Every new idea needs to get past that 'reluctance barrier' - especially with those of us not rocking a 4090, or 24GB 3090.
Thank you for your feedback! Yeah, i agree about all of this being "too technical", i sometimes find myself in that situations when it's easier to learn and make it yourself, rather than to understand what's going in someones 250nodes workflow. At this point i'm still exeperimenting myself, it's fairly new concept, but i think it has some real potential, so i'll simplify and update it as much as can.
I'm myself runing on 3080ti 12gb and this wflow actually easier on hardware than ip-adapter or\and controlnets, because you just don't need to upload them in your memory.
Thank you! This is a good workflow.
I feel dumb... Where is noise to image seed?
Thank you for kind words, it's near 1st sampler in "Inputs" tab right above "steps" node. Also there's no randomness involved in the 2nd sampler so there's no seed. It's pretty raw for now, but i'll update it in a few days.
Seems like something i would love to try, but i opened the workflow and i have no idea what to do... I couldn't make much of the notes unfortunately. Is there any chance you might explain the workflow a bit more?
This is the main problem with the punk-AI movement, although the need to 'skill up' will create many new AI engineers in the future. One problem is the 'easy' use of Python based code machines really only happens on Linux, an OS that many of us avoid like the plague for very good reasons. Using these tools on Windows is usually doable, but requires methods 'alien' to those only familiar with usual Windows methods for deploying code.
Youtube videos CAN help, but are often made by people unskilled in teaching.
OTOH, those nice people creating one-click apps on the back of these code machines usually limit the exposure of interface control variables to the point of uselessness in the name of simplification.
Still I think all authors here, LORAs or Workflows, would gain more enthusiastic users if they included a little essay talking about their project, rather than leaving potential users to break thru the obscurity themselves!
Hey! I uploaded v2 and edited description, feel free to ask if you have anymore questions.
@NexarCG Thank you so much!!