Feature Request: Negative ControlNet Model

This may be a simple project for a someone who can train ControlNet models, or it may require the help of the devs, or both (or either). I don't understand the system well enough to say, but I wanted to get the idea out there, for people to consider. Please forgive me if this idea is impossible, as I lack a deep understanding of how CN works. I'm just not smart enough to grasp the algorithms and other peer reviewed research paper stuff, so I'll just have to put this out here and see if it's feasible or not.

I propose a dedicated ControlNet preprocessor/model for negative image prompting. In the same way that other ControlNet models instruct the diffusion using images, rather than prompts, this would do so for negative prompting. I envision this working one of the following ways:

• Providing an image full of exactly what you don't want in your generation with this model would cause SD to diffuse images without these things.
• It could work like 1p2p and just be super good at removing what you specify with negative prompts.
• It could work like the Inpainting model, but you would draw over any representative area of what you want gone and it would remove that from the entire image generation. In other words, if you draw a mask on a person's face in an image of a crowd of people, the generated image would remove all of the faces, so everyone would have heads with no eyes, mouth, nose, etc.

Depending on the method, some things, such as color and basic shapes would have to be ignored, or we would get blank images, but if LoRA models can be trained to add details, or fix just hands, I don't see why this wouldn't be possible.

Imagine getting something in all your images, like halation, noise, pseudo-text, or people (I make landscapes and abstract stuff and people pop up in my generations ALL the time). All the negative prompting in the world, even with SDXL, can't seem to solve your issue. This could be an super useful tool in the 'ole toolbox, for folks who get this problem a lot.

I would love to hear someone weight in on the feasibility of this idea. I can train LoRAs, but I've never done a CN model. Any knowledgeable input or advice appreciated.

Apologies if this falls outside the scope of a feature request.

-V

The issue proposes a dedicated ControlNet preprocessor or model for negative image prompting, with possible approaches involving negative images, 1p2p-style removal, or inpainting. Start by assessing the feasibility of these approaches alongside the stated ControlNet, LoRA, and diffusion-model context. No files or tests are identified, and completion criteria are not defined beyond determining whether the idea is feasible.