tanmaypatil123/krea-2-controlnet

Krea-2 Depth ControlNet-LoRA is a tool that adds depth-guided control to an AI image generation model called Krea-2. Give it a photo and a text description, and it will create a brand new image that keeps the same three-dimensional structure and composition as the original photo, while changing the content and style to match whatever you described. For example, feeding in a photo of a robot and asking for a tiger produces a new image shaped like the original but rendered as a tiger. Under the hood, the tool first extracts a depth map from your input image using a separate model called Depth-Anything-V2, which figures out how near or far each part of the image is from the camera. That depth information then guides the image generator so the new picture follows the same layout and perspective as the original. This is done with a small add-on file called a LoRA, a lightweight adjustment layered onto the existing Krea-2 model rather than a full retrained model, which keeps the download size to under a gigabyte. The project works with both the full Krea-2-Raw model and the faster eight-step Krea-2-Turbo version. It reports strong consistency between the depth of the input and the depth of the generated output, meaning the 3D shape is preserved closely. Users can also skip the text prompt entirely and let the depth map alone drive the output, which is useful for testing how much structure the depth control provides on its own. To use it, you install the code and download a shared checkpoint file from Hugging Face, then run a command line script pointing at your photo and prompt, with options to adjust step count, guidance strength, and how tightly the output follows the original structure. The project also includes code for training your own version of this control system on other kinds of visual guides beyond depth, such as edge outlines. It works best with photos or 3D renders that have real depth variation, since flat illustrations do not give it much to work with. The model weights are released under a separate community license from Krea, not a standard open-source license.