Image Edit

Inpainting, image-to-image, and regional editing setups built on Ideogram 4 and Krea 2 GGUF quants.

Inpainting

Encode the source image to latent space with the VAE. Feed a mask into the sampler so only the masked region regenerates. Decode the result back to pixel space.

Image-to-image

Encode the source image to latent space. Run the sampler with a controlled strength so the output stays close to the input. Decode the result.

Regional editing

Chain a mask into the conditioning stage. The GGUF diffusion model regenerates only the selected region and leaves the rest of the latent unchanged.

Models and formats

Model	Repository	Recommended formats
Ideogram 4	molbal/ideogram-4-gguf	Q8_CR, Q8_0
Krea 2	molbal/krea2-gguf	Q8_0, Q4_K_M

Loading instructions

1. Download the GGUF files from Hugging Face.
2. Load the diffusion model and text encoder with the GGUF loader nodes from comfyui-gguf-reboot.
3. Encode the source image with the VAE before sampling.
4. For inpainting and regional edits, feed the mask into the sampler.
5. Decode the result.

Showcase

Image examples are on molbal's Civitai profile.