ComfyUI nodes for outpainting images with diffusers, based on diffusers-image-outpaint by fffiloni.

Updates:

• 20/10/2024: No more need to download tokenizers nor text encoders! Now comfyui clip loader works, and you can use your clip models.
• 10/2024: You don't need any more the diffusers vae, and can use the extension in low vram mode using sequential_cpu_offload (also thanks to zmwv823) that pushes the vram usage from 8,3 gb down to 6 gb.
• 10/2024: If your text_encoder and text_encoder_2 names contain .fp16. or other things before safetensors, you need to remove it (see the table below).

Installation

• Download this extension or git clone it in comfyui/custom_nodes, then (if comfyui-manager didn't already install the requirements or you have missing modules), from comfyui virtual env write cd your/path/to/this/extension and pip install -r requirements.txt.
• Download models in the comfyui/models/diffusion_models folder, following the grid below (you can use the links to download the suggested models; you can also change the main model, but you need the specified controlnet since the extension is hardcoded to use it, for now):

  main model	controlnet model
  Diffuser Model folder (you can change this model)	Diffuser Controlnet folder (you need this model)
  model_index-json	config_promax.json, diffusion_pytorch_model_promax.safetensors
  Unet folder
  config.json, diffusion_pytorch_model.fp16.safetensors
  Scheduler folder
  scheduler_config.json

Overview

• Minimum VRAM: 6 gb with 1280x720 image, rtx 3060, RealVisXL_V5.0_Lightning, sdxl-vae-fp16-fix, controlnet-union-sdxl-promax using sequential_cpu_offload, otherwise 8,3 gb;
• As seen in this issue, images with square corners are required.

The extension gives 4 nodes:

• Load Diffusion Outpaint Models: a simple node to load diffusion models. You can download them from Huggingface (the extension doesn't download them automatically);

• Paid Image for Diffusers Outpaint: this node creates an empty image of the desired size, fits the original image in the new one based on the chosen alignment, then mask the rest;

• Encode Diffusers Outpaint Prompt: self explanatory. Works as clip text encode (prompt), and specifies what to add to the image;

• Diffusers Image Outpaint: This is the main node, that outpaints the image. Currently the generation process is based on fffiloni's one, so you can't reproduce a specific a specific outpaint, and the seed option you see is only used to change the UI and generate a new image. You can specify the amount of steps to generate the image.

• You can also pass image and mask to vae encode (for inpainting) node, then pass the latent to a sampler, but controlnets and ip-adapters are harder to use compared to diffusers outpaint.

Credits

diffusers-image-outpaint by fffiloni