inpaint_global_harmonious preprocessor

[axelblaze88]

Do these nodes include any preprocessor like inpaint_global_harmonious from automatic1111?

That feature works wonders for image restoration and I need to switch to ComfyUI for more flexibility

[valery016]

It is indeed a very special preprocessor that can be used for a number of purposes beyond image restoration (where SILVI2 is the most brilliant example). I use it for my current artistic project as the only suitable method of the kind, will try to demonstrate that with examples on the main discussion page. According to my brief research, there is no node for, or any other code implementation of inpaint_global_harmonious preprocessor in Comfy, and SwarmUI doesn't support it either (the developer doesn't even know what it is). Still, I would love to have support for inpaint_global_harmonious in Krita Ai plugin too, if possible.

[axelblaze88]

Yes, that's exactly what I was trying to do, implement SILVI in Comfy. Wanna add some intermediate processes to the workflow that'd be impossible with AUTO1111's rigidity. A real shame they don’t implement it in Comfy, honestly :(

[valery016]

I am getting very good results with Hyper-SDXL-8steps-CFG-lora in Forge / SDXL, equivalent or actually better than those with SILIVI. So hang on, I intend to publish results here and on reddit SD sub, also on CivitAI (obviously).

In the meantime, here is what I designed for Krita in this regard. Feel free to play with it, see if you can reproduce it and/or perhaps improve it even further.

{
"name": "HyperScale (XL)",
"version": 1,
"sd_version": "auto",
"sd_checkpoint": "RealVisXL_V4.0.safetensors",
"loras": [
{
"name": "Hyper-SDXL-8steps-CFG-lora.safetensors",
"strength": 0.85
}
],
"style_prompt": "best quality, highres",
"negative_prompt": "bad quality, low resolution, blurry",
"vae": "Checkpoint Default",
"clip_skip": 1,
"v_prediction_zsnr": false,
"self_attention_guidance": false,
"preferred_resolution": 0,
"sampler": "Alternative - Euler A",
"sampler_steps": 8,
"cfg_scale": 2.0,
"live_sampler": "Realtime - Hyper",
"live_sampler_steps": 8,
"live_cfg_scale": 3.5
}

Before generating (Strength 100% (!)), create Unblur control layer with the settings as in the screenshot below.
Also, since the method adds significant red color cast to the output, use the Luminosity blending mode for the generated layer to get rid of it.

Here's my best result with Krita for the SILVI test image: