Others

GitHub Repo: https://github.com/TinyTerra/ComfyUI_tinyterraNodes

v1.4:

Added Nodes:

• textDebug - displays text that is input into it, with optional print to console.

• INT > TEXT - converts int to string

Bug fixes:

• fixed text concat outputs when boxes change to inputs had no connection

• fixed models from pipeLoader being reloaded every generation

• fixed rembg not being installed properly

• updated config logic

V1.3:

• Added imageRemBG node - Background Removal node with optional image preview & save.

• Added imageOutput node - Preview or Save an image with one node, with image throughput.

Image shows an example workflow for photobashing directly from generations. JSON can be found in \workflows

v1.2:

• fixed pipeLoader having image output instead of clip

• Added clip output to pipeKSampler

• Added original pipe throughput to pipe>basic_pipe and pipe>detailer_pipe nodes

  ________________________________________________________________________________

• added config.json for auto update (set to false by default) -

  • you will need to install the nodepack via git clone https://github.com/TinyTerra/ComfyUI_tinyterraNodes.git from inside the ComfyUI\custom_nodes folder, and set autoUpdate to true in the config file for auto update to work properly.

  ________________________________________________________________________________

v1.1:

Added 'pipe > basic_pipe' and 'pipe to detailer_pipe' for better compatibility with ltdrdata's ImpactPack

v1:
Adds pipeLoader and pipeKSampler (Modified with merge of Efficiency nodes and Advanced Clip Text Encode) for even more efficient Loading and Sampling.

Along with pipe Utils (IN, OUT and MERGE)

The guide is short as you will see: The number of people who do not know that safestensors can be opened as txt to extract the metadata is impressive. They can steal everything, even the number of images used and repeats, the only thing that is inaccessible is the database

I trained the singing voice clone AI on voice clips of the Sentry Bot from Fallout 4. I used the default training settings meaning 10000 epochs- though with the simplicity of the Sentry Bot's voice that was probably overkill...

Anyway, It works pretty well and the AI holds on to the "audio detail" of the Sentry Bot's voice, and where it messes up it still sounds believable (since the Sentry Bot's voice is already "noisy" and imprecise in a way). However, if you want those pitch changes found in the Sentry Bot's voice it needs to be included in the input audio. More on output quality, slow clear speaking is recommended for the input audio, since that's how the Sentry Bot speaks as its voice is hard to understand otherwise.

I'm not sure whether or not CivitAI allows uploading models of other open source AI tools, but I mean if they don't its their loss I guess, if they do then thats a whole other dimension the service can take advantage of.

Here is the source:

Image source: https://www.nexusmods.com/fallout4/mods/56150

You will need ControlNet (I was using version 1.1.166) with a segmentation model. Optionally you will need some editing software. I will be using Photoshop but you can use whatever program you want.

Model: Counterfeit-V3.0, but it should work with most models.

Links

Counterfeit-V3.0 model: Counterfeit-V3.0 - v3.0 | Stable Diffusion Checkpoint | Civitai

VAE I used: civitai.com/models/23906/kl-f8-anime2-vae

EasyNegative embedding: civitai.com/models/7808/easynegative

ControlNet: github.com/lllyasviel/ControlNet

ControlNet models: huggingface.co/lllyasviel/ControlNet-v1-1/tree/main

How to install: ControlNET 1.1 - What you NEED to know!

For CLIP and VEA Settings add this to Settings -> User Interface -> Quicksettings List:

CLIP_stop_at_last_layers, sd_vae

1. Set up ControlNet and txt2img

First we need to make segmentation image for ControlNet. To do this you can use an existing image from txt2img/Internet or draw it by yourself. Depending what angles or how many views you want to have it will look differently (other examples I made and tested are on the end of this post)

Drop your image and set Preprocessor to ‘seg_ufade20k’. Click Allow Preview, then click red explosion button. After it ends processing save this image with download button in the top right corner.

For some images you will need to clean it. After that, drop your clean image to ControlNet and set Preprocessor to ‘none’. Select ‘control_v11p_sd15_seg’ model and click ‘Enable’.

Colors: Red: #96053D, Grey: #787878.

I suggest you set Control Mode to ‘My prompt is more important’ and Ending Control Steps to 0.6-0.9.

After that it’s time to set up txt2img. I suggest you use 3:2 ratio.

Resolution: 768x512 or higher if you have a good graphic card. I’m using 1152x768.

Higher resolution should give you better quality.

(These settings are just an example, feel free to play with it but I suggest you leave white background, simple background, reference sheet, from below, from above, multicolored background, multiple views, variations, mismatched clothes)

Prompt example:

Positive prompt: masterpiece, best quality, ultra-detailed, absurdres, 1girl, blue eyes, black hair, long hair, blunt bangs, detailed eyes, wide-eyed, eyelashes, looking at viewer, upper body, white background, simple background, reference sheet, standing, white t-shirt

Negative prompt: EasyNegative, monochrome, mismatched pupils, symbol-shaped pupils, heterochromia, multicolored eyes, no pupils, slit pupils, asymmetrical pupils, empty eyes, asymmetrical eyes, mismatched eyelashes, asymmetrical eyebrows, asymmetrical eye shading, two-tone hair, streaked hair, colored inner hair, multicolored hair, gradient hair, earrings, hair ornaments, boy, asymmetrical breasts, grabbing, text, from below, from above, multicolored background, multiple views, variations, mismatched clothes

Other settings are not that important - you can use whatever you want.

I used these settings:

2. Click Generate and take image you like most.

This is the image I’ve generated (seed: 3355877752):

Now it’s time to edit this image. (if you like your result you can skip this step)

Image after editing:

3. Set up img2img

Upload your image to img2img.

Copy settings from txt2img, but this time don’t use ControlNet.

I suggest you use higher resolution than txt2img but like before - if you don’t have good gpu just use 768x512. I’m using 1536x1024 here.

Change prompt.

Positive prompt: masterpiece, best quality, ultra-detailed, absurdres, 1girl, blue eyes, black hair, long hair, blunt bangs, detailed eyes, wide-eyed, eyelashes, looking at viewer, upper body, reference sheet, standing, white t-shirt, outdoors, blue sky, clouds, trees

Negative prompt: EasyNegative, monochrome, mismatched pupils, symbol-shaped pupils, heterochromia, multicolored eyes, no pupils, slit pupils, asymmetrical pupils, empty eyes, asymmetrical eyes, mismatched eyelashes, asymmetrical eyebrows, asymmetrical eye shading, two-tone hair, streaked hair, colored inner hair, multicolored hair, gradient hair, earrings, hair ornaments, boy, asymmetrical breasts, grabbing, text, from below, from above, multiple views, variations, mismatched clothes

Resize mode: ‘Resize and fill’, denoising strength: 0.4-0.65.

My settings:

4. Generate until you get a good result

For more details you can take your output image, drop it to img2img and use it again with less denoising strength.

My output (seed: 2666038464):

___________________________________

Examples of different segmentation images: