Koda KOLORS Pack - Controlnets,12K Upscale, IPAdapter in ComfyUI版本v2 (ID: 660052)

v5 pack -- (install info, show more, scroll down)

v5 pack --

v4 pack --

v3 pack --

v2 pack

v1 pack

KODA Collection

New Kolors Base Model

with IP adapter plus in ComfyUI

v1 pack

- base txt2img and img2img workflow
- base Kolors IP adapter-plus

v2 pack

- Advanced IP Adapter workflow with SUPIR Upscaler

- Base workflows for running Hyper Kolors /LCM Kolors

v3 pack

- ip adapter embeds

- all in one workflow

- SUPIR upscaling

v4 pack

12K, 3 stage upscaling
SDXL checkpoint merge, UNET Block Weighted
Sampler Garden (Tuned Grid)

v5 pack

all-in-one type workflow
controlnet canny + depth (toggle)
IP Adapter advanced (toggle)
Sampler Garden (toggle)
12K upscaling (2 stage toggle)

v6 pack

all-in-one type workflow
controlnet canny + depth (toggle)
IP Adapter 2 stage (toggle)
Sampler Garden (toggle)
Face model creator & swapping toggle
12K upscaling with color match/remix (2 stage toggle)
auto prompt enhancer variant workflow

How to use the V6 Workflow?

The main control panel has ben refactored for easier operation.
Here we can use the txt2img basic Kolors and get simple output images from prompts
When we have placed images into the IP Adapter, we can enable that like so:
As shown above, model switch (2) is used to activate the IPAdapter. You can use the controlnets with or without the IPAdapter, by using that switch.
When you have finished making some good images, you can activate the Upscaling section:

all dimensions are being driven by AspectSizeV2, fully explained in the video.

With the new Upscaler, you have two options, Color match and Color Remix:

You can choose to use the same image for the color match or experiment with other images to alter the color range used.

The second new workflow used "Auto-Prompt" Enhancers for both positive and negative. Enter your simple prompt into the left most box, and use the options to enhance your prompt with magic prompt, one button prompt and auto negative.

it is identical to the new all-in-one workflow, just with this added as an example of what you can do quickly with this. It will introduce many obscure tokens and has both style drop down list features and insanity levels for positive and negative. Full previews are included so you can see what was added.

How to use the V5 workflow?

txt2img with Cnet+Ipa
(Koda-Kolors-Zero-Legacy)

Base Kolors mode

model switch (1)

Kolors Basic enabled only

Functions as only txt2img
IPadapter mode

model switch (2)

Kolors basic & (2) IPadapter enabled
Controlnet Mode

can be used with both model postions, as you choose.

enable the Depth & Canny groups together, for default behaviour
All perturbed modes (sampler garden)

these can be use with or without controlnet or IPA modes

you can copy these and customise them to build a dynamic Grid
12K Upscaling

stage 1, 2, 3 each is performing 2X upscale, you can use your best upscaling model too

comparer located at the bottom for easy access, but beware 91mb images can cause UI lag ! recommend to minimise/collapse/remove all previews in stage 2 and 3 ;)

How to use the V4 workflow?

below pictured: Stage 4 active, switch position 4, [4] combine Embeds.

the new all-in-one workflow is multifunctional, you use it in stages 1,2,3,4 however it's much easier to use. Simply match the model switch to the active group. All is explained in the v3 video, and in notes in the workflow.

Step 1. txt2img only

model switch 1

used when all groups are bypassed

This is Kolors txt2img

Generate images and place your best ones into the IP Adapter image load nodes

The rightmost image loader will also be used as the negative embedding

Step 2. a) IP Adapter

model switch 2

requires group [2] IP Adapter, to be active

Test the output of the chosen IPadapter image batch, remember this is separate from the prompt entirely. If you are happy with the results;

Step 2. b) Save Embeds

model switch 2

requires group [2] IP Adapter to be active

saves the positive and negative as IPA embed files

these are saved to \models\output\

run once and then copy the files to \models\input\

then click refresh in comfyUI

Step 3. Load Embeds

model switch 3

requires group [3] Load Embeds, to be active

once you have created your embed, we should test it.

load the positive and negative, then test the generations.

repeat the process with characters, styles and scenes so you have some elements to be mixed, then we can head to the final stage.

Step 4. Combine Embeds

model switch 4

requires group [4] Combine Embeds, to be active

Select the embeds and add or remove nodes as needed

interpolation and operators are available, with all the usual IP Adapter nodes.

This can work like Aesthetic Embeddings worked with SD1.5

For most purposes, we want to anchor in a style or concept and then use text prompts to generate, LORA is the best for this, however this can get very close within the constraints of the model memory.

the SUPIR group can be toggled with any stage for easy upscaling

You will need:

install any and all missing custom nodes & use the latest version of ComfyUI
watch the v1 video for more setup info

https://huggingface.co/Kijai/ChatGLM3-safetensors/tree/main

download one of the models, i will use the fp8 version, you might use another.

place inside ComfyUI\models\LLM\checkpoints\

https://huggingface.co/Kwai-Kolors/Kolors/resolve/main/unet/diffusion_pytorch_model.fp16.safetensors

place inside ComfyUI/models/unet/diffusers/kolors

https://huggingface.co/Kwai-Kolors/Kolors

I also added all of these files from diffusers here.
we already downloaded the "diffusion_pytorch_model.fp16.safetensors" and put it into the /models/unet/kolors/some of the nodes look in a different location, it's up to you which method you want to use as both will work, this is discussed in the video.

We do not need any .bin of the contents of text_encoder because we took the ChatGLM3.fp8.safetensors instead.

structure of ComfyUI\models\diffusers\Kolors\

ComfyUI\models\diffusers\Kolors> tree /F

│   model_index.json

│

├───scheduler

│       scheduler_config.json

│

├───text_encoder

│       config.json

│       tokenizer.model

│       tokenizer_config.json

│       vocab.txt

│

└───unet

        config.json

        diffusion_pytorch_model.fp16.safetensors