I Made a Comic with Consistent Characters (No lora training)

Intro

Hello there. I'm still kinda new to this, but want to show people this comic I made of my worst date ever lol.
Hopefully some of you will find this interesting.

I also made a video about making it going over the workflow, and shows the comic, with a voiceover, panel by panel at the end: https://youtu.be/yqSxxORksLE

Overview

So this took me a couple months to do, mainly cause I had to learn ComfyUI. It took me a while to get comfortable in it. I tried many different workflows from others to try this before I made my own that best suited my needs.

(Originally I was using an SDXL workflow with a load of ipadapters and controlNets, but I found a simpler workflow yielded better results.)

All the images were generated with Stable Diffusion, but I composed each panel in photoshop. Each character, background, object (and even the snot lol) were generated separately.

I was using SD1.5 for the speed, (model: Arthemy Comics on civitai), as I had to make a lot of images, and each one took many many iterations of trial and error to get good enough image.

It took 4s for an SD1.5 to generate an image, where as SDXL took 40s, so was too slow to quickly iterate. So then I would only upscale the images I wanted to use in the comic. (Just bypass upscaler section when iterating.)

(Apart from the sneezing images which are SDXL on a different model, because I did these first and couldn't get good enough results in SD1.5. So that's why the girl looks slightly different.)

The comic panels are all 1920x1080 because I made it more for visuals to go along with a video (linked below) to help tell my horrifying story, than for it to be a stand alone comic. I would of tried a more dynamic flow for the layout of the panels if it was to be a standard comic page.

There was loads of different challenges from getting decent poses, getting hands to be somewhat respectable (though still pretty bad haha), getting good expressions on the faces.
But the main one is consistent characters. (And I did it without training my own character loras.)

Consistent characters

So the way I used to get consistent characters was using a mix of:

img2img (denoise: 0.7) with the same model posed into the position desired.
Then using a weighted down character lora and a weighted down celeb name, to help give a consistent base.
And finally a prompt with all the same details each time. (Well apart from changing expressions in the prompt each time.)

Workflow

I'll add a screenshot of the workflow so you can see, and a generated image that you can pull into comfy to get the workflow. I'll also attach a json file as an alternative way to get the workflow.

Workflow screenshot 1