V2.0 Notes

This builds upon the previous version, I've added so much to it, there's now an extra 2 prompt groups, 4 refiner groups, an extra 2 Upscalers and a SEGS Detailer.

This is designed to be fully modular and you can mix and match. Instruction nodes are on the workflow.

Overview -

This is in group blocks which are colour coded. By each block is an input switcher and a bypass toggle control to quickly turn groups on or off and switch between them. Some have extras such as Seed Control etc. These are bright orange so they stand out and are easy to spot, same for the Notes boxes which are bright pink.

First main block is green and these are the Positive Prompt groups. There's a standard manual prompt box, a combination of BilboXPromptGeek and Portrait Master nodes, an Ollama Generate group and a Florence2 group with an image loader and this one will use the desrption of the image to make the prompt.

There is then 3 model loaders - Flux, Kwai-Kolors and Stable Diffusion, they're used by more than one group so and by keeping these separate you can turn groups off without them breaking other groups.

The next block is purple and here are 2 choices, Flux Txt2Img or Im2Img.

After this we have the blue group which has a minor upscaler and a choice of 4 refiners, Flux, Kolors, Stable Diffusion and Stable Cascade.

The final block is black and these are the 4 upscalers you can choose from, AuraSR v2, UlitmateSD (connected with Flux so it does take longer than what you may be use to), SUPIR and a Model Uspcaler and Kolors refiner.

Final group is a SEGS detailer. I normally have this switched off during generation and there is a Load image node connected to it and you can choose to either switch it to run with the generation or load an image. So I'll turn all the other groups off bar the Stable Diffusion model loader, drop in an image I like that's completed in and use a detailer. You can repeat if you want to detail different areas of the same image such as face, eyes, hands etc.

My personal preference atm, is Flux for the generation and 1st refiner and then the model uploader with a second refine using Kolors.  

Just seen about ControlNet, so if that's working I'll probably be putting in an updated version soon.

V1.0 Notes

This is designed for dev (which I much prefer over Schnell), if you are using Schnell just drop the steps to 4.

If you haven't set Flux up yet follow this and make sure your ComfyUI install is up to date.  https://comfyanonymous.github.io/ComfyUI_examples/flux/

It's colour coded for blocks. The first block is green and is for prompting. There are 2 options here: Manual Positive Prompt as per standard or BilboX PromptGeek and Portrait Master. There's a switch at the right also in green, set it to 1 for manual prompting and 2 for BilboX/Portrait Master.

The next block is in purple and is for the Flux generation. I've set the weight type in the Load Diffusion Model node to fp8_e4m3fn and the clip_name1 in the DualClipLoader node to t5xxl_fp8_e4m3fn so it should run on most systems, I have a 3090 and 64GB of system RAM so I've set clip_name1 to txxl_fp16 on mine. I recommend leaving the weight type as it is as my generation times take much longer when that is set to default.

There's then a couple of bright orange nodes (so they're easy to spot), an input switch and a fast group bypasser. These are for choosing the Upscale method, AuraSR (I'm using v2) or UltimateSD. Set the switch to 1 and turn off the UltimateSD group to use AuraSR and vice versa to use UltimateSD. I'm using AuraSD as it takes around 30 seconds and UltimateSD takes around 30 minutes with no real difference in quality. I left UltimateSD in for anyone who likes to tinker but it's switched off by default.

For AuraSD you will need to create a folder in the models folder in your ComfyUI directory called Aura-SR and download models.safetensors and config.json from here  https://huggingface.co/fal/AuraSR-v2/tree/main  and put them both in that folder. As I found out when I first tried it, it will not work without the config file in there.