This is an advanced workflow designed for those familiar with ComfyUI who want pro grade results. It is custom node heavy and has numerous options designed for those who want granular control over their end result. I encourage you to take the time to learn it. If you feel overwhelmed, you can start with my simple workflow HERE
也有中文说明
Instagram: https://www.instagram.com/synth.studio.models/
Buy me a☕ https://ko-fi.com/lonecatone
This represents hundreds of hours of work. If you enjoy it, please 👍like, 💬 comment , and feel free to ⚡tip 😉
Note: If you have ANY issues with nodes not downloading, read the notes or reach out. There's nothing that special about any of them that aren't core modules. Most issues can be resolved by updating everything
⛔⚠️🛑✋ Read the notes completly before using. Most common install and node problems are listed in the directions
A user and troubleshooting guide including a list of common error messages can be found here https://civarchive.com/articles/24678/how-to-use-my-workflows-a-comprehensive-breakdown
也有中文说明
V4.2
Man, I hate detailers. I really do. They either make an image or tear it up. I spend more time trying to find the right combination of these that actually do something, only to find each model is different.
I've upgraded these to Sam3 detailers. They are much cleaner, especially on the face and eyes.
You can also use them for inpainting.
added a spare detailer
added blend mode sliders
I improved the LLM instructions.
Added word ✂️ nodes in teh Florence area to get rid of watermarks, logos, etc.
Improved CivitAI 🚩 filters
Made a few minor tweaks.
V3.1
Not enough to change the version number, but important enough to re-release:
Repair hi res fix slider get/set break
Redid sigma math on Karras (try it....)
Added a link for the 🚫abliterated Text encoder
Adjusted some settings for out of the box realism
V3.0
Fixed an error in the Image Edit area 🙄
Added LLM instruct prompt selector 📝 (modified from @FaboroHacks )
Added a really cool style selector node
setting tweaks
updated notes.
V2.0
made some tweaks
added Image edit
Note: This is not as good as I thought it woud be. I need to find a model or LoRA that helps it. If someone finds one, let me know.
Note: I've used several .GGUF versions of this, especially from this site. They are not good. I left the GGUF option there for it, but it is a very VRAM friendly model
Updated notes
V1.0:
Pompt Assist
Includes Florence Prompt to extract a prompt from an image
Includes a GGUF abliterated version of Qwen Promper Enhancer.
Image Options:
Select the aspect ratio or set it to match the original
Includes I2I 🖼️
Advanced Sampler section
Custom schedulers (unique setup) that allow you to choose between preset sigmas
Detailer Daemon
Denoise slider
simple one location setup for your base settings
Ultimate Upscaler with VRAM assistance
Full array of advanced SFW & NSFW detailers
Full post production Suite
multiple photo effects
Color adjust to a loaded image or to the original image
Clarity FX
Lens FX
Arcane Bloom
📷 Optical Realism Suite
Inline crop & edit suite
Description
Not enough to change the version number, but important enough to re-release:
Repair hi res fix slider getset break
Redid sigma math on Karras (try it....)
Added a link for the abliterated Text encoder
Adjusted some settings for out of the box realism
FAQ
Comments (11)
And of course ComfyUI-Conditioning-Rebalance releases a full-on overhaul that breaks the Conditioning Krea2 Rebalance node. The error is: TypeError: scale_conditioning() got an unexpected keyword argument 'per_layer_weights'
Specifically the "per_layer_weights=plw" call. Comfy apparently doesn't know what that means.
Stay on version 1.1.0 for the flow to work; 1.3.0 or higher breaks things. I think this version may break the NSFW portion? I'll poke about and find out why NSFW prompts are not working.
Sage Attention.... looks like that was causing my NSFW problem. Even when changing the prompt, the output image seem to still be pulling information from the loaded image... And it was even in your notes. :P
And dude, the input/output tags on the ( . Y . ) Detailer? Oh man, I must ask if I can steal that! :P
Sage attention does not work if you are running image edit.
Other than that, the workflow runs just fine on a full update just 5 minutes ago. Yo need to run on nightly, not the official version
Yank the conditioning node out and put a fresh one in its place, bypass it, or empty it out so it skips it.
lol, and sure
@lonecatone23 It's working on the 1.1.0 version so I will leave it for now, and isn't this node the workaround for making NSFW in Krea? (New to this model.)
Anyhoo, as I said, it's working all working now and VERY impressive. Thanks, as always, for the work you put into these!
@maxoverride871 Not so much the workaround for nsfw per se, as that'd be an abliterated version. The model taps the text encoder 12 times, it helps balance that out. It also helps with prompt adherence as well as Ip dadaptor like image references that you could only get from teh closed source version. As a bonus it does NSFW better.
V3.1 is soooo much better. I hope you update soon
ngl bro this might be the best all in one workflow for krea2 if u bypass things u dont need then customize and fine tune things with good loras, you will get awesome results.
That's why I have so many switches. Thank you 🙏
can you also make the workflow use [ComfyUI-Krea2T-Enhancer](https://github.com/capitan01R/ComfyUI-Krea2T-Enhancer) node cause i think, not only it has conditioning rebalance but it also makes the model understand the prompts better.
Yeah, I was looking at that last night. I like the author's work a lot. However, I noticed very little difference between the two. I may mess with it more this week. We'll see.
It's a new model. Everyone throws their work in the ring.
@xvcy3w I added it.
I finished understanding your workflow. Yes really you made something AMAZING !
The only thing i don't like is your I2I flow. You could make a way better one using controlnet union for krea2.
The main issue for I2I is the lack of preserving original subject and with/without background to edit via prompting the result.
It's the same for image editing.. for example, if i use Flux2Klein 9B with for example a comix image, it can make a real life photography out of it without too much changing the subject and pose (around 95~98% same as original) without using controlnet at 1.00 denoising.
In your workflow the image edit transform too much the original picture. like it's a completely different one expect if i reduce a little the denoising like 0.80~0.90.
But expect for that, the real power of your workflow is for detailing and upscaling. Very good node formations. I took some of your workflow parts to mine because you had very good ideas.
Big thanks to you !
That is a limitation of the Krea 2 model. not my workflow. You will not get that out of the Krea 2 model. This is closer to Qwen Image