Hello there and thanks for checking out this workflow!
What's new in v27? : Major rework with new SubGraphs to streamline and compact the workflow even more, removal of outdated/broken nodes + fixes and optimizations all over!
—Purpose—
Built to provide an advanced, versatile and modular workflow for Flux with focus on efficiency, structure and information.
It comes with many notes explaining node settings and recommendations, as well as general guides from instructions to troubleshooting.
—Features—
Convenient loaders for all common versions of Flux and Clip models
Full metadata; recognized by CivitAI
LoRA support
SageAttention, EasyCache + Model Compile acceleration
ControlNet with Union support
Flux Tools LoRAs Canny + Depth (alt. to CNet based on LoRA)
Flux Redux (similar to IPAdapter)
PuLID (SVDQuant version only)
Wildcard prompting
Installation and download guide for models and nodes
multiple passes with optional upscales
— 1st : Detail Daemon + Variation Seed
— 2nd : DD. + Tiled Diffusion / UltimateSDUpscale
— ADetailer with dedicated LoRA Loader
— Inpainting
—Custom Nodes—
ComfyUI-nunchaku — SVDQuant version only
All of which can be installed through the ComfyUI-Manager
—Troubleshooting—
If nodes show up red (failing to load), check the 'Install Missing Custom Nodes' tab of the ComfyUI Manager for the missing node packs and install them.
Please check if all custom node packs load properly after installing, i.e. no
(IMPORT FAILED)messages next to any of them in the console upon ComfyUI startup.Always reload/drag'n'drop the original, downloaded workflow file into ComfyUI to reload an intact version of the workflow.
→ The last opened workflow that appears on startup shows a cached version of the workflow, "remembering" group nodes that failed due to missing nodes as failed, keeping them broken even after having everything installed correctly.
—Thanks—
The workflow would not be possible as is without these custom node packs. If you want to support the custom node creators, give them a ⭐ on their github repos! Thank you!
Feel free to ask questions, share improvements and suggestions in the comment section!
Let me know if you encounter confusing points I can elaborate on in the next update!
Description
v24
- complete rework without group nodes
- custom node pack changes :
- : ComfyUI-Allor
- : ComfyUI-List-Utils
- : ComfyUI_LayerStyle
+ : comfyui-mxtoolkit
— v24.1
- fix for dedistilled-model check causing needless rerun of passes when no parameters have changed
FAQ
Comments (28)
getting multiple import failed's with some dependences i can trace back and some that i cant . I've come to accept that troubleshooting these things is going to take longer than the benefits of most workflows. And my new way of doing things is simply to keep trying different workflows until i get ones with either one error to solve, or no errors to solve. I can save myself thousands of hours this way all totalled up. Fixing workflows is not really the creator of the workflows fault, its just all very temperamental --- and the time it takes to fix something is often not worth it.
Hey there, ComfyUI can indeed be very stubborn with all its troubleshooting. Compared to a year ago though, all the python dependency hell has gotten much better. But I know the frustration all too well, if by "temperamental" you mean the urge to punch a hole in the screen 😄
In general a lot of issues stem from:
- Changes in ComfyUI's core causing conflicts especially on less maintained node packs
- Less experienced node authors setting required dependencies in stone on a specific version, rather than keeping the versions as flexible as possible (bad practice issue)
- Incorrectly installing node pack requirements manually by using the system's python instead of the standalone one of the Comfy location via "pip install..." (a mistake I made in my first months of Comfy 😅)
Also important from a system setup perspective:
- Having GPU drivers somewhat up to date
- Compatible CUDA version for PyTorch
- Visual Studio Tools
- Environment variables
- Ffmpeg (niche)
In your case, checking up on the respective node packs' github pages might solve your import failures and oftentimes only requires fixing once.
When I decide on which node packs to include I try to use the most known/popular/maintained ones, which most users have installed already. So when there's an issue with one of them, there will also likely be a lot of reporting done with a quick fix following.
I also try to keep my workflows as similar in node pack choices as possible, so that getting one of the larger ones set up means many others are mostly good to go too, so there shouldn't be too many headaches at that point. The work I put into my workflows likely does amount to thousands of hours of experience building them, so I might be a little biased, but I think they're worth checking out.
Troubleshooting is a tedious aspect for sure, and I try to provide as much insight as I can to make things work smoothly and help out, but at the end of the day, each user has a unique setup and on top of that ComfyUI is in constant devolopment, adjusting to daily changes in the world of AI.
We miss the old workflow, the old node layout will return ?
So do I. Did you read the "General Info" inside the workflow?
As long as group nodes are not replaced by the SubGraph feature, there is nothing I can do to further reduce visual clutter.
Great workflow! the vllm node would not import because there is no windows wheel for that version had to do a manual install someone posted on the issues page on github. i think same is happening with pulid wont import it's again a wheel error i guess any chance you could help with that? because i could not find a manual install/fix for that.
Hey there, thank you very much! 🙌
I am not sure why the ComfyUI_VLM_nodes pack would even need a wheel to install though. I don't remember having to do that, same with PuLID... are you running a somewhat "exotic" setup? I'm just a casual windows 11 using bozo with regular gaming hardware I suppose 😅
So I'm not very familiar with intricacies different systems might have.
In your case though a regular install through Comfy Manager should have worked just fine.
Does comfy show any other issues, errors, (IMPORT FAILED) messages upon startup in the console log?
Hi! I was looking through your workflow and noticed that it uses a node called SimpleMathDual+.
However, I couldn’t find this node in ComfyUI_Essentials or any other common custom node packs.
Could you please let me know where I can get this node?
Hey there!
That is a weird bug other users ran into as well, but the fix is simple. For some reason specifically node packs made by Cubiq/Matteo fail to install completely/correctly through the Comfy manager. But uninstalling the pack, deleting the eventually remaining folder from ComfyUI/custom_nodes and then manually installing it via git clone should solve this.
1. Delete ComfyUI_essentials and close ComfyUI
2. Delete potential remnant folder from ComfyUI/custom_nodes
3. Open a command console/terminal window within ComfyUI/custom_nodes
→ (right-click empty space within folder and select "Open in Terminal")
4. Execute the following command to download the ComfyUI_essentials node pack repository from github:
```
git clone https://github.com/cubiq/ComfyUI_essentials.git
```
5. Enter the following to move the terminal to the newly downloaded ComfyUI_essentials folder
```
cd ComfyUI_essentials
```
6. And now install its dependencies from the requirements.txt file within that folder by executing
```
..\..\..\python_embeded\python.exe -m pip install -r requirements.txt
```
7. Done! 🤞
Very nice workflow, thank you.
A small comment: maybe I missed something, but I had to remove the logic related to dedistilled models because it forced all the stages to restart despite of the same seed and input used. It was very annoying when I changed something in upscaling params, but the workflow kept running 1st stage over and over.
Hey there, thank you very much!
That is of course not supposed to happen. The only cases where it can't be avoided is when using PuLID or the random image loaders. Those unfortunately always rerun. But the dedistilled check has no excuse to do that.
Looking into that right away. 👌
Are planning to add ComfyUI-TeaCache node? it's gives significant generation speed boost
Hey there, thank you for the suggestion.
Personally I found TeaCache to hurt the quality too much and remember that I ran into some issue as well, but don't remember what exactly.
But I'll look into adding it as an option in the next version 👌
@Ultra119 Added in v25 and it works much better and smoother than when I checked it out around the time it came up. They made a lot of progress for sure. Thanks again for the suggestion! 🙌
ComfyUI node graph theory in practice
Does it support seed variation with strength 0-100?
Hey there!
What do you mean exactly? If it's an interesting feature, I might be able to build it in.
The seed handling in this workflow is using a central seed value to supply all seed-using nodes. However I built in an offset calculation into all those nodes to have unique seeds and prevent running into fried generations caused by sampling on the same seed twice.
Additionally by locking in the general seed to "fixed", all other seeds are automatically locked in as well, when their respective offsets are kept at "fixed"
@RedPinkRetro I learned this concept with Automatic1111 before but it doesn't support Flux so that's how I got to know your workflow. It's described here: https://stable-diffusion-art.com/automatic1111/ (scroll to Variation seed and Variation strength parts). I believe standard KSampler node accepts variation seed parameter, it is rather a matter of providing an input field for it plus the wiring.
Basically, why I need it is if I want to create a new AI character, I would set up my prompt and iterate through random seeds. But when I find the golden shot, I want to fix the seed and then slightly modify the output and take another 20 shots so I can try to create a LoRA of this character.
@elejama I see, it does have a similar effect to splitting the sampling process with a noise injection, like I had in some early versions of this.
Noise injection would basically allow keeping an unfinished first half of the generation constant, while having a second sampler going from there with any seed you want, being forced to stay closer to the first half depending on the step you split the sampling and place the injection.
That would effectively give you a very similar control, but without being a mere interpolation between two seeds, but with more variability and freedom to adjust.
I'll look into it and try adding the feature next update 👌
@RedPinkRetro Cool, thanks! Hopefully others will love such feature as well.
@RedPinkRetro I tested the new variation seed option, it works as expected. Thanks!
In v24 I can't seem to use Inpaint alone. It requires at least 1st pass to be enabled. Otherwise it will print:
Failed to validate prompt for output 21022:
* (prompt):
- Required input is missing: dictionary
Also, after I have inpainting done, cannot upscale the result in 2nd pass, as it only accepts 1st pass as input
@elejama Yes, this workflow really is intended to be used left to right in one go, rather than inpainting mid way.
I can safeguard the error you got there to no longer occur if you turn off the 1st Pass, but it would imprint default values as metadata when saving. Because of that mostly the 1st Pass is supposed to always remain active, as the paramters of it are the initial generation parameters.
Can you describe the typical procedure you would expect for your use case?
If it's something I can make possible within this workflow without negatively impacting it, I can do so. If not I could design a dedicated workflow specfically built for it.
@RedPinkRetro The most flexible way for me would be a routing box where I could link outputs from each pass to others. But for now it would be enough if I could upscale inpainted image, either as an output from inpaint or from an input file.
@elejama That will be part of v25, as I add SVDQuant support, which is not generating 100% deterministically, so even fixed settings and same parameters always have slight differences.
Because of that it makes sense to have that feature to avoid regenerating and worse outcomes on repeated tries.
@elejama All added in v25 alongside the variation seed feature. I found a node that does exactly that and it's great to work with. Thank you again for the suggestion! 🙌
My inpaint prompt is not saved in output file metadata :( Instead, pass 1's prompt is being saved. I have to use save_workflow_as_json which is cumbersome. Hopefully you could fix this?
Hey there, this is intended and necessary as only the 1st Pass prompt is part of what would generate the image in the first place.
Inpainting is usually an iterative process where different parts are marked and replaced with focused prompts on specific areas, so the prompts would not represent the overall image at all.