This is the workflow that was made with my Lustify models in mind, although other models would obviously work with it too. Please, read all the notes I've left in the workflow!
The goal here was to create a setup that utilizes custom nodes only when absolutely necessary. Let's be honest - you don't need 100+ custom nodes to generate great images, and I know exactly what it's like to have obscure nodepacks installed just to use them for a single workflow.
It also utilizes subgraphs, keeping the UI very easy on the eyes and hiding all the noodle spaghetti inside them.
Krea 2 edit related:
Get the node from there (I think it's not available via ComfyUI manager right now).
Get the identity edit LoRA from there (any of the newer ones).
Otherwise the edit functionality wouldn't work.
SDXL related:
WARNING! If you use DMD2 workflow, you'll need to do a little bit of manual tweaking, as it utilizes NAG (Normalized Attention Guidance) node for negative prompts to work. And the node hasn't been maintained by it's author for some time already. It's not hard to fix it:
1) If you don't have the node installed yet, open your ComfyUI\custom_nodes. Open terminal there and do "git clone https://github.com/ChenDarYen/ComfyUI-NAG.git"
2) In the file "custom_nodes\ComfyUI-NAG\chroma\layers.py", change line 5 to: from comfy.ldm.flux.layers import DoubleStreamBlock, SingleStreamBlock
3) That's it, the node is now working.
NON-DMD2 Workflow: My personal favorite. It’s slower, and you really need to use most of the custom tools I added to the workflow to get it right. But honestly? It gives the absolute highest quality and most creative results.
DMD2 Workflow: Super fast (6-8 steps) and actually supports negative prompts thanks to NAG. The downside? DMD2 has its own specific visual style, and you do lose a bit of diversity in faces, backgrounds, and textures. Some people prefer this purely for the speed and the vibe.
Description
FAQ
Comments (60)
Hey, I just bought your model and wanted to test it using the recommended workflow. But even though I’m following all the instructions, I’m getting an error and the Groups Bypass node isn’t being recognized. Where does that come from?
Does the template also work with a basic workflow using just KSampler and VAEDecode?
Hey. Which of the workflows are you using - DMD2 or NON-DMD? Groups Bypass node comes from rgthree nodepack, do you have it installed? Is your comfy updated to the latest version?
Not sure what you mean by "the template"? If you mean the model - yes, absolutely, it's just this workflow was made to compliment it, but any other WF will work as well
@coyotte
Yes model. Everything has been updated, including rgthree. I'm using NON-DMD. That's weird, but oh well, I'll just stick with my basic workflows. Thanks for your reply.
While I'm at it, I love your work, all your models are amazing.
@Asticotboy well, in that case you can just add the "fast groups bypasser" node by yourself. It will use the one you have (Comfy does have that problem sometimes) and you'll be able to enjoy the model with this workflow.
thank you for the kind words!
Such an awesome workflow, it works really well with your Lustify models and others too. thanks for your hard work 🫶
Thank you!
Hello, what about A1111? thank you !
A1111/Forge don't have much customizations. The only advice I can give is to use highres.fix and adetailer, and possibly tiled upscale (I believe SeedVR is not supported in these UIs)
what about it? it's deprecated
@billybob420 I also still use it cuz I despise ComfyUI's spaghetti mess.
@KaiserWilhelm it doesn't have to be a spaghetti mess. It's just many creators don't bother making their workflows look easy on eyes. There are tons of tools in Comfy for that (e.g. subgraphs, which my workflow utilizes heavily)
@coyotte I think that new Forge Neo is also pretty helpful? Since it runs ComfyUI in the back but has a A1111 layout? Sadly my pc is kaka and I have to rent a GPU off Runpod. And they don't have Neo. And their ComfyUI is severly outdated.
On a sidenote. I tried out Lustify V8 in A1111 but the eyes are all wonky. Im assuming this is due to using A1111 and not your model. Any advice on how to fix that?
Great work as always on the model btw.
@KaiserWilhelm for the best possible quality, you need to use highres.fix -> upscaling -> adetailer. Not just with Lustify, but with any SDXL-based model, the 4-channel VAE makes it impossible to have good facial details on medium-to-distant shots.
Also, if you aren't using DMD2 lora, the amount of steps should be around 30-35.
Alternatively you could try using DMD2 lora, it somewhat helps with that issue you've mentioned, but not completely
@KaiserWilhelm forge doesn't have comfyui in the background. you're thinking of swarm
v8 apex works great. Love your workflows. Very easy to understand. However, this time, there's no bypass for main gen. So im unable to use inpaint?
Hello, thank you for this great contribution. I have loved and exclusively used your previous "Bare Minimum" workflow. It looks like "Apex", albeit taking longer, is more potent and much better organized. It may be user error, but it turns out the results are a little inferior (and slower) combined with LoRAs. I would really appreciate if you could write an overview of the improvements and added functionality. Such "propaganda" would convince me that it is indeed a superior workflow—which I suppose, since you managed to revolutionize T2I multiple times in the past. Maybe this can be appended to the description. And thank you for the buzz tip yesterday!
Not sure why would the newer workflow output worse results than the previous one did, because it's basically the same thing, but compacted into a more organized version, and now incorporates SOTA tech like SeedVR2.
It is slower because it is set to make use of V8's 1536 resolution training (the optional FaceDetailer and HandDetailer are working at that exact resolution, but you could lower it to 1024 in subgraph settings, and maybe the "upscale_by" value in the highres pass, depends on what you have it set to).
But if the older workflow works best for you - just go for it, it's not a hard requirement to use that one :)
@coyotte Thank you for your reply. Apex actually is better (+ super well organized), just with LoRAs the results seem less creative/flexible. Using dmd2, but I have previously also used dmd2 due to hw. I will probably stick to 1024, thanks for the info. I don't want to stay stuck in the past, but will go with your developments! You are the goat here. Огромное спасибо ))
@Solovyova Взаимно!
[17:03:37.507] 🚀 Materializing DiT weights to CPU (offload device): C:\Users\micro\Downloads\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\models\SEEDVR2\seedvr2_ema_7b_fp16.safetensors
C:\Users\micro\Downloads\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable>echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe
If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe
C:\Users\micro\Downloads\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable>pause
Press any key to continue . . .
Well, have you tried doing what the log prompts you to?
ive got the latest nvidia drivers, the workflow gets stuck on the seedvr2 step. Other than perhaps too small a paging file, im at a loss.
@majorboner what gpu do you have? or VRAM amount?
ok thanks @coyotte ,update, I'm now using the '7b Q4 ~4.5 Gb VRAM' and that's working now. I was using the '7b fp16' before which didn't work. I'm quite new to all this. I have a 5080 16gb VRAM and 32gb RAM.
Hi, xinsir-controlnet-union is marked red in first pass. Also 4x_NMKD is marked red in second pass. Should I download them manually?
hey! The "4x_NMKD" can be changed to any other upscaler you have/prefer, like Remacri, or Superscale, or Ultrasharp, etc. The "xinsir-controlnet-union" needs to be downloaded manually, that's correct https://huggingface.co/xinsir/controlnet-union-sdxl-1.0
Alternatively, if you're not going to be using controlent at all, you could open the first subgraph and remove the controlnet nodes completely.
@coyotte Thnak you very much, is there a tutorial for impaint part? I assume there is a way for a faceswap in this workflow?
@marek69857 actually no, faceid for faceswapping isn't present there. It might be tricky to install because of insightface dependencies, so I chose to not to add it.
As for the inpainting - the note near inpainting graph describes what you need to do
@coyotte Thank you, so is it possible to generate images with input face?
I'm asking because i want to generate images with consistent face.
@marek69857 if you add the FaceID nodes manually - yes, it should be!
Alternatively you could add some faceswaping nodes like Reactor.
I'm also finding there is a noticeable difference between the 'First Pass' & 'Highres Pass'. First image looks good, then the high res pass will change things, sometimes considerably, like swap a woman for a man or change breast size. How can I prevent this?
Lower the denoise value on the highres pass. Sounds like it's too high for the image you're working on
@coyotte thanks, ive not changed any values from the workflows you provided by the way, I just used them as is and noticed this when generating images. i want the highres pass to essentially take the exact image from the first pass as is, but its not doing, its changing things considerably as I say
@majorboner I understand. Some images require adjustments of the values. Try setting denoise to 0.4
Fantastic workflow! I'm new-ish to image generation, and this workflow is not only simple to use when you just want to generate content, it's also very useful to pick apart if you want to learn how everything works and should be hooked up.
Thanks so much for posting!
thank you!
I don't know why, but my lora doesn't work. btw my checkpoint and my lora works in others workflow
What checkpoint are you using? First guess is that it's not SDXL-based
The high pass mess up the body and the whole picture, it adds breasts, new legs, and weird rhings.. at first it was working good, i dont know why it went like that... Even with a 0.4 denoise, it still messing up everything, is it a seed issue ? i just disabled the ksampler for the moment
EDIT : Solved, this was because the seed by default on high pass was fixed, put it to 0
Can this workflow be used with a character lora? Should we train the lora on Lustify Apex or base SDXL for best results?
I'm getting very blurry images using the DMD2 workflow as is. Do I need to wait for the DMD2 version of Apex V8 first?
you need to use the dmd2 lora.
Works great but the inpainting requires manually disabling nodes and it doesn't seem to work for me. It's processing but the output is the same image.
open the inpainting subgraph and see if the nodes there are bypassed or not
@coyotte Thanks for replying, I already checked the subgraph and nothing is purple in there. I still can't figure out how to make it work it's always outputing a very similar image, the only way I found to finally get changes it to have 1.00 denoise but I'm getting poor results by doing this. I think I just don't understand too much how it's supposed to work.
@Notahumanagain It wouldn't be a drastic replacement if you're using denoising values like 0.3-0.4. Something around 0.45-0.55 is as high as it should go if you're using a regular version of SDXL checkpoint (in some occasions 0.6 max). For value of 1.0 you'll be better off using a inpainting version (Lustify has multiple)
@coyotte I did try the inpaint versions it works better but it's still hard to get decent result, I believe I just don't understand enough how AI works
@coyotte but thanks for help
Where do I put the prompt please?
i have error: Cannot execute because a node is missing the class_type property.: Node ID '#11'
Great Workflow. I just wish it would support wildcards to create certain variants in a batch!
works great but face detailer is not following my character lora and changing to another random face
You might need to connect the lora model output to facedetailer model input, if it's not connected that way
@coyotte yeah thats the first thing i've checked and its already connected, idk whats hapenning, it ignores completely the lora, can pretty much change straight from asian face to caucasian :shrug:
@AIVORY3D does the lora work for the first pass tho?
@coyotte y it does, I have a friend on discord facing the same issue with face detailer. so idk i will investigate furthermore
you should lower denoise. 0.4 is quite hefty. 0.20 - 0.25 should do the trick