This is an advanced workflow designed for people familar with Comfyui and who want pro grade results. I encourage you to learn it, but if it is overwhelming, the simple version can be found HERE
A complete user and troubleshooting guide inclusing a list of common error messages can be found here https://civarchive.com/articles/24678/how-to-use-my-workflows-a-comprehensive-breakdown
Each Flux model has it's own combination of text encoders. Read the notes carefully. You will get MATT errors of it is incorrect.
也有中文说明
Ver 6.0
Controlnet/ Image reference area, gut & rebuild:
you can now set the image ratio based on your original image. Huge timesaver for Controlnet & I2I
Base image 🖼️ controls everything, so no need to set multiple images
You can even use it as a color reference in post processing
Better Prompt generator (Florence)
Huge update to detailers. I basically yanked them out and replaced them with my V5 units.
You an use Flux or use an SDXL based model to detail. Just switch it.
Updated 🎨 post production suite that includes realism enhancers and an inline crop tool
Updated the main area
Huge update to 🗒️notes.
Ver 5.0
Added controlnet
Added a sentence remover to prevent CivitAi Flags
Ver 4.0
Total overhaul
Optimized for Klein_9b
Options to control the scheduler
Upgraded detailers
Upgraded Post Production Suite
📷 Optical Realism upgrade
Arcane Bloom FX
LensFX
ClarityFX
也有中文说明
Flux 2D I2I, T2I, Upscaler, Detailers, and a hefty Post Production Suite
Optimized for Klein as well as Flux 2D
Include the Klein Enhancer (Similar to a detail daemon) to improve the quality of the Quant images
Prompt Assist
Allows up to 10 images to be loaded for reference
Ultimate Upscaler package
Seed VR2 Upscaler
Face, Hand, Eye, and Expression Detailers
Add lighting, color correction , and specicl effects after generation in the Post Production Suite
Flux 2D Base Model
Flux 2D Model (You have to agree to their terms first):https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
Flux 2D GGUF Model https://huggingface.co/gguf-org/flux2-dev-gguf/tree/main
Flux 2D Mistral text endoders https://huggingface.co/Comfy-Org/flux2-dev/tree/main/split_files/text_encoders
Flux 2Dev Turbo Lora https://civarchive.com/models/2266313/flux2-dev-turbo-lora?modelVersionId=2550958
Flux 2 Klein Model & Text encoder
Klein 4b
Text Encoder qwen_3_4b.safetensors
Flux 2 Klein diffusion_model Flux-2-klein-base-4b-fp8.safetensors
Klein 9b
Text Encoder Qwen_3_8b.safetensors
Flux 2 Klein 9b FLUX.2-klein-9B
VAE (for both)
Instagram: https://www.instagram.com/synth.studio.models/
Buy me a☕ https://ko-fi.com/lonecatone
This represents hundreds of hours of work that I give away for free. If you enjoy it, Please make sure to 👍like and post to my workflows. That is how you support me. Also, feel free to ⚡tip 😉
Description
optimized for Flux 2 Klein
Upgrades
added draft mode
FAQ
Comments (43)
Can't get past the error:
mat1 and mat2 shapes cannot be multiplied (1024x7680 and 12288x4096)
Works with the 4b model, but the 9b model doesn't work.
Same error
I put it in the notes; You have to update ComfyUI
@2thecurve Chedk which version you are using. it should be 3.0.2. I had to modify the clip source for the detailers, but I did that yesrerday
Same error on my site with 9B gguf model (Q4)
@lonecatone23 I use comfyui desktop version, unsure if that matters,
@2thecurve Yeah, I've got bad news for you. ComfyUI broke the desktop links about a month ago. I was on Desktop and switched to Portable>i'm never going back.
Step 1: Update ComfyUI
Step 2: Turn everything off except the "Save Draft" and see where it breaks and I'll create a worksadound. I've spent days working around this It should not be happening
@johanneskrieg254 Check to see that you are using ver 3.0.2, Then switch everything off but "save draft", then turn on and tell me where it crashes. Also, let me know which version of ComfyUi you are using
for me solution is to use the same type of klen as your text encoder. for example both as gguf. that fixed my error of the same typee
@2legsRises357 Yeah, It is very touchy with its quants
you are using wrong VAE . I was there too. name of file was ae.saftners but flux ae safners was diffrent. both are diffrent. i renamed one as flux.ae.saftners
@2thecurve you are using wrong VAE . I was there too. name of file was ae.saftners but flux ae safners was diffrent. both are diffrent. i renamed one as flux.ae.saftners
@amgjustsaiyan595 I have the right VAE listed in my notes. The issue is the text encoder. Klein 9b requires Qwen_3_8b. I didn't realize I forgot that note. I've updated my notes in the workflow accordingly
what about NSFW? does flux 2 (base model) loras work?
Yes and no. They work on the basic flux 2 model. Nothing works on the Klein models right now. It has a different weight structure. However, it does not censor I2I (at least as far as I can tell)
If you are looking to push out fast, use the Lightning Lora with the GGUF model and 8 steps. That still works with Loras.
使用了太多不常用的插件节点
Are you using Desktop? You need to update. Everything works fine.
@lonecatone23 First of all, thank you for your reply. This is my first time using this model, and I've encountered many problems, which I'm currently working on. It might be even better if you could release a basic version with a simplified interface and easier workflow; it would be more suitable for beginners, haha.
@markvin 我这样做是为了让你学习，但我确实考虑过。
@markvin I will work on one soon
Should make it clear it doesn't work for windows
It works fine for windows. I have windows
it werent be useable at all. it just gives me error
SamplerCustomAdvanced
[WinError 2] The system cannot find the file specified
Did you download the models, clips, and VAE? That is a user error
@lonecatone23 it starting with Flux2KleinEnhancer get red with error message
device() received an invalid combination of arguments - got (bool), but expected one of: (torch.device device) didn't match because some of the arguments have invalid types: (!bool!) (str type, int index = -1)
@lonecatone23 disabling klein conditioning other error comes up SamplerCustomAdvanced
[WinError 2] The system cannot find the file specified
@shithook are you using 4b or 9b? also, are you on desktop or portable?
That is typically due to a mismatch of model/clip or GGUF/clip or not using "default" in the loader.
I'm also not a huge fan of the enhancer. I think it works better without it and I personally never turn i on, so you can totally delete it. It won't wind up in a later update.
@shithook did you download the models? Click on the diffusion model dropdown, then the clip, then the VAE (for later) and make sure the path to each is correct (i.e. Flux\Flux2_Klein-4b.safetensors). it can't find either the model or clip
@lonecatone23 no it was shit attention problem...
@shithook yeah, that'd do it
Possibly asking a stupid question here : is not having an input for negative prompts by design ? Been trying to make a negative prompt box and wiring it to the rest of the system, but it is quite difficult with a flow this complex. *steams coming out of my head figuring out what I broke this time* lol
so Flux and ZIT are Classifier Free flow tranformer models (they don't use a CFG so it is set to one.) You have no reason to use them on Flux or Flux 2. Instead, you either use an empty clip text encoder or a conditioning zero out node. However, Klein does, and I probably have it minimized
Every Ksampler and detailer wil have a positive and negative connector. In the GETSET area of the user input, I should still have a hegative text encodr, expand it (it will be in red right below teh green nodes)
FYI, I hate Flux2 and flux2 Klein.
@lonecatone23 I can kinda tell it's annoying from my little experience with it having different structure like a pick-me model x'D
Sadly I'm having the most successful photo-realistic results with Klein9b and the LoRAs I dug out from this website, so I'm still beholden to using it for the time being.
So since I'll be using Klein9b, replacing the ConditioningZeroOut to every negative input with Get_Negative is good, is that okay ?
Really appreciate the answer ! Slowly getting the hang of it now.
@lucilucilu27765 There's a negative Clip text encoder in the gestset node that is global. You can message me on discord if you need further clarification
@lonecatone23 I think I've got it. Thank you very much! Tho if you don't mind me asking, what's the model you like to work with best ? And for what reason ? Ease of use ? Best results ?
@lucilucilu27765 oh lord 🙄
ZIT I hated at first, but as it was insanely popular, I built workflows that worked for each purpose. Now I go to it quite often. People want to say it's best for realism, but an image that looks like it was shot form a 10 year old phone isn't realistic. Every pretty woman has the latest 22mp phone. It's so obvious it's not even funny anymore. I purposely posted real images on reddit and they were all "Plastic, looks fake" and when I used I2I on the same images, they thought they were real.
I like Illustrious. I'm a semi-realism/ Anime/ Art kinda guy. 90% of the time when I'm generating, I use this.
Flux has its purposes. good for artsy realism shots.
I like Anima. It's good for Anime. The prompt adherence is insane.
Klein 4b has its benefits. Its fast and okay, but I have RTX5090 and 256gbRAM. speed and size are not an issue for me.
Flux 2 is garbage. If i need somerthing similar, I'll use Nano Banana Pro. It serves zero purpose for everyday generation.
I want to like Qwen, but I need to mess with it more. I started when I had my old machine, so it was sloooooooow.
Flux Krea pumped out hyper realistic pictures that put everything else to shame with zero LoRas. I was shocked as I hated it at first until I made a workflow that wteaked the settings. It's stupid good.
I'm about the workflows. I'm not trying to achieve anything but to produce high quality professional highly adaptable workflows.
@lucilucilu27765 I just trained my "Better passion" on Klien_4b. I'm probably going to build an essentials workflow tis afternoon after I push V8.0 for Zit out today
@lonecatone23 People like the low megapixel photos cause it can make convincing old pics for nostalgia, I guess ! Must be the older crowd. I've tried Anima, and the only downside is the barebones LoRA choices for now, I guess ? But it IS very easy to get most characters and tags without even needing LoRA so true I agree ! I tried ZIT but I can't see much difference from Klein, personally. Maybe I'm just not refined enough to notice. Flux Krea ? Never tried it before. I'm working with a hand-me-down 4070 Super, so my options are more limited :')
I'm still working to learn the nodes (hugely supported by your workflow uploads!) and try personally generate more details with realism first, but definitely will mess a lot with Anime style after !
@lonecatone23 A Checkpoint ? Is it for lighter use purposes ? My Drive Disk is slowly getting filled up with more and more LoRA and Checkpoints from window shopping these websites ToT
@lucilucilu27765 Delete what you do not use. A checkpoint is basically all in one (model, cl;ip, vae) and takes up space accordingly
@lucilucilu27765 I did a thing https://civitai.com/models/2414019/klein4b-essentials
@lonecatone23 Oh a Workflow on Klein4b ! Are there a lot of differences between the Klein9b and 4b ?