Jibs Qwen workflow uses a custom-trained Wan VAE to remove the hash/grid lines visible in Qwen outputs.
Here is the GitHub for nodes: https://github.com/spacepxl/ComfyUI-VAE-Utils?tab=readme-ov-file
and the Custom Wan/Qwen VAE model: https://huggingface.co/spacepxl/Wan2.1-VAE-upscale2x/blob/main/Wan2.1_VAE_upscale2x_imageonly_real_v1.safetensors
V6 a Qwen 2512 Version of Jib Mix Qwen.
+ Has better fine details than previous versions.
+ Less of a same face issue.
+ Less nosie artifacts.
- Not quite as good at NSFW as V5. (You can use my NSFW lora to bring it back: https://civarchive.com/models/1943554/jibs-qwen-nudity-fixer-lora)
- Not quite as fine good details as Base Qwen 2512
More European faces and fewer Asian by default.
The Pruned Model nf4 (13.91 GB) is actually a Q5_0.GGUF
V5 -More realistic, less plastic skin by default, with more imperfections.
My Skin detail/imperfections lora can be used to add more or even help remove them if used at a negative weight.
The Version 5 Q5. GGUF (marked nf4)
V4 - Much more natural pretty faces (Much better at Asian faces), less noise, cleaner (slightly less photographic look by default but adding LORAs can make that stronger again)
NEW: The fp8_e5m2 model is twice as fast as the Q5.GGUF on my 3090.
There is a Q5 .GGUF (marked nf4)
Tune for Clownshark dpmpp_3s/Bong_tangent sampler this time instead of Euler_ancestral/Linier_ Quadratic.
V3- Important I have uploaded 2 different versions of this model:
High Noise version (marked fp16) that I think is better at lower steps and single stage workflows (but can sometime show grid/scan lines).
Lower noise version (marked fp32) That makes cleaner/less noisey images on the first gen but is better suited to 2 stage Hi-res fix workflow (this is my preferred method)
There is a Q5 .GGUF (13.91 GB) that is the Higher Noise version
The Q6 .GGUF (15.6 GB) is a lower noise realistic skin version.
Small Q5 .GGUF (13.91) is a lower noise realistic skin version.
V2 - I tried to fix the big bobble heads from the previous versions. It is better but they can still be a bit big sometimes.
The V2 fp8 model listed is now actually a Q8 .GGUF (Thanks to export_tank_harmful for the conversion when mine was broken)
V1 - Adds a more photorealistic/amateur look to Qwen.
The fp8 model listed is now actually a Q6_k .GGUF
and has fewer grid lines than the fp8 (But I still recommend the fp16 with ram offloading, for better quality)
Description
Initial Release
FAQ
Comments (45)
Hey bro.
Why is ControlNet not working?
This is one of the best image generation models out there. It's impressive how much of an improvement this is over Qwen Image (and often over Z-Image too). Unfortunately it makes way more errors (usually with things in the background) than Qwen Image and Z-Image, so that can be annoying. And it's clearly mostly meant for amateur photography, which isn't always what I want.
Qwen does seem better at some prompts, especially when they are more complex, it is slower than Z-Image and also has some issues with same face and more plastic skin, I might have to try some 2 stage workflows with Qwen for initial generation and Z-image as a 2nd pass as I think that will look amazing, but be pretty slow.
why i get just black images with the fp16 model?
and with the wan vae i got an error, also with the custom node.
If you have Sage Attention turned on you get black images, it used to be fixable with a KJ Sage Attention patch node but I think a recent ComfyUI update broke that fix so now you have to load ComfyUI without the Sage Attention flag.
@J1B ok i will try later thx. and why the wan vae dont work?
@J1B dosent work, with the fp8 everything is fine, but the fp16 give me black images.
@blackbeard30 strange, not sure what would cause that. what GPU do you have? The fp16 is very big, I have 24GB of vram and 64GB of systems ram and that is on the edge of running out.
@J1B rtx3090 with 32gb ram
@blackbeard30 Hmm if the fp8 works and the fp16 not, then maybe it is running out of memory and it needs more than 32GB to offload to, worst time ever to buy more system RAM! (as prices have shot up).
Maybe just stick to the fp8, the fp16 is a bit more lora compatible but looks pretty similar.
@J1B yes maybe thats why. im happy with the fp8 🙂 thank you for ur help, and ur awesome model.
the Price's are crazy, do u think it would be better some times?
I'm excited for jibmix qwen image 2511. I hope it gets released.
I am working on a Qwen 2512 version now, I am getting OOM issues though, so might have to order an RTX 5090! 😬
@J1B Great news, this model is the best of all for me even after trying the zit version and tons of others !
Can't wait.
@J1B Hopefully we won't have to order an RTX 5090 to use
The provided workflow references this "Jib_Mix_Qwen-Image_V6_A_0001_.safetensors" This does not exist. Can someone clarify where to find this diffusion model?
Sorry, that is just an internal Beta I have only locally, I often start working on the next version as soon as the newest version is uploaded if I think I can make it better.
The upload of V6/Qwen 2512 should be coming within a week or so, it has taken a while as I have been having technical issues making a new version.
is the V5.0 QWEN 2512?
No, not yet, I have tried to make a 2512 model but I am having a technical problem with ComfyUI as the latest version seems to require more memory to merge/save BF16 Qwen models than it did before and it causes an OOM on my system right now, but the Comfyui developers are going to look at it for me I think.
@J1B great, waiting for it!
@J1B Oh man, that would be phenomenal! I would even pay for it! ❤️
Still hoping to see a 2512!
It's been long enough with Klein and ZImage that anyone serious is already starting to see there isn't really a replacement for high parameters once you get past making a profile of two giant tits with a girl attached.
@fohowop124538fdfdfdfds Yeah I still think Hunyuan image 3.0 follows prompts the best out of any open source model, but it recommends 320GB of Vram so quite hard to run locally. I do keep thinking about making a Qwen 2512 finetine but have got distracted playing with Z-Image base for a while.
@J1B 2515 needs you homie
@J1B I am also humbly waiting for a possibl JibMix on 2512. Face down to me Qwen2512 is significantly the most sophsticated open model out there, that can be realistically run locally. I really like Zit and F2 for their speed and accuracy ... but man, tell Qwen 10x to output a "character" ... and you get 10 different ones. The others give you 10 versions of the same... nice and sometimes desireable, but in general way too narrow.
Jib you managed to heave 2509 up to a significantly better level ... I'd love to see you do your best of magic on the top notch out there. :)
@lamohag405 I have been working on a new Jib Mix Qwen 2512 version properly the last few days, and it is looking pretty good. One problem is Qwen 2512 is a much higher bar visually to try and beat than the original Qwen-Image! 2512 does has that weird green filter on a lot of images (probably from training on ChatGPT Images, I do really want to get rid of that as well as increasing NSFW/beauty without it being to AI looking.
If you are looking for a model that does good variety in the mean time Z-Image base (without any speed loras) is very good at variety but needs a good negative Prompt and maybe some loras to look photo realistic.
Man, your model is simply perfect, the sharpness, adherence to the prompt, versatility of body types, heights, among others is incredible.
Excellent work, thank you very much.
Seems to be overtrained with woman nipples and vagina. Basically men look femine or transgender.
If you only generate women, it great.
Yeah, that's probably a fair comment, I do make 99.5% females, I will try and do some more testing on males with my next Qwen 2512 version. If I can get some technical problems sorted.
Just use a custom sampler with res4lyf. res_2m with sgm_uniform, 4steps with the 4 step lora, cfg and denoise at about 1. It shocks me how many people don't know this. Practically all the qwen generated images on this site are jacked except the ones that use this or similar loras, but literally all you have to do is read the documentation. A normal KSampler fucks it up.
>but literally all you have to do is read the documentation
Can you point us to this documentation you're talking about? My search skills aren't so good I guess.
EDIT: You mean the readme on the repo?
https://github.com/ClownsharkBatwing/RES4LYF/blob/main/README.md
How to use this with Snapshot Lora Img2Img? Im noob :((
Have you installed ComfyUI? : https://github.com/Tavris1/ComfyUI-Easy-Install
You can load that lora into the power lora loader node in this workflow: https://civitai.com/models/1936965?modelVersionId=2436685
It is a rather complex workflow, but you can just use the Prompt box.
@J1B when u say use the prompt box, u mean the prompt box for text-image?
I wanna use this model for image 2 image, just like in the comparison here because the results look so real and I need that! But not text to image. image 2 image.
Maybe you'll recognize this comparison.?
comparison: https://civitai.com/articles/21920/qwen-realistic-loras-a-quick-and-dirty-comparison
lora: https://civitai.com/models/2022854/qwen-image-smartphone-snapshot-photo-reality-style?modelVersionId=2289403
Greetings! can you tell me Jib Mix how to remove skin oiliness, shine? I admire your model, as I wrote below in the comments, this is the best model for drawing today, it surpasses both Z-image and FLUX.2-klein. But there are shortcomings, an error, namely that dirt appears on the fabrics during detailing. I use Clownshark - dpmpp_3s + Bong_tangent, and I get good results on res_2s + beta57, but I'm constantly struggling with oily skin and dirt, small dots. Would you like to hear your advice on this? I tried using your LORA Jibs_Qwen_Skin_lora_v5 with a value of -0.20, but this is not the result I need. I also tried to add LORA detailz_qwen_000024000, it adds details, but alas, it spoils the result.
Yes it is an issue with my model, I have had some luck by using this oily skin lora: https://civitai.com/models/2048380/oily-skin-qwen-v2-gmr at a negative value, but it doesn't seem to take it all away. and change sthe image quite a bit as well.
My 2512 Qwen Version should hopefully fix this but it still needs some more work until I am happy with it.
Rumours are that Qwen V2 might be releasing from the Dev soon, so I better hurry up!
@J1B I love the soft moist skin, its a feature dont remove it ! BTW This is still the only model I use and love. it seem to be the perfect evolution and successor to DreamshaperXL Turbo in style and prompt !
can i used this as an edit model?
inclusively, there is some error on the nodes though, especially the one at image saver, may i enquire how do i resolve it?
@leonviz The easiest way is just to replace it with a standard image saver node.
No, it is just a Text to image model. you could use it with inpainting, but it is not quite the same.
@J1B thank you!
@J1B I tried to use this as a image edit model, it seems not to be following the prompt instruction compared to the base qwen model... just curious on the logic behind this , how come ?
@ydo93172 The model V6 is based on Qwen 2512 is not an image edit model, it is a text-to-image foundational model.
You would need to use Qwen-Image-Edit-2509 or a model based from that, I think.