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
02/04/2026 - Seems the Wan 2x VAE node is broken after a ComfyUI update and makes dark images, use the updated workflow without it: https://civitai.com/models/1936965?modelVersionId=2824176
Workflow Used to make showcase images
FAQ
Comments (34)
This is beautiful. I cannot wait to try it. ^_^
so is the v5 model based on from qwen image 2512? also where is the fp8 version of the model?
V6 is biased on Qwen 2512 (V5 is the old original Qwen Image tuned for realism)
The FP8 file is uploading now, should be there in about 20 mins.
fp8 model is there now: https://civitai.com/api/download/models/2753478?type=Model&format=SafeTensor&size=pruned&fp=fp8
It can have some slight noise artificats compared top the fp16, but is faster than the Q5.GGUF
I think the model is more artistic than the v5 version.
Does original Qwen LORAs work with Qwen 2512? Great work as always. Have you seen anima that looks like a great model too.
From my limited experience (before I quickly gave up). No. Not a single one worked for me personally.
Some apparently will work to a limited degree (I'm told), but it's so rare that you probably shouldn't bother.
@ShadowCell I have found style loras to be very compatible between the 2 models, but I haven't tested character loras, as I don't really use them.
@J1B I think I tried some various ones on here; 2509, 2511, etc.......it might have been the 4 step LoRA causing issues, however. In the few times I tried, it was often just generating completely broken images though.
Why isn't Qwen way more popular btw? Is it expensive/hardware restrictive to train with it or what?
Also, what's all the different styles you've found can be generated, & via what prompting? I'm trying "strong penciled lines" to get a comic style, but given it's more literal (less pretensious prompting style), I'm having a harder time generating a massively diverse range of styles with just the base model. Particularly, how do I get more abstract, dream like, lucid images? Also, all comic styles? An overt Western anime type style would be nice.....or is it just not in the base model to pull out?
@ShadowCell I think the reason it is not more popular is that it is heavy on system recourse and slow to use, the smaller quants (and turbo loras) are not the same quality as the full bf16 model and even on my 3090 it takes around 3 mins per image (even more if I do a final upscale) when ZIT can make a similar or better 1girl image in less than 1/2 the time, I think Qwen image could have a good future (especially if they ever release Qwen Image v2 opensource) it probably just needs more loras or finetuning to push art styles more, I might give making more loras for it a go as I am liking what I am getting out of it now (even if it is really quite slow).
@J1B I see what you mean about popularity, & system resources. I saw someone else's image uploads, & corresponding generation times; & immediately started to wonder if they set it up wrong....... (even when they used 2 or 4 steps LoRAs the times were insane by comparison to what I was getting).
https://huggingface.co/Wuli-art/Qwen-Image-2512-Turbo-LoRA/tree/main Give V3 a shot if you haven't already. I'm getting amazing quality outputs/speeds from it. It'll do 1500x1500 4 step images in less than 20 seconds for me (though I realise this varies massively on different hardware). I keep meaning to upload images to demonstrate (but just haven't quite got around to it).
Seems Qwen 2512 is so dangerously close to generating extreme NSFW material from the base model alone that it's actually pretty hilarious (surprisingly so). Some selective/creative prompt wording offers surprising outputs to say the least......lol.
@ShadowCell Yeah that V3 Turbo lora does seem good, takes my 3 stage workflow down from 320 seconds to 90 seconds, with very little quality loss. I just need to do some more testing on pose variation and prompt adherence, but it is looking good. I might try and merge it in as that can be a little faster than using it as a lora sometimes.
Can't wait to try out V6 (Qwen 2512). It seems Qwen doesn't get anywhere near the amount of love & attention it deserves - 2512 base model seems massively superior to FLUX; amazing natural language understanding/prompt adherence (& it's way less censored, too).
Edit - it was less of a change from standard Qwen 2512 than I was hoping for tbh ;(
I'll say this, I didn't like the v6, it's like a basic 2512, one-on-one, what's the point of this model at all? Jib Mix Qwen all previous versions were different from the basic ones and I liked the realism, even with a bit of FLUX skin, but v6 is not what I expected, so I stayed on v5
I don't think it is one-to-one the same (I would show some side by side testing but Civitai image galleries are very broken right now).
But yes it is quite closed to Qwen 2512, as that is a very good and photo realistic model.
I did make some early versions of V6 that were much closers to Jib Mix V5, I could upload those as an alternative model if you would like?
@J1B Yes, with pleasure, please download
Here is an example of the difference in the look with my model vs base Qwen 2512: https://civitai.com/images/124355223
Mine is more sexy/Jib Mix looking, but it removed a little of the realism, which is why I have not got too far away from Qwen 2512 as I like the realism.
and better NSFW: https://civitai.com/posts/27284078 as always.
@J1B I would like, v5 is the holy grail for me. I would love to get it!
what lora(s) to do porn shots?
Is there any GGUF of V6 version?
Yes, it is this one: https://civitai.com/api/download/models/2753478?type=Model&format=GGUF&size=pruned&fp=nf4
Q_5_K.GGUF
I still cannot believe Civitai have not got proper .GGUF support in there upload labels
@J1B Thanks - i see that's the Q5_k version. The Q6 link in your model page points to a C:\ drive path - do you have the link to the Q6 version?
nevermind i see that was for V3
@EchoHeadache Yeah, I do not have a Q6.GGUF for V6, I suggest you use the 19GB fp8 with 16GB of vram and system offload as it is 2-3x faster than using a .GGUF format in my testing.
Do you plan on doing a qwen2?
It doesn't look like the weights for Qwen 2 are going to be released online. I have done some testing on their API and it doesn't seem much better than using a good workflow and Qwen 2512.
Really wanted to love it but it seems to produce five faces for men and two for women.
Yeah, I do know what you mean, it annoys me too sometimes. Hopefully Qwen 2 will be better and will be here soon...
Really impressive work, my go-to checkpoint for Qwen today. Would you be open to disclose how you did the fine-tune? what tools and setups you used and what Loras you merged? I want to try my hand on this myself and its always good to start learning from one of the best ;)
V6.0 is absolutely brilliant!👍