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
Better realism and clarity
FAQ
Comments (33)
the example images looks very good, nice
Looks amazing. Really need Nunchaku int4 version for 4steps.
The Nunchaku team hasn't updated their Deep Compressor Repo to support Qwen yet:
https://github.com/nunchaku-tech/deepcompressor
not sure if they plan to Open-source that or not.
For me v3 doesn't seem to be playing nicely - it doesn't like Ultimate SD upscaler, it adds grain and jpeg compression like artifact - while with the same setup default Qwen doesn't do this.
Also, this feels slightly overburned, as it barely can create Asian faces, even when I prompted 'Japanese woman'.
I have found in my testing that SD Ultimate Upscaler does work ok if you use Euler_Ancestral/Liner_Quadratic.
Yes, I know Asian and other ethnic faces are hard to get right now I will look at fixing improving that in a future version.
I used jibMixQwen_v30.safetensors version; if I try the model in default weight dtype mode, it already creates minor faulty looking patterns with UltimateSD Upscale, I use "Custom Sample" version of this node, and euler_ancestral + linear_quadratic, cfg 1.0, tried denoise 0.0.5-0.2, steps 4-10. In this case quality looks more like there was lying sigmas node in use with too much strength (but I don't have it here), but when I change the model dtype mode to any of the 8-bit versions, upscaled image gets really nasty grainy pattern on each area where there normally would be some details.
If I only replace the model with qwen_image_fp8_e4m3fn (don't change anything else), image is like one would expect, but of course people will look like plastic dolls, like Qwen output usually does, but there are no visual artifacts.
Looks great! Any chance to get the Qwen Nudity Fixer Lora update?
Respect to you! thanks for Q5 .GGUF, it fits exactly 12 gig of vram memory, we will test it.
Dear J1BC, tell you which model you like the most for generating, Flux or WAN or Qwen, which do you use more often?
do you use 16gb of ram?
Hi mind to sharr the workflow for the GGUF model?
Can you do Q5 Lower noise?
I have uploaded a Q6 Lower Noise version now. (I think this is the most realistic)
I can do a Q5 as well if it is useful
@J1B Q6 is good! The main thing is that it's not 38 GB.
Q5 Higher Noise is too noisy.
Is there a way to make her chest bigger? Will I need a separate lora for that. I'm pretty new to Qwen image I've mainly been using Illustrious.
Prompting should be able to get you results but you might find this easier https://civitai.com/models/1994714/breast-slider-qwen
Finally managed to get not over exposed images and asian girls. Using your model as high noise one for first 15 steps, then original qwen for 5 steps. Results are promising.
I don't have the original qwen, but I've got nunchaku qwen image edit 2509 instead. Thanks for the tip bruh! @novarchibald174 @J1B
Mind share the eorkflow?
@claudex810953 sure, let me get to pc
@TonyRex mind to share the workflow?
@claudex810953 I am away from the pc, but I used the Qwen template wf except I replaced the ksampler with an advanced ksampler and generate 3/6 steps, then pass the latent to another ksampler (qwen or qwen image edit 2508) ... but then again they have their own prob with repeating face syndrome
https://civitai.com/models/2066148?modelVersionId=2337981 my workflow for 2 step gen
So far, excellent model. Skin doesn't look plastic and cartoonish like on the base model. Would love to see a SVDQuant, FP8 quality at around 12gb. Qwen image flies on a SVDQ + Lighting LoRA. Maybe I'll figure it out. I don't mind loading some money up in Runpod.
Same. They are relatively less lossy than going lower GGUF Quants in my experience.
Great model! Thanks. I’ve been using your v3.0 for my generations. However, I noticed that the high-noise and low-noise files are actually the same. I realized this because I was getting identical outputs from both. To confirm, I checked the SHA256 hash, and indeed, they match: 2432d89bc694efeee0feabed18f2d7adff422accaff13ce5bf9b984867c5cf7c.
That said, it’s still a big improvement over the original.
I'm downloading using 32 at the end of the link
......&format=SafeTensor&size=full&fp=fp32
cause I downloaded the fp16 (I think)
does this work with the 4 steps lighting lora?
Yes, 4 steps is a little low for my taste but if you do 4 steps and then a 4 step Hires Fix /Latent upscale, you can get good images without using any lightning lora or use it at a low weight around 0.35-0.5 for the best quality images.
Really appreciate the detail and realism in Jib Mix v3! I am using the Q5 13GB. One thing I’ve consistently run into: it’s still very difficult to generate accurate East Asian faces—even with exp licit prompts like 'Japanese woman', 'Korean', or 'Chinese female,' the results often default to non-Asian or mixed features. On top of that, the model tends to produce the same facial structure repeatedly limiting diversity. Excited to hear you’re working on this. Good on ya m8
My next version Jib Mix v4 Qwen is so much better at Asian faces, it should be out today or tomorrow.
V4 is out now, it does have much better Asian faces: https://civitai.com/images/106210176
@J1B I'm seeing an interesting behavior with v4 where the KSampler preview shows facial diversity in early steps (I've observed distinct asian faces in early steps), but by the final output, the model consistently converges to the same phenotype regardless of seed, prompt, or scheduler settings. This happens even at 4-12 steps. The face is the same woman in the samples. It always collapses in the later denoising stages into the same face. I tried injecting the latent with asian faces (as in Qwen Edit wf) it worked for a bit then collapsed, leaving the same face only with a different clothing. This is really interesting because the original Qwen leans towards asian faces than latinx/mediteranian/western. Ima conclude this is a qwen image limitation than jib mix'