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
Merged Qwen models with my favourites loras at precise weighing to produce the usual Jib Mix quality.
FAQ
Comments (53)
at 38gb, I doubt my 5090 could run it, which is a shame T.T
fp8 (19 gb) has been added
I am running it on a 3090 (24GB of Vram) in Comfyui, it spills into system RAM and makes it a bit slower and also a 19GB fp8 version is now available , I have not done much testing with that yet, so not sure how much worse it is.
@2P2 the 19GB fp8 does have a lot of banding artifacts unfortunately. Maybe, I could try and make a Nunchaku Quant with deepcompressor, that may solve the lines.
@J1B When you say "a bit slower" what does this mean? what frame times are you getting?
the base with lightning lora I run at under 10s for most gens.
I'm thinking of testing this on runpod A40.. will report back
@QualityControl I am doing really big images now 2160x1536 @ 20 steps with Eular/Linear_quadratic
That is taking around 200 seconds on my RTX 3090, but no need to upscale afterwards :)
planing to convert this into nunchaku just like u did with flux? btw i am using fp8 for now of this qwen of urs
I am going to try, but the official nunchaku Qwen model is very bad at skin , so I am not sure if I will have much luck. The fp8 version of my model has very bad artiface on most images I am afraid that the bp16 version does not.
@J1B yes that model seems vary bad at quality also
,but do u convert new jibmix flux model to nunchaku coz we have old version now
@markhassain3712 I just loaded up my Runpod account with $50. I am going to give it a go for Qwen and Flux.
@J1B thnx thnx ,i love u
something about this merge keeps making bobbleheads. most of the gens have a weird pony-esque fakeness and overly large heads. It's weird. Any idea where this might be coming from?
you can see it in most of the gallery examples here too. most of the people have heads that are too big
hmm, yeah now you mention it , I see it, I have been having issues with the faces and lack of making blondes (which are my favourite!) so something is a bit weird.
@J1B yeah. it gets even worse when I stack a character lora that was working on base qwen on it. super enlarged head
@QualityControl which lora? If you put it in as a negative weight does it shrink the heads?
@J1B just one of my own custom trained ones. simple bare basic character lora that used to work on base but inflates heads on this. I'll try negative but I've since deleted the 30gb checkpoint so I'll dl it again later to test
I have found if you put "big head" and "large face" in the negative it helps a lot, but you have to have your cfg above 1 for that and it doubles the generation time :(
@J1B I see v2 - what'd you do to try fix the bobbleheads?
Was this model trained or is it a merge?
Merge with 8-10 different loras, I am working on a V2 that is finetuned in Runpod on my aesthetic dataset right now.
@J1B I asked just because it's marked as checkpoint and not checkpoint merged
@NRDX Oh I didn't tick either, it must be a bug in their UI, now corrected.
@J1B What do you use to merge LoRAs into a model? Every time I tried with Flux, the LoRAs would be too weak to even notice.
@DaifaGarda Just ComfyUI, load a model load the lora and selected the weight, save the model.
It is the hours of comparing different blends with a metaphorical magnifying glass and checking for artifacts/overtraining before and after merging that is the time consuming bit.
I recognize these faces, I saw them in another model, funny
Yes, the quality of the fp8 is very poor... It's a pity, because the model turned out to be really photorealistic, but neither my GPU, nor RAM, nor SSD capacity can pull a full 38Gb.
Qwen draws people well, but it's a very, very heavy model.
Yes, maybe someone can fix the way it quantizes to fp8 for Qwen in Comfyui? There is still hope of a Nunchaku model from me you might be able to run, I am just testing out training on Runpod instead of locally as I need the extra Vram.
@J1B Yes, it might work... Or someone will fix the quantization. GGUF quantization of Qwen models works fine in principle.
Have you tried making a scaled FP8 version instead of a standard one? They seem to have almost Q8 like properties. I believe KJ nodes has one to use as a base, but it's e2m5 rather than e3m4.
@dysphorai I will check that out, thanks as Qx .GGUF's are a lot slower than fp8.
Successfully converted to GGUF formats - Q8 and Q6 using City96's tools. I can confirm that there are no artifacts like in F8 version. That solved the problem of the big size. If anyone wants, I can post the steps (Flux is straightforward but QWEN is a bit complicated). Thanks for the model. Images look a bit more natural.
Yes, I would be interested in the Qwen GGUF conversion process you used, if you don't mind sending the details.
I have city96's nodes updated but when I try to run the converstion I get this error:
File "C:\ComfyUI_For_Wan\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\custom_nodes\ComfyUI-GGUF\tools\convert.py", line 154, in is_model_arch
assert not invalid, "Model architecture not allowed for conversion! (i.e. reference VS diffusers format)"
^^^^^^^^^^^
AssertionError: Model architecture not allowed for conversion! (i.e. reference VS diffusers format)
@J1B if you are at this stage, it's easily solved.
The project has 3 branches. The tools to read QWEN is only in the auto_convert branch (not the main branch). Use this command to clone that particular branch only.
git clone -b auto_convert --single-branch https://github.com/city96/ComfyUI-GGUF.git
Then extract the tools folder and follow City96 instructions.
(You need to install everything in requirements.txt plus 'safetensors' and 'packaging' in your virtual environment)
This text file contains all the instructions needed for a beginner for the conversion of QWEN image to gguf format.
https://drive.google.com/file/d/1QdmtXijMROZC8nTpKPJGQ_Xlt8OIJK6U/view?usp=sharing
@puretna5320 Can you share gguf Q6 ?
@puretna5320 Could you share gguf Q8?
@soyv4 the author says he'll be uploading.
@puretna5320 Great instructions, I managed to create the .GGUF a Q4_K_S , but the quailty when I generate images with it is even worse than the fp8 I created before! :(
@puretna5320 This worked, thanks!
@J1B I only did q6 and q8. They were great.
But I only had the fp8 to compare. The original 38gb model wouldn't even load.
@puretna5320 I made a Q_6_K now and it looks great, just uploading now. Also I am trying a Q4_K and will test.
Q6 .GGUF is there now:
https://civitai.com/api/download/models/2192261?type=Model&format=GGUF&size=pruned&fp=fp8
I will test a Q4 later.
@antey3074 I think he forgot. That was nearly a week ago.
@BRai4now there's the v2 instead now to work with as base anyways
@BRai4now what were you looking for? There is a Q6 and a Q4 for V1, I will try and make .GGUFs for V2 in the next few days.
@J1B Oh no worries, I was waiting for a Q8_0 and saw @puretna5320 had made one. I was kinda expecting y'all to hook up so you didn't have to make it and you could just post his.
I should pro'lly just follow his instructions and City96 to be able to make my own.
@BRai4now I can make a Q8 as well, it just takes about 30 mins to generate each Qx and takes up a lot of disk space and I have been too busy generating images :)
Greetings! will there be a GGUF?
Yes, the fp8 was a failure, but I plan to make Q8 and Q4 .GGUF and upload them tomorrow.
@J1B can't wait!