ver.0.4 Anime (Qwen2512)
This is an anime model based on Qwen-Image 2512 （FP8/GGUF Q4KM）.
When using the GGUF Q4 version, please increase the Shift amount in ModelSamplingAuraFlow. 8.0 or higher is recommended. Increasing the Shift amount will improve the unevenness of the paint, but it will also reduce the amount of information in the image, so be sure to strike a balance.
ver.0.3 Anime（FP8/Q4KM/BF16）
This is based on ver.0.3 Real (a more advanced version of Real) with animation elements added. I was planning to make it available for shared use, but that failed. It is almost an anime model. There may be some photographic elements left, so adding anime to the prompt will stabilize the drawing.
ver.0.3 Real(FP8/Q4KM/BF16)
This model is based on ver.0.2 (which has undergone further training) and adds Asian elements.
The previously quite harsh Caucasian facial features have been toned down a bit.
The GGUF has been changed to Q4KM, which prioritizes accuracy. If you prefer Q4KS, which prioritizes speed, I can add it. There is not much difference in accuracy or speed, though.
ver.0.2 Real (FP8/Q4KS/BF16)
This model uses similar materials to 0.1, but it has more data and less training. Compared to 0.1, I think 0.1 produces cuter girls. 0.2 has darker colors and improved accuracy in NSFW representation.
ver.0.1 Real (FP8/Q4KS)
I've also used a FineTuning model for the Real series. The dataset is quite different from 0.0, but in a nutshell, it's a typical photo model.
I'm still testing FineTuning.
It doesn't have very high reproducibility with existing prompts like SDXL's, but there are often problems with the prompt itself. I'd appreciate it if you could report any scenes that don't work well.
I'm currently investigating the cause of the vertical stripe noise in the image.
ver.0.0 Anime (FP8/Q4KS)
This is an anime model using Finetuning. It takes a very long time, so currently it's only been 5 epochs. I'll decide how far to train it based on the progress.
Looking at the output, there are problems with body stability and prompt following, so it looks like some adjustments to the training data will be necessary.
This is my first time using FT with Qwen. It seems to be more effective than LoRA, partly because there was a large amount of data.
Of course, you can also adjust the style using Anime LoRAs, which has already been released.
I have prepared FP8 and GGUF Q4KS versions.
ver.0.0 Real (FP8/Q4KS)
This is my first model with Qwen-image.
To be honest, I haven't fully adjusted it yet, so it just happened to look like this.
I'm hoping to retain the good parts of the original Qwen-image while adding some new elements.
Description
FAQ
Comments (5)
is there a Q8 gguf? The fp8 model doesn't like certain startup args in comfyui so it looks a bunch of unfinished pixels. Strange thing is I can use other models' fp8 version just fine, it's only this one that's problematic.
Thank you for the information.
I checked the FP8 file I have and found no major issues. CIVITAI's uploader is very unstable, so I re-uploaded the FP8 version. I also uploaded the Q8_0 version of GGUF, as requested. Since it is difficult to upload models with the same bit count simultaneously in CIVITAI, one of them is listed as the Full version, but the content is the Pruned version. In theory, the image quality of GGUF should be better than FP8, but there is not much difference visually.
@aueki4g467 I'm sorry I was quite vague. When using the --fast argument it generates broken images. This isn't the only model now, after extensive testing, but base models are perfectly fine, either gguf or safetensors. Thank you so much for the Q8, I figure I will simply turn off the --fast arg where required, too :)
Do this model support edit or can just t2i?
No. This model only supports normal Qwen-Image model. I don't know about Edit version training or control images at all. I searched for it quite some time ago, but I couldn't find any information about it at the time.