FireRed-Image-Edit-1.0 - WOW... Qwen image 2.0 in advance ? need AIO
Hello @Phr00t , you've probably already seen it, but the “FireRed-Image-Edit-1.0” model has been released. Based on my analysis, it appears to be an independent EDIT version, fine-tuned on Qwen Image 2512. The model is very good. Do you think it would be possible to make a quick AIO with the LORAs you had already identified for the EDIT 2511 model, possibly removing the LORAs based on skin realism? Or would you prefer to wait for the QWEN image 2.0 version?
thx
As much as I wish a new version of This finetune, FireRed is a rare case where the finetune is worse than the base version. They promissed a model that could beat Nano Banana level but it cannot beat 2511, which is a shame...
You're absolutely right. I was excited at first, but over time, it's not as good. After testing it extensively, I noticed that it's very dependent on the input image. It can be the best or the worst, so it's very inconsistent. The consistency of the model is poor in all cases, probably because the dataset used for fine-tuning contains too many imposed faces. On the other hand, it seems that the vast majority of poor results are attributable to the very poor FP8 that's hanging around on Hugging Face. The creator has reported that he's working on a much better one.
In the meantime, we'll wait for Qwen Image 2.0, which looks much better, especially in terms of skin realism.
I tested FireRed-Image-Edit-1.0 with the FP8 version for an hour today. I'm not sure if it's due to quantization or if it still can't compare to Qwen Image-1.0, but the skin texture isn't realistic enough. I also increased the CFG values, but FireRed-Image-Edit-1.0's offset was too large, while Qwen Image-1.0 did a much better job. Okay, I'll stop here. We're fortunate to be using the best model currently available. Of course, I'm still looking forward to the release of Qwen Image 2.0.