can you share some style image demos to help us reproduce the results ?

[t00350320]

Sorry for can not upload images, please click this url

https://www.imagehub.cc/image/re.bvSNaz

we prefer

https://www.imagehub.cc/image/%E6%B0%B4.bvSODj

as you can see, we want to transfer the user input to a water style, but the csgo results seems not good as expected.
How to improve the result in csgo?
can you share some style image demos?
Or can we make a plugin Lora quickly based on the original CSGO.bin in future?

[xingp-ng]

There are actually no constraints on style images. You can input images similar to the journeyDB dataset as candidates for style and content images. I think this is related to the base model SDXL. SDXL's ability to generate realistic images may be relatively weak. You can try changing the base model. In addition, we hope to develop a FLUX-based version of csgo, which is in the plan. Besides that, we lack more effective style feature feature extractors focusing on style features. This is a pressing issue for the community. If you have batch images, I'm sure fine-tuning on CSGO can yield higher quality results.

[xingp-ng]

Sorry for can not upload images, please click this url

https://www.imagehub.cc/image/re.bvSNaz

we prefer

https://www.imagehub.cc/image/%E6%B0%B4.bvSODj

as you can see, we want to transfer the user input to a water style, but the csgo results seems not good as expected. How to improve the result in csgo? can you share some style image demos? Or can we make a plugin Lora quickly based on the original CSGO.bin in future?

I think it would not be difficult to reproduce the results in the paper. hhh, you can try these parameters。

[john09282922]

Hi, can you tell me how to fine tune with this method? It is not working well to specific style on real images.