This is fantastic! But, can it use FLUX.1-dev's lora?
Furthermore, unlike VLM, image generation models don't have benchmarks, making it difficult to see improvements mathematically. Its functionality seems similar to Qwen-Image-Edit? I don't quite understand the concept of "combining image generation and editing functions into one."
However, are its hardware requirements lower than Qwen-Image-Edit? Since FLUX.1-Kontent-dev has much lower hardware requirements than Qwen-Image-Edit, I estimate that FLUX.2-dev's hardware requirements aren't very high either?
Compared to Qwen-Image-Edit, which requires 40GB of RAM to run, 32GB is more attractive, and if it can be run on 16GB (consumer-grade graphics cards) after optimization, that would be even better.
I discovered it has a 64GB model file and also includes an LLM chat template.
Therefore, I deduced that it's significantly larger than Qwen-Image-Edit.
If that's the case, it doesn't seem very competitive.
Because it's so large.
can it use FLUX.1-dev's lora?
No, it's a new model made from scratch
Simply, no it problably cannot, unless u find a way to reformat the lora. This is problably because the model architecture changed, and now requires its own pipeline. However, if you know what data previous loras were made with, maybe you can try to make a copy of the lora? ¯_(ツ)_/¯
Simply, no it problably cannot, unless u find a way to reformat the lora. This is problably because the model architecture changed, and now requires its own pipeline. However, if you know what data previous loras were made with, maybe you can try to make a copy of the lora? ¯_(ツ)_/¯
Forget it, I'll look for another way.