Flux1-Fill-Dev model support

[zhou-linpeng]

Hi, thanks for your work. have you tried to change the base model to Flux1-Fill-Dev? Or are there any plans to support this base model?

[Orannue]

@zhou-linpeng Hi, Flux1-Fill-Dev is included in our plan : )

[zhou-linpeng]

Thank you. I'm looking forward to your new work.

Furthermore, I have a couple of questions:

• First, is it possible to achieve In-Context Learning through channel concatenation rather than spatial concatenation, in order to enhance the quality of the results during inference?
• Second, can negative pairs introduced as a form of prior information, effectively serving as a type of negative In-Context?

[Orannue]

Thank you. I'm looking forward to your new work.

Furthermore, I have a couple of questions:

• First, is it possible to achieve In-Context Learning through channel concatenation rather than spatial concatenation, in order to enhance the quality of the results during inference?
• Second, can negative pairs introduced as a form of prior information, effectively serving as a type of negative In-Context?

Regarding the first question, I believe channel concatenation is a plausible approach—ACE++ adopts a similar paradigm. However, it requires modifying the model’s input layer and additional training. Spatial concatenation remains the most straightforward method in this setting. Alternatively, you may decouple the query pair and the example pair and then integrate them within the attention block, similar to the approach used in EasyControl.

As for the second question, when using a supervised dataset, the model can typically learn the relationship between the example and query pairs quite easily. However, the relationship between a negative pair and a query pair tends to be more ambiguous. Without an additional mechanism to explicitly model this, it becomes difficult for the model to learn effectively. In my view, the task is already sufficiently defined by the example pair alone, and introducing negative pairs may not provide substantial benefit.