Suggestion: make an edit model based on MOSS Image
Make an edit model from the MOSS Image base. https://huggingface.co/datasets/inclusionAI/ConceptEdit-12M can be used as the dataset, seems rather good. Ideally, editing with many inputs would be good, but there is no dataset for it
The edit/ref method doesn't need to train the whole model from scratch. can use the community approach—like how they trained Krea 2 Edit. Plus, the text encoder (TE) used by Moss has a built-in vision encoder.
The edit/ref method doesn't need to train the whole model from scratch.
Indeed, that's how edit models are often trained, by adding image conditioning to a T2I model and training on a proper dataset
can use the community approach—like how they trained Krea 2 Edit.
I'm asking for the official edit model
Plus, the text encoder (TE) used by Moss has a built-in vision encoder.
This isn't really editing. While this works for models that use VLMs, it's just crude reference
can use the community approach—like how they trained Krea 2 Edit.
I'm asking for the official edit model
I wasn't replying to you. I was just leaving a message. 😃
Thank you so much for your interest and the great suggestion! We really appreciate the community's feedback. Currently, we are actively exploring reference-related features and capabilities to improve the model. We'll keep you posted as we make progress. Stay tuned!