I conducted a test using a demo, but I felt that the results were not very satisfactory, especially regarding the characters. The limbs were prone to errors

[libai-lab]

[Xinjie-Q]

Thanks for trying the demo and sharing your feedback.

Character generation and editing are among the most challenging scenarios for image models, as they require accurate handling of human anatomy, poses, and limb consistency. The demo currently uses our 4-step Turbo model, which is optimized for fast inference and interactive use, so image quality can degrade in particularly difficult character cases.

For higher fidelity results, we recommend using our RL model, Mage-Flow. It is designed to deliver stronger visual quality and generalization, and typically achieves much better results with around 20 sampling steps, especially for human-centric content, complex poses, and challenging editing tasks.

[libai-lab]

Mage-Flow

Thank you for your reply. Will multi-image editing be supported in the future? Will there be scripts for training and fine-tuning

[Xinjie-Q]

Thank you for your interest! Our current editing model already supports simple multi-image editing scenarios, such as style fusion, and can take up to three input images for editing. We are also actively exploring more advanced multi-image editing capabilities and plan to support more complex multi-image editing tasks in future releases. Regarding training and fine-tuning, we plan to open-source a lightweight fine-tuning framework in the future to help the community quickly adapt and customize the model for their own use cases. Please stay tuned for upcoming updates and releases! We appreciate your support and feedback. 😊

[bleakj]

Your images came out a lot nicer than my edits of my dog at least lol

[Xinjie-Q]

Your images came out a lot nicer than my edits of my dog at least lol

Would you mind sharing the example(s) you tried? We'd love to take a look at the prompts and outputs to better understand where the model struggled.
Also, if you're running the models locally, I'd recommend giving microsoft/Mage-Flow-Edit a try as well. It's our RL-tuned editing model and generally provides stronger editing performance on more challenging cases. For best results, we usually recommend using around 20-30 denoising steps.
We'd be very interested to hear how it compares on your dog-editing examples! 😊

[libai-lab]

I recommend this model. Its characters exhibit good consistency, and the model itself is not overly large. Could we consider adopting its approach to improve the current issue of poor character effects，Boogu-Image