Thank you for this, most closed models can't do it
I was stunned to see it working so well. Other than Seedance, which is ridiculously censored to the point a simple character kiss is blocked, only Kling V3 can do something similar. I'm not aware of any other model, closed or open source, that can allow you to input a background image, two characters, and get a pretty good result out of it with almost no bleed.
Think about using ggufs for the model, your workflow works well on my RTX 3060 12gb but uses over 60GB of ram, but HD resolution video can be done in under 8 minutes.
Tip: you can create their "reference sheet" for characters using Flux 2 Klein 9B or Dev, and also Nano Banana, which provides decent results as long as you input a couple of good images of the subject you want to include in the videos and caption them well using a llm. Vertical videos require you to make the reference sheets and background vertical or include a white margin to match the video's aspect ratio.
I was stunned to see it working so well. Other than Seedance, which is ridiculously censored to the point a simple character kiss is blocked, only Kling V3 can do something similar. I'm not aware of any other model, closed or open source, that can allow you to input a background image, two characters, and get a pretty good result out of it with almost no bleed.
Think about using ggufs for the model, your workflow works well on my RTX 3060 12gb but uses over 60GB of ram, but HD resolution video can be done in under 8 minutes.Tip: you can create their "reference sheet" for characters using Flux 2 Klein 9B or Dev, and also Nano Banana, which provides decent results as long as you input a couple of good images of the subject you want to include in the videos and caption them well using a llm. Vertical videos require you to make the reference sheets and background vertical or include a white margin to match the video's aspect ratio.
Bernini is able to.
it even works with eros/sulphur lora or merged model