beta7 ignoring my ref while base ref h3 model works good with refs
i tried video and image refs of a girl, none worked.
in the below example prompt, depthmap pose is retained very well but the girl's face is different:
subject_definitions:
<Subject 1> is the girl in <Video 1>.
<Video 1> is the source video the target video continues from; its last frame is the exact starting point of [Shot 1].
<Picture 1> is a depth map defining the seated pose and the final camera framing for [Shot 1]; the depth map's grayscale appearance and silhouette body volume are not inherited.
retention_analysis:
<Subject 1> (appears in [Shot 1]): fully_preserved - the girl's identity, build, and full outfit are retained.
<Video 1> (continuation starting point): fully_preserved - [Shot 1] begins exactly at the last frame of <Video 1>, preserving the pose.
<Picture 1> ([Shot 1] final pose and camera framing): fully_preserved - the seated pose and the low three-quarter rear-side camera angle are followed exactly; the grayscale depth-map tone and silhouette volume are not inherited.
detailed_description:
The target video continues in the same bright bedroom from <Video 1>,...
It's one of beta 7's shortcomings, the input references become diluted to a general reference rather than a 1-1 likeness and even if you have a custom workflow using REF2VA multi-ref, VSA, 2-pass, etc with multiple input images for a single subject they tend to, in my testing at least, fight one another for control. Even attempting to overpower the model for control by trying to brute force likeness with an image collage, (2-3 images max), might gain you back a little bit of likeness but overall it will still be a general reference rather than close to base.
The main area beta 7 excels in is speed and that's the main reason to use it, but if you need the absolute 1-1 or close to likeness stick with beta 5 or the base model with loras.