One continuous 39-second shot from H3 segments, notes from a 12 GB card
#124
by LIOSAM007 - opened
Sharing what worked for us in case it helps. Identity holds if every segment after the first runs in reference mode, with the face reference and the establishing image as references and the full outfit repeated in the text. Over 10 segments and 53 seconds nothing accumulated. The model also plays a small action in about 1.2 s and then stands still until the segment ends, so we size segments by the number of actions: 2-4 per segment, at least 3 s. Negations pull the named object into the frame ("no cars" gave cars), and a cup has to be described by its state in each segment, or a second one appears. Write-up with videos: https://civitai.com/articles/38141