You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I am trying to use the Transfer model for world-scenario-map-controlled video generation. Specifically, I provide the first frame as the image condition, and use world-scenario-maps rendered from different trajectories as the control condition to generate the subsequent video.
However, I found that the generated video quality is quite poor. The video often shows abrupt changes in style and structure after the first frame.
I am trying to use the Transfer model for world-scenario-map-controlled video generation. Specifically, I provide the first frame as the image condition, and use world-scenario-maps rendered from different trajectories as the control condition to generate the subsequent video.
However, I found that the generated video quality is quite poor. The video often shows abrupt changes in style and structure after the first frame.
vision.mov
side_by_side.mov
Relevant configuration
model_mode = "video2video"
resolution = "720"
aspect_ratio = "16,9"
num_frames = 240
fps = 10
num_video_frames_per_chunk = 121
num_steps = 20
guidance = 1.5
control_guidance = 3
seed = 2026
num_first_chunk_conditional_frames = 8
num_conditional_frames = 1
condition_frame_indexes_vision = [0]
condition_video_keep = "first"
share_vision_temporal_positions = True
wsm_weight = 1.0