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
{{ message }}
Repository navigation
Generated results of DreamForge-DiT and DreamForge #64
I succeed to run DreamForge-DiT along with DreamArena with APIs for TrafficManager and DrivingAgents.
But for the generated results, it looks weired.
For the first few frames, it looks perfect, but for the later frames, the generated frames look noisy, "seems like past frame contents be directly passed to the next frame, making it look noisy".
I use dockers to create enviroemnts, so I think library version for the 3 modules in DreamArena including DreamForge-DiT should be mostly correct.
Here I share some samples:
1st frame:
3rd frame:
5th frame:
7th frame:
9th frame
You may see some clear noises since 5th frame.
Am I wrong in any step? If there is any cue, I would appreciate if you could share with me.
I would be appreciated if you could answer me just a simple question first.
I wonder if this situation happened during your experiments as well? Or just for me?
If latter applies, I will try to find possible solutions myself first.
Reacted by Yongjia Zou, Handsome-Fun and Peiyi Wang
I also encountered this problem. After five frames, the image gradually filled with noise. By the 30th frame, the image looked like this, and the scene remained completely unchanged.
I face a different issue for DreamForge (w/o DiT).
The scene appearance changes drastically every 10 frames, while the object position remain expected.
So I think the TrafficManager is keeping sending correct control signal.
But the temporal consistency for DreamForge only maintains within the 10 frame clip?
Here I attach some samples (Look at the changes of 10th->11th, 20th->21th):
1st frame
10th frame
11th frame (sudden change of scene appearance, but car position looks normal)
20th frame
21th frame (sudden change of scene appearance, but car position looks normal)
I also meet the same problem. I use the DreamForge-DiT and ran TrafficManager and Uniad synchronously. There is a lot of noise from the first frame.
WARNING:root:You did not provide box_latent for infer. I will sample w/o random seed!
I have noticed that WorldDreamer has been issuing warnings all along, and I am not sure if this is the reason.
Hello, thank you very much for your sharing. Currently, when I run the DreamForge-DiT + UniAD closed loop, I find that there are bugs in the DreamForge-DiT code itself. Could you please tell me if you directly reproduced the above result according to the Readme at that time?
Activity
yzli-0910 commentedon Jun 9, 2025
I would be appreciated if you could answer me just a simple question first.
I wonder if this situation happened during your experiments as well? Or just for me?
If latter applies, I will try to find possible solutions myself first.
jiakaa commentedon Jun 12, 2025
I also encountered this problem. After five frames, the image gradually filled with noise. By the 30th frame, the image looked like this, and the scene remained completely unchanged.
[-]Generated results of DreamForge-DiT[/-][+]Generated results of DreamForge-DiT and DreamForge[/+]yzli-0910 commentedon Jun 13, 2025
I face a different issue for DreamForge (w/o DiT).
The scene appearance changes drastically every 10 frames, while the object position remain expected.
So I think the TrafficManager is keeping sending correct control signal.
But the temporal consistency for DreamForge only maintains within the 10 frame clip?
Here I attach some samples (Look at the changes of 10th->11th, 20th->21th):
1st frame

10th frame

11th frame (sudden change of scene appearance, but car position looks normal)

20th frame

21th frame (sudden change of scene appearance, but car position looks normal)

gaoshanliushui58 commentedon Nov 25, 2025
CHJK-reinforce commentedon Apr 28, 2026
Hello, thank you very much for your sharing. Currently, when I run the DreamForge-DiT + UniAD closed loop, I find that there are bugs in the DreamForge-DiT code itself. Could you please tell me if you directly reproduced the above result according to the Readme at that time?