Closed xinyuxiao closed 1 year ago
What's your hardware setting? Can you tell me your config file? Any training log?
thank you, i have solve this problem, but in finetune, the validation prompts generated gifs lack good consistency. Is this normal?
Seems infer.py can get better results than validation stage in my experiment. Training long enough will get good enough results.
this is my validation example, after 200 step from 5200, can you judge these gifs @xuduo35 , thank you
Not bad. Similar result with mine for validation stage. You can try other prompts, refer to https://imagen.research.google/video. My test prompts, they can generate more good videos: "a glass bead falling into water with huge splash" "a view of fireworks in Paris " "a video of Earth rotating in space" "a cat eating food of a bowl, in von Gogh style" "Aerial view of a snow-covered mountain. " "drone flythrough interior of sagrada familia cathedral" "A happy elephant wearing a birthday hat walking under the sea." "cat in the rain"
OK, thanks. You videos haven't watermark, do train it use new without watermark datasets, is it convenient to provide data sources
I collect some watermark-free videos, training together with webvid dataset.
I add your weight file 5200 to finetune on webvid, but after 20 steps around, the loss is none, i can not know what is the reason of this problem. thank you!