._. o_o O_O
Kling 3.0
It’s time to talk about the current flagship in video generation. As you’ve probably already heard, the release of Seedance 2.0 has been postponed.
Detailed version first, short and dry summary below ^_^
The model handles multi-scene videos really well, and now it also does lip sync well in 5 languages (Chinese, English, Japanese, Korean, Spanish). It can generate videos up to 15 seconds, but there are some tricks under the hood — based on tests, it seems to use something similar to FramePack. That means in your generated video, around 3–5 seconds may have little to no active movement/camera motion, and lip sync may become inaccurate in the final seconds.
As a result, I recommend generating videos up to 10–12 seconds.
Elements — you can now use them both for product placement and for making characters more consistent. The input supports not only images, but also video references (3–8 seconds), and all elements can be linked inside prompts.
They improved not only the main model — Omni quality is also noticeably better now. But without references, the result still isn’t perfect, so don’t be stingy with them — add references and you won’t regret it.
Now the approach to generation and prompting can be very different compared to previous models. In essence, there are 3 approaches:
Multi-Shot Off (one continuous video, no editing/cuts, like before)
Multi-Shot Auto — one prompt where you describe scenes (Shot 1: ..., Shot 2: ..., etc., all in one prompt)
Custom Multi-Shot — each scene has its own separate field, where you set timing, write the prompt, and attach the elements I mentioned above
Finally, we can say goodbye to the uncanny valley — Kling 3.0 is incredibly good at emotions, and they no longer trigger that internal rejection.
Now for the downsides — the model loses badly to Seedance 2.0 in action scenes. Trading blows / hand-to-hand exchanges basically doesn’t work at all. Gunfire looks better now, but censorship is still present. For the hostage action scene I attached here, I used a prompt from the PsyEyes
My overall feeling: you can now generate short films with real enjoyment and style. While we’re waiting for Seedance 2.0, it’s a great time to level up your multi-scene skills in Kling, since the prompt structure is almost identical.
o_O Short version O_o
15-second video generation with excellent quality (in practice, 10–12 seconds is better)
Support for 5 languages for speech/lip sync (Chinese, English, Japanese, Korean, Spanish)
Multi-Shot lets you create multiple shots/scenes in a single video or do split-cut editing
Works well with Elements (assets/references), and now you can also use 3–8 second videos as references — product placement is even easier now
The Omni model has been significantly improved, but if you don’t cheap out on adding elements/references, the final video will be even better
Kling 3.0
It’s time to talk about the current flagship in video generation. As you’ve probably already heard, the release of Seedance 2.0 has been postponed.
Detailed version first, short and dry summary below ^_^
The model handles multi-scene videos really well, and now it also does lip sync well in 5 languages (Chinese, English, Japanese, Korean, Spanish). It can generate videos up to 15 seconds, but there are some tricks under the hood — based on tests, it seems to use something similar to FramePack. That means in your generated video, around 3–5 seconds may have little to no active movement/camera motion, and lip sync may become inaccurate in the final seconds.
As a result, I recommend generating videos up to 10–12 seconds.
Elements — you can now use them both for product placement and for making characters more consistent. The input supports not only images, but also video references (3–8 seconds), and all elements can be linked inside prompts.
They improved not only the main model — Omni quality is also noticeably better now. But without references, the result still isn’t perfect, so don’t be stingy with them — add references and you won’t regret it.
Now the approach to generation and prompting can be very different compared to previous models. In essence, there are 3 approaches:
Multi-Shot Off (one continuous video, no editing/cuts, like before)
Multi-Shot Auto — one prompt where you describe scenes (Shot 1: ..., Shot 2: ..., etc., all in one prompt)
Custom Multi-Shot — each scene has its own separate field, where you set timing, write the prompt, and attach the elements I mentioned above
Finally, we can say goodbye to the uncanny valley — Kling 3.0 is incredibly good at emotions, and they no longer trigger that internal rejection.
Now for the downsides — the model loses badly to Seedance 2.0 in action scenes. Trading blows / hand-to-hand exchanges basically doesn’t work at all. Gunfire looks better now, but censorship is still present. For the hostage action scene I attached here, I used a prompt from the PsyEyes
My overall feeling: you can now generate short films with real enjoyment and style. While we’re waiting for Seedance 2.0, it’s a great time to level up your multi-scene skills in Kling, since the prompt structure is almost identical.
o_O Short version O_o
15-second video generation with excellent quality (in practice, 10–12 seconds is better)
Support for 5 languages for speech/lip sync (Chinese, English, Japanese, Korean, Spanish)
Multi-Shot lets you create multiple shots/scenes in a single video or do split-cut editing
Works well with Elements (assets/references), and now you can also use 3–8 second videos as references — product placement is even easier now
The Omni model has been significantly improved, but if you don’t cheap out on adding elements/references, the final video will be even better