Introduction to Dream Machine from Luma AI

Dream Machine is an AI model that makes high quality, realistic videos fast from text and images. It is a highly scalable and efficient transformer model trained directly on videos making it capable of generating physically accurate, consistent and eventful shots. Dream Machine is our first step towards building a universal imagination engine and it is available to everyone now!

[reifuTD]

I really like Dream Machine so far, It's had some untestable issues with traffic and queue. My big complaint there should be a daily allowance of free credits and not monthly, it could be as low as one or two day that build over time and caps out at like 30 or something. This would give Dream Machine a bit more like a Suno model to generation.

[NegariaDesign]

I found you couple days ago. Your works are incredible. I wish I could be part of your team

[ReflectionOcean]

By YouSum Live

00:00:08 Introduction of Dream Machine model
00:00:34 High-quality text-to-video generations
00:00:47 Emphasis on reasonable generation times
00:01:09 Dynamic and coherent motion generation
00:01:15 Consistent characters and physics understanding
00:01:31 Innovative camera motions beyond static angles
00:01:50 Image to video capabilities enhance creativity
00:02:30 Excitement for user creativity and future developments

By YouSum Live

[wholeness]

Can't wait for prompt guides, deep dives, and UI features!

[gumvue.studio]

please remove the video from website background, its slow down all . its also heavy load for your website. no need that video

[chenyangwu]

How can I delete the generated videos? Is there no such function?

[zhangzhang]

I just received this email message from Luma, I
thought I should share it: At this time, it is not
possible to delete Dream Machine generations.However, we are actively working on adding this feature. A Ul update that will introduce the ability for creators to delete their generations is currently in development.
Rest assured, your creations remain private and
cannot be viewed by others unless you choose to
share them.
Stay well,
Navi

[lucase.garcia]

I appreciate and am pleased with the success of your Dream Machine in recent days. A great success.
If possible, try to expand the processing spectrum somehow, the creations take a long time to wait, plus they are now limited to only five for free accounts.
I recommend adding options such as negative prompts and giving the AI the possibility to identify "eyes, mouth, nose, legs and arms, cars, trees, etc", also an option to "stay true to the shape" of a given object or character face so that it varies as little as possible.
This could help AI, not only to improve results but also to perceive better when composing something.
It is true that the more one uses it (a lot in the last days) the more it learns and the better things it does.
But I think these additional options would give it an interesting boost.
Keep it up. This is very promising.

[izsit_ai]

Wow some of those generations look amazing!

[FrappuccinoAlfredo]

Thank you for releasing this just weeks before the presidential debates 😃😃😃

[kareemebaid]

@lumaAI, Team, I want to make an observation about frame drops in Release 1.5. The videos become laggy and have severe frame drops in all motions. I could not find a support link to report the issue.

[Veneglock]

Luma no me deja entrar dice que debo suscribirme que paso con la versión gratis?

[Kailamptey]

It has mastered physics, but not biological movement. I think you're going to need RLHF, to provide the biological feedback. You have a facility where it can generate videos frame-to-frame. It's basically gonna have to use Q-learning to estimate the score it gets from human estimation of the quality of the generated video (you're basically gonna have to find social media type photos which are relatively easy for the algorithm to transition between, which don't have any actual video frames for a diffusion model to learn. You use a "many heads" human feedback mechanism to get an estimate on the quality of the generated video (above 9 is perfect, indistinguishable from real life), then you use q-learning to have the model estimate the score. Over time it learns to estimate the score it's videos will get, and learn to produce videos indistinguishable from real life. Then you have to reduce the temporal resolution, increase the temporal length, train the traditional diffusion model (to predict the same frame at time step t as the RLHF original model, but lower temporal resolution, higher temporal length), then when it's trained, you repeat the RLHF process in the same way, q-learning to get it to understand biological movement. Then the same again, etc, etc. until you hit the model's limit in temporal resolution/temporal length frame prediction. With this process, I believe you will get around the problem the model has with biological movement.

[denniswhite]

I did a text to video. It was a simple bird in a clear blue sky, slowly flapping its wings. It has been in the queue now for 15 min. Have I done something wrong or is this a typical time frame?

[ChosenOnesUniverse-ehgw]

is there an altarnative to use luma free plan for monitization

[starjupiter]

Hello, I used Luma AI for 2 months, I wanted to spend my credits yesterday because the money was too much and then I canceled the subscription. When I download the videos I created to finish my credits, I see a watermark. But I paid you 39 dollars to make a video without seeing a watermark. Now I'm downloading with a watermark. Can you explain the reason for this?

[AIForHumansShow]

first, great job y'all.

second, would love to know what it looks like on the backend meaning super curious to see how your spikes in traffic have affected times + generations etc

[AnthonyAntmanAcosta]

How do you delete the videos after you are done with them?

[WNR_Wiki]

I am telling you that if these non-free plans were removed and the waiting list was also removed.. it would be better for all users of your site and specifically your dream machine.. so that your website would be 100% free :)

[SimplyMartin]

I honestly didn’t see a single normal prompt so far. It’s not bad but defo not what the ad shows.