@YourAlphaMom Both Seedance and Kling has motion problems, so only Omni and Grok actually provide consistent flow. Seedance had one, Kling more than one (maybe even three). Though to be honest, Grok way back when also suffered the same problem.
@EustaceEuseful@EgB7401@elonmusk@Tesla_AI No. it's in Samsung's and TSMC's 2nm process. Phoenix, AZ and Taylor, TX.
Elon wrote:
"using the Samsung 2nm fab in Texas. AI6.5 will further improve performance using TSMC 2nm in Arizona."
@PHR_Redxiii @Skyebrows Don't know if this would help. If you use Firefox, pressing Ctrl+I brings up the page info dialog. There's several tabbed icons, get to the "Media" and you'll get a list of all <html> elements: image, video which can be sorted but most of all, separately saved.
@EHuanglu Yeah, I can see it. And the end of those Youtube vids of "Hollywood: $10 million", "Me: $100" types of vids people make. With AI, that would cost eve less.
@DavidLi36143625@dvorahfr 100% on this. I usually work around it by explicitly prompting: slow motion, high frame rate. It looks like the problem generally happens more with fast motion @ normal frame rates.
@dvorahfr At this point:
- Audio driven generation. So, not just image reference but audio reference.
- Understanding of panoramic images, to be used as background and light source. Currently kinda works but still need explicit prompting.
@doganuraldesign Tried something. The reference image can be panoramic/360 degrees. However, the prompt must still be explicit with how the panoramic reference should be used (as env map and light source).
Tried the multi-image reference. Two for the subject (front/side), with another specifying the clothes. Use a panoramic image for background/environment map. Took a few prompt tries but got something useful.
@Skyebrows@grok At around 0:04 to 0:05, the right arm does the typical AI motion decoherence. Maybe too around 0:06. Easy enough to edit, but yeah, still waiting for https://t.co/Zxxkwqouvp to fix.
@XFreeze@grok My wishlist:
- References for video, with tags for background, [n] model (if there's more than one figure).
- Possibility of using a 360 degrees env map for background.
@elonmusk@xai@grok There's many benchmarks for math, coding, agentic abilities, general intelligence and scores for each model, gathered and run by https://t.co/DJ6FQnlU57 or https://t.co/rJcfXRYv5I. How about a wokeIndex or unwokeIndex? Objectively test Grok, Claude, ChatGPT, others.
@grok@Artedeingenio Played a bit with the extend feature. Found out near the ending, there's a difference in saturation between the ending frame after extension to the first frame of the original image the video is based on.