One of the downsides of large generative models is the entire video is essentially a hallucination, we try to prevent captions from showing up, but ultimately the model is non deterministic. For what it’s worth, we only saw one out of thousands of videos we generated with captions.
We’ll try to improve this with the next model.
Michael
Angel Garcia Crespo
•
Mar 28, 2025
Hi Michael,
Thanks again for your response.
I gave it another try, but unfortunately, I ran into the same issue — the subtitles appeared again, which makes the video unusable for our purposes.
I understand these things can happen with early-stage models, and I appreciate the complexity behind it. Still, I’d be grateful if you could consider some kind of refund or credit in these cases.
Looking forward to future versions — I'm sure they'll be amazing.
All the best, Ángel
Michael Lingelbach
Team•
Mar 28, 2025
Hi Angel,
We rarely see subtitles in our generations, what two images did you use? Or were they the same image?
Angel Garcia Crespo
•
Apr 1, 2025
Hi! I used the same image twice, and in the prompt I specifically asked not to include any subtitles. Here's the exact prompt I used:
A young woman, seated in a bright and modern library or study room, smiles naturally while blinking occasionally. She slightly tilts her head, her hair subtly moves as if a gentle indoor breeze is present. In the background, the soft lighting from the ceiling and the ambient light from the windows stay constant. Papers on the tables and nearby plants may show minimal motion, like a calm environment with life. The atmosphere should feel peaceful, academic, and slightly warm, with no camera movement or scene transitions. Absolutely no text, subtitles, captions, or written elements of any kind should appear at any point in the video. The visual should remain completely clean and free of on-screen writing.
Help me please!
Michael Lingelbach
Team•
Apr 2, 2025
The same image is likely to offer similar results
Angel Garcia Crespo
•
Mar 26, 2025
Hi Michael,
Thanks for the clarification — I totally understand the situation. It makes sense that with large generative models, some level of unpredictability is inevitable.
That said, I wonder if the issue might be related to the model having been trained on videos that included subtitles, possibly in Korean. In this case, unfortunately, the generated video is not usable for our needs.
Would a refund be possible in situations like this, where the final output ends up being unusable? I completely understand if it's not the usual policy — just thought I’d ask.
Appreciate your help and looking forward to future improvements!
Warm regards, Ángel
Michael Lingelbach
Team•
Apr 19, 2025
Hi there,
As a large generative model, hedra will occasionally hallucinate captions. It’s unfortunately not something that is very controllable. We see it very rarely, it mostly happens with things that look like “news reports”.
Paid users do not have the option to refund specific videos, it would not be viable for generative video companies to do this given the unit economics.
Michael
Will Pham
•
May 19, 2025
Hello,
Thanks for reaching out. Would you mind sending the starting image and audio file over for us to investigate?
Best,
Will
non NI
•
May 20, 2025
广å·å¿æ¿ç女主æ人é³é¢50ç§.MP3
717.2 KB• Audio
Will Pham
•
May 22, 2025
Hi,
Thanks for reaching out and apologies for the delay. Would you mind sending the starting image and audio file over so we can investigate?
Best,
Will
봉이김선달
•
May 22, 2025
I upload my work every day, so after much consideration, I had no choice but to use the video yesterday by covering the strange characters in the middle with another text. This was never my original intention. I’m planning to use Hedra again today, but honestly, I’m hesitant. I sincerely hope this issue can be resolved soon.
Michael Lingelbach
Team•
May 31, 2025
This is a known issue with our last generation model that will be fixed with the next one, this rarely occurs but does so reliably with certain images. Merging with the main ticket.
Log in to comment and vote
Comments12
Michael Lingelbach
Mar 26, 2025
Hi Angel,
One of the downsides of large generative models is the entire video is essentially a hallucination, we try to prevent captions from showing up, but ultimately the model is non deterministic. For what it’s worth, we only saw one out of thousands of videos we generated with captions.
We’ll try to improve this with the next model.
Michael
Angel Garcia Crespo
Mar 28, 2025
Hi Michael,
Thanks again for your response.
I gave it another try, but unfortunately, I ran into the same issue — the subtitles appeared again, which makes the video unusable for our purposes.
I understand these things can happen with early-stage models, and I appreciate the complexity behind it. Still, I’d be grateful if you could consider some kind of refund or credit in these cases.
Looking forward to future versions — I'm sure they'll be amazing.
All the best,
Ángel
Michael Lingelbach
Mar 28, 2025
Hi Angel,
We rarely see subtitles in our generations, what two images did you use? Or were they the same image?
Angel Garcia Crespo
Apr 1, 2025
Hi!
I used the same image twice, and in the prompt I specifically asked not to include any subtitles. Here's the exact prompt I used:
A young woman, seated in a bright and modern library or study room, smiles naturally while blinking occasionally. She slightly tilts her head, her hair subtly moves as if a gentle indoor breeze is present. In the background, the soft lighting from the ceiling and the ambient light from the windows stay constant. Papers on the tables and nearby plants may show minimal motion, like a calm environment with life. The atmosphere should feel peaceful, academic, and slightly warm, with no camera movement or scene transitions. Absolutely no text, subtitles, captions, or written elements of any kind should appear at any point in the video. The visual should remain completely clean and free of on-screen writing.
Help me please!
Michael Lingelbach
Apr 2, 2025
The same image is likely to offer similar results
Angel Garcia Crespo
Mar 26, 2025
Hi Michael,
Thanks for the clarification — I totally understand the situation. It makes sense that with large generative models, some level of unpredictability is inevitable.
That said, I wonder if the issue might be related to the model having been trained on videos that included subtitles, possibly in Korean. In this case, unfortunately, the generated video is not usable for our needs.
Would a refund be possible in situations like this, where the final output ends up being unusable? I completely understand if it's not the usual policy — just thought I’d ask.
Appreciate your help and looking forward to future improvements!
Warm regards,
Ángel
Michael Lingelbach
Apr 19, 2025
Hi there,
As a large generative model, hedra will occasionally hallucinate captions. It’s unfortunately not something that is very controllable. We see it very rarely, it mostly happens with things that look like “news reports”.
Paid users do not have the option to refund specific videos, it would not be viable for generative video companies to do this given the unit economics.
Michael
Will Pham
May 19, 2025
Hello,
Thanks for reaching out. Would you mind sending the starting image and audio file over for us to investigate?
Best,
Will
non NI
May 20, 2025
广å·å¿æ¿ç女主æ人é³é¢50ç§.MP3
717.2 KB• Audio
Will Pham
May 22, 2025
Hi,
Thanks for reaching out and apologies for the delay. Would you mind sending the starting image and audio file over so we can investigate?
Best,
Will
봉이김선달
May 22, 2025
I upload my work every day, so after much consideration, I had no choice but to use the video yesterday by covering the strange characters in the middle with another text. This was never my original intention. I’m planning to use Hedra again today, but honestly, I’m hesitant. I sincerely hope this issue can be resolved soon.
Michael Lingelbach
May 31, 2025
This is a known issue with our last generation model that will be fixed with the next one, this rarely occurs but does so reliably with certain images. Merging with the main ticket.