I'm 100% convinced Bytedance is nerfing 2.0 to push 2.5... 2.0 is just refusing to follow my prompts that worked 100% before 2.5. R2V is even outputting 9:16 or 4:3 even though all the assetsI used for reference, even the video is 16:9.. and I even specified it in the prompt and output settings on Venice.
Workflow Venice AI Workflows and Models (DON'T POST REAL PEOPLE!!)
- Thread starter LaniChuck
- Start date
You have to match the output with your reference image, video and audio. if you want a 16:9 output. Use a 16:9 reference image. video duration is 4-30 seconds. Setting it to -1, afaik, has to be in code prompts. I think -1 is random or will use 4 to 30 seconds depending on the prompt. 2.5 is very specific.
Seedance 2.5 Tips
Seedance 2.5 Tips
- • Seedance 2.5 is a director, not a search engine — write a shot brief, not a description
- • Real cuts come from explicit "cut to" or "Shot 1: / Shot 2:" labels — bare timestamps only set pacing inside one continuous take
- • One action + one camera move per shot; stacking camera moves muddies the motion
- • Wrap dialogue in double quotes with a tone note and keep lines short: she softly whispers "I am free"
- • For image-to-video: describe motion and camera only — the model already sees your image
- • Reference-to-video takes up to 50 assets (images, video, audio, 3D white-box) — give every @Image1 / @Audio1 an explicit role
- • Lens type + shallow DOF steer the look more than camera-body names; always close with the spec line (resolution, aspect ratio, duration, fps)
Bruno6
Lurker
- Joined
- Mar 18, 2022
- Posts
- 127
- Reaction score
- 0
depends on the error you're getting, could be an issue with the reference aspect ratio, could be an audio length issue, could be that you're trying to use a celebrity or someone who looks too young, could be you're reference video is too high or low quality
Joel Rook
Lurker
- Joined
- Mar 10, 2026
- Posts
- 9
- Reaction score
- 0
This question is somewhat vague because you didn't say what model you want to use. But you can do this with Seedance.
When you are on Venice, go to the Studio section, then find Seedance 2.5 R2V or 2.0 R2V (Seedance 2.0 R2V Enhanced on venice no longer supports video references)
Once you upload the video, upload the AI Generated Person you want to replace the person in the video with.
Once that's done a section will show up below your references that says "Try". Click on the box that says "Modify with image" It will autofill the following prompt:
"Replace in @ Video1 with @Image1, with motion and camera preserved."
After that you will have to add the rest of the prompt, the hardest part.
Here is an example of a prompt used with Seedance 2.0 for a BJ vid using a Reference image of a AI Generated woman and a blowjob reference vid:
Yes it does take some work to prompt, but laying out exactly what you want helps. It makes for a good output video. Plus with a prompt like this, you can get 30-40 uses out of a prompt with the same reference vid and It only had to be written once.
(Because there are no body or facial characteristics to add or delete.)
When you are on Venice, go to the Studio section, then find Seedance 2.5 R2V or 2.0 R2V (Seedance 2.0 R2V Enhanced on venice no longer supports video references)
Once you upload the video, upload the AI Generated Person you want to replace the person in the video with.
Once that's done a section will show up below your references that says "Try". Click on the box that says "Modify with image" It will autofill the following prompt:
"Replace in @ Video1 with @Image1, with motion and camera preserved."
After that you will have to add the rest of the prompt, the hardest part.
Here is an example of a prompt used with Seedance 2.0 for a BJ vid using a Reference image of a AI Generated woman and a blowjob reference vid:
Replace woman in the complication video in video1 with woman @Image1 , with motion and camera preserved. Woman is extremely close to the camera, performing a deep throated sloppy blowjob on the penis of a unseen man.
Copy the motion and structure of video1 with no deviation from her actions, the events are as followed:
@Image2 is the starting frame. @Image2 shows the subject with a penis fully submerged in her mouth, the man's testicles are in front of her mouth. In the first second, she lifts her head up removing the penis from her mouth and throat until it is fully out.
At .01 the subject is holding the penis in her left hand beside her mouth. she runs her lips up and down the penis as she strokes it. the penis is covered in spit and slobber.
At .02 seconds cut to the next part with the woman squarely in frame with the head of the penis in her open mouth. She quickly lowers her head onto the penis until fully throated. reaching the starting frame again in @Image2 with the man's testicles at the bottom of the frame pressed against her nose at around 2.7 seconds.
Immediately after at or around (2.9 seconds) another cut to the same camera angle, now show her with her hands underneath the man's testicles, lowering her face on the penis still full of spit, quickly deep throating the shaft until at 4 seconds. At 4 seconds the final cut to the same angle, the woman's hands are now removed, she's looking directly in the camera as the man face fucks her mouth for the rest of the video. She has half of the man's girthy penis in her mouth fully lowering her head down to the base of his penis and raising it back up.
@Image2 eyes and facial likeness are preserved with no alterations.
Generate sounds of moans throughout as she performs the blowjob. You can hear the sounds of spit, wetness, and slight gagging as she performs the act.
Copy the motion and structure of video1 with no deviation from her actions, the events are as followed:
@Image2 is the starting frame. @Image2 shows the subject with a penis fully submerged in her mouth, the man's testicles are in front of her mouth. In the first second, she lifts her head up removing the penis from her mouth and throat until it is fully out.
At .01 the subject is holding the penis in her left hand beside her mouth. she runs her lips up and down the penis as she strokes it. the penis is covered in spit and slobber.
At .02 seconds cut to the next part with the woman squarely in frame with the head of the penis in her open mouth. She quickly lowers her head onto the penis until fully throated. reaching the starting frame again in @Image2 with the man's testicles at the bottom of the frame pressed against her nose at around 2.7 seconds.
Immediately after at or around (2.9 seconds) another cut to the same camera angle, now show her with her hands underneath the man's testicles, lowering her face on the penis still full of spit, quickly deep throating the shaft until at 4 seconds. At 4 seconds the final cut to the same angle, the woman's hands are now removed, she's looking directly in the camera as the man face fucks her mouth for the rest of the video. She has half of the man's girthy penis in her mouth fully lowering her head down to the base of his penis and raising it back up.
@Image2 eyes and facial likeness are preserved with no alterations.
Generate sounds of moans throughout as she performs the blowjob. You can hear the sounds of spit, wetness, and slight gagging as she performs the act.
Yes it does take some work to prompt, but laying out exactly what you want helps. It makes for a good output video. Plus with a prompt like this, you can get 30-40 uses out of a prompt with the same reference vid and It only had to be written once.
(Because there are no body or facial characteristics to add or delete.)
No, not for frontier models. If you have a decent GPU, you can use open source models like Wan, but access to models like Seedance is always going to be paid in some form since the model providers charge for access. Venice is probably the best access I've seen of all the providers since they give you NSFW when possible.
Chaz1988
Lurker
- Joined
- Apr 30, 2022
- Posts
- 147
- Reaction score
- 0
Still in beta, and there's a queue. WaveSpeed has some of the most traffic so maybe longer than some other sites either way 1-3hrs is normal rn for any generation with wan3. Full release should be by end of month. And hopefully there won't be long queues anymore