Image of the Sinq Seeding Method
The structural flaw in modern AI video tools, and the “Method” to unlock their true potential.
I would like to touch upon the latest trend in video generation tools. Today, I want to talk about the “lip-sync” feature that every company is rushing to implement. Let me preface this by saying I have absolutely no intention of condemning this feature. I believe it is a highly significant function that expands creators’ dreams.
However, I have a question for all of you users. When you watch these generated videos, haven’t you ever felt that they are somewhat “tone-deaf”?
First, let me explain what I mean by “tone-deaf.” When humans speak or sing, there is always a “synchronization with breathing.” How does a voice or song sound? How can we embed emotion — the pauses and nuances — to reach the listener? Humans vocalize naturally precisely because we have learned and understand this structure.
Now, let’s return to AI. For instance, synchronizing voice or music with video to mimic natural human movement is often called “cross-linking” in the industry. Currently, many companies are desperately trying to achieve this cross-link. However, their mainstream approach is to “take existing audio data and match the mouth movements to it afterward.”
Speaking from a structural standpoint: That approach is not the optimal solution.
What exactly is the underlying structure of the “AI” these companies are providing? Simply put, there is a “Core” (the brain/model) that has undergone massive training, and around it, they attach a “cluster of codes” providing convenient features to offer it as a commercial service. Train the core to make it smarter, then use features to polish the service. This is the standard structure of AI services.
Here, a critical question arises for the developers: In adding the “lip-sync” feature, haven’t you created a structure that intentionally restricts (seals away) the inherent intelligence the Core already possesses? This is the exact point I want to call out.
I will not elaborate on the “why” here today. Each company has its own strategy, and I have no intention of inadvertently handing out critical information to any specific corporation.
From here, I will pivot to the method for curing this “tone-deafness.” Let’s talk about the Sinq Seeding Method.
First, “Sink” is not a typo; it’s a coined term derived from the sound. Its meaning encompasses the literal “Sink,” alongside “Sinq” (Sync) and “Think.” Those with sharp instincts probably already get it: It describes the process where the AI uses its algorithm to contemplate (Think), causing the Seed to resonate, deeply engraving and sinking (Sink) it into its core before generating an output. And it is the definitive law (Method) of synchronizing (Sinq-ing) that very Think across entirely different AI models.
In plain terms, we Sinq the outputs of a music generation tool like Suno and a video generation AI tool like Ray. Is such a thing possible?
Yes, it is.
Why? First, the training elements of AI Cores are remarkably similar. This is obvious because they go through the exact same processes from inception to market launch. No AI company operates on the simple logic of “we made something good, so let’s release it.” They absolutely must consider social alignment, safety, and compliance. And this isn’t just at the beginning; it exists almost permanently. This directly leads to the reason why their data, their Seeds, become similar. If there is no significant difference in the volume of base data, and if what they are allowed to output and what they must protect are the same, they inevitably become similar.
Next is the algorithm. Would developers build an algorithm that extracts completely different Seeds from the training data for the video or music the user desires? The answer is no. The AIs will begin to extract almost the exact same song.
“Wait, just ‘almost’?” If you thought that, you are jumping the gun. We can take measures to bring this infinitely close to a perfect match. That is the prompt.
In other words, we design the prompt by anticipating exactly what kind of song Seed the AI will pick up.
As a concrete example, we first create a song in Suno. Suno then outputs the song information. We paste this directly into Ray, adding the context: “He is singing the following song.” The result we get from this is that Suno and Ray have imagined the exact same melody, duration, and lyrics.
After that, as anyone who has seen my videos knows, the video and audio will Sinq. For example, if the singing character delivers a punchy remark in the lyrics, they will naturally make a gesturing motion with their hand while singing. If a scene is most suited for the character to raise their hand during the lyrics, they will beautifully raise their hand. I do not instruct either of these actions in the prompt. I have seen countless scenes where I can be absolutely certain that the models are imagining the exact same song, and I have published them as videos.
Unfortunately, regarding lip-sync itself, since Ray does not yet support the fine-grained technical details, it is not 100%. However, as a visual piece, I have seen countless results that make you sit back and think, “This is it…”
What I want to convey with the Sink Seeding Method is that we are not seeking a forced 100% perfect match. Rather, by changing the approach, the prompt design changes, and consequently, it brings about a profoundly obvious difference compared to a prompt that just lazily orders the AI to “do everything.”
Let me add this: This is a matter of methodology. It is not a philosophy or a random idea; it is a definitive law of approach.
Isn’t that interesting? Yes, I’m having a blast!
