Create a video using only the start and end point images。The most cost-effective practical method for AI video generation

AI video generation is rapidly evolving、In actual operation, the generation cost becomes a surprisingly big problem.。

Especially when producing videos for SNS.、The number of trials becomes more important than the finished product.。

No matter how good the model is、It is rare that the ideal image is output in one shot.。Rather, search for direction while generating multiple times.、It will be a process of gradually increasing the degree of completion.。

That's why I've been using it a lot lately.、1080"Video 2.5 Turbo" allows you to generate p.5 second videos at low cost.。

Not the highest quality model。

However, it has sufficient quality for SNS posting and short video purposes.、Above all, production costs can be significantly reduced.。

In AI video production, it is often more important to be able to try again and again than to be able to hit it with one shot.。

In other words、In AI video production, the number of trials may determine the quality of the work rather than performance.。

If you think so、This low-priced model is very good.、become a realistic option。

and、What I feel is particularly practical about this cost-effective model is、This is video generation that specifies the "start point" and "end point" images.。

Video generation that makes AI think about the process

This method is very simple。

・Specify the first image (starting point) of the video

・Specify the last image (end point) of the video

・Let AI fill in the gaps

That's all there is to it。

For traditional video production、Humans had to design camera work and subject movement from start to finish.。

However, in the case of AI video、Just give the start and end points、AI will infer and generate the story and movements during that time.。

This is a completely different idea from traditional video production.、It can be said that this is a production method unique to AI.。

The starting point image and ending point image used this time

The image used this time is not a live action、Both are AI-generated images。

First, generate an image that shows the whole body.、After that, a close-up image of the face was generated.。

In other words、

“A person is walking towards me.”

The starting point image is、

"Close-up of a person"

I prepared an end point image called。

Input these two images to the video model、AI automatically considers the movements in between。

come walking。

gradually approach the camera。

stop somewhere。

becomes up。

AI automatically creates this series of flows.。

Of course, it doesn't work out exactly as I intended.。

but、It often assembles natural images without detailed human instructions.。

In fact, the AI ​​is doing some pretty unreasonable camera work.

This video is quite difficult to actually shoot.。

the subject is walking。

The camera is also moving。

Furthermore, there is a transition from a pulled-down image to a close-up.。

In some cases, gimbal or dolly photography may be necessary.。

Distance management to the subject。

focus management。

Camera movement speed。

Synchronization with walking speed。

All these elements must be fulfilled。

I think it will not be an easy shoot even for experienced cinematographers.。

However, AI does not understand such complex shooting processes.。

Simply look at the start and end points、It just fills in the gap。

nevertheless、Creates camera work that looks like it was actually taken by a human。

This is where the fun of AI video lies.。

Rather than recreating the technology、As a result, images that look like those are generated.。

I failed the first time

However, a failure occurred during the first generation.。

My costume has changed。

The depiction of hair was also unstable.。

This is a very common phenomenon in videos of people.。

Especially like this time、

・Large amount of movement

- Large change in angle of view

・Move from full body to close up

When conditions such as、Increased difficulty in maintaining consistency。

AI does not understand people themselves。

Therefore, the shape of clothes and hairstyle may change during the process.。

My basic policy is to leave it to AI first.

The method I use in situations like this is simple。

First, create a starting point image and an ending point image that can be used by AI.。

After that, no detailed control、Leave it to AI once。

Look at the results first。

Hire if good。

If it doesn't work, fix it。

This is the flow。

Don't try to control it by writing a bunch of prompts from the beginning。

because、This is because AI can sometimes produce good results that humans would not expect.。

Check your abilities first。

Then fix only the missing parts.。

This is a bit similar to photo retouching.。

Don't just touch on the details from the beginning、It's like looking at the whole thing and adjusting only the necessary parts.。

What was interesting was、The AI ​​has made a gesture where it turns around and looks to the side.。

It was a movement that a human would want to add as part of the performance.、This time I didn't use it because I wanted to show a simple flow to the final image.。

5In the short span of seconds、Rather than introducing a play that moves the gaze or looks back.、I felt that the video would be more cohesive if the amount of information was concentrated on the movement of the subject and camera work.。

They also come up with ``unexpected but attractive proposals'' around this point.、I think this is an interesting part of AI videos.。

This modification prompt

The correction prompt we added this time is very short.。

``With natural movements、To avoid image corruption。Be careful about the depiction of clothes and hair.''

That's all。

I didn't write a long prompt.。

However, in the second generation, it was greatly improved.。

Long text prompts are not always advantageous in AI videos。

In fact, it is often better to clearly communicate just the problem.。

This time was also a typical example.。

Let the audio take its course

Audio is basically left to AI.。

Of course, you can also add sound effects later.。

You can also add background music。

But first, let's check the generated sound。

This time, in the second revised version, environmental sounds like the sound of waves were generated naturally.。

It also went well with the video。

Therefore, this time we are using it as is.。

In AI videos, the coincidence of not only the images but also the sounds is interesting.。

Unintended environmental sounds and atmosphere can sometimes add to the appeal of a work.。

The number of trials is important in AI videos

What I felt again this time was、What is important in AI video production is not the perfect prompt, but the number of attempts.。

Decide the start and end points。

first generate。

See results。

Fix any problems。

generate again。

This repetition is the most efficient。

Especially if it is a low-cost model, you can try it as many times as you like.。

AI video comes from an era when humans designed the details.、We may be moving to an era where humans decide the direction and let AI interpret it.。

give only the start and end points、Let AI imagine the time。

I feel that this method will become even more practical in the future.。

▶︎[60fps slow motion production workflow that solves the 30fps problem of AI video with Premiere Pro]