🎓Seed, Aspect Ratio, and Duration

Haiper Settings Guide

Steps, Guidance, Sample, Schedule, Image weights, Aspect Ratios, there's seemingly no end to the number of small parameters that can be dialled in on your quest for the perfect AI generated video. Fortunately, with Haiper video gen we've gone and done most of the hard work for you, leaving you to concentrate on the only thing that really matters, the prompt!

However, we do still have some settings to play around with, and some important quirks for AI generation to make note of. For most, if not all modes of creation on Haiper, there are three settings you can adjust that will have the greatest effect on your result, that's the Seed, the Aspect Ratio, and the Duration.

The Seed

Every generation begins with a pattern of random gaussian noise that is defined by the seed number, with no two being alike. A denoising pass is then applied to this noise by a model specifically trained to understand the patterns of noise in images and predict what a clean version of the image should look like.

This process involves training the model on pairs of noisy and clean images, allowing it to learn the relationship between the two and develop the ability to remove noise effectively. Once trained, the model can then apply this learned knowledge to new, unseen images, effectively reversing the effects of noise and producing entirely new images, and in turn, new video!

So in theory, there are an infinite number of first-frame images of "giant stone statues" inside the video model, and an infinite number of videos that can be generated from that first-frame. Videos with the same seed will be generated from the same noise pattern, meaning that with a comparable prompt, things may turn out to look rather familiar.

The best use of the seed is to "lock-in" a generation you like so that you can make more fine tuned adjustments. Notice in each of these examples how the rough position and angle of the statue is the same and how the smaller islands in the distance are roughly in the same location.

In the next example, we lock the seed in a standard image generator and gradually increase the complexity of the prompt. Note that the longer the prompt, the more subtle the changes become. It's also important to remember that not all prompts are made equal or applied equally, so key phrases like "depth of field", and "bright vibrant colour" have a far stronger influence on the final image. Single, popular words like "wizard", "glasses", "forest" that are incredibly easy for the model to apply also have a strong influence.

The Aspect Ratio

On the surface, the aspect ratio is about as simple as it sounds. But there are a few things to take note of for the best results. Remember how the seed number dictates the noise pattern the image is generated from? The aspect ratio of your image is going to impact that pattern, and in turn, your result. So locking the seed will not generate identical results between different aspect ratios, only similar ones. Below is an example video of the same prompt "The red fox runs in the mountains and rivers, especially cheerful" using the same seed, but different AR's.

Secondly, the Haiper model is trained specifically on common video aspect ratios such as 16:9. This means that the videos generated at these exact AR's will be the highest quality, since they can more accurately rely on the data from their training. Videos generated outside of these AR's (21:9) for example, will be more likely to contain more errors, such as stretching, or duplicates.

This also has a significant effect on the composition of your image beyond detail and clarity. The model will often try and centre-frame most of your subjects and avoid any awkward cropping or extreme angles. (Note the foxes generated in different aspect ratios from above, and how the wider aspect rations showed more of the foxes body to fill out the frame, whereas the taller videos had the fox directly facing the camera). For example, a prompt for a "tall stone tower", a "giraffe", or a "very big skyscraper" may generate far higher quality in 9:16. So make sure to experiment with how different aspect ratios affect your image, and consider how you might use that to your advantage.

This also has a significant effect on the composition of your image beyond detail and clarity. The model will often try and centre-frame most of your subjects and avoid any awkward cropping or extreme angles. (Note the foxes generated in different aspect ratios from above, and how the wider aspect rations showed more of the foxes body to fill out the frame, whereas the taller videos had the fox directly facing the camera). For example, a prompt for a "tall stone tower", a "giraffe", or a "very big skyscraper" may generate far higher quality in 9:16. So make sure to experiment with how different aspect ratios affect your image, and consider how you might use that to your advantage.

For Text-to-Video there isn't too much to worry about, as the generations are locked in to specific AR's. For Image-to-Video the model will first stretch out your input image to the nearest usable size before generating the video and stretching it back. Whilst the difference in quality is negligible, it's advised to always input your images at an aspect ratio listed in the Text-to-Video settings. (16:9, 4:3, 1:1, 3:4, 9:16, 21:9).

The Duration

Much like the change in aspect ratio, a change in duration will also affect the noise pattern produced by the random seed, meaning that a longer generation (4 seconds) will not merely be an extension of what you have already seen in a shorter generation (2 seconds).

However, since the noise used to generate the first frame is the same, the two videos will share the same starting point. See below three examples of two second generations, and their four second counterparts.

You can adjust all of these parameters in the settings, located next to the prompt window over at haiper.ai/creations. For support and further guidance in generating AI video, you can reach out to us on Discord. We have a thriving community of talented creators that are eager to share their work as well as great tips on creating high quality content.