Introduction
Recently, the Runway Gen-2 model has made significant breakthroughs in its image-to-video generation capabilities, with some of the results being quite impressive. Therefore, I am writing this minimal usage guide for everyone to have fun with.
Usage Guide
This guide will introduce the costs, usage methods, and generation patterns of the Runway Gen-2 model. It is very easy to get started with Gen-2; the images needed for generation can be existing high-quality images or images generated by MJ or SD.
Runway website:
https://runwayml.com/ai-magic-tools/gen-2/
Usage Method
The usage method of the model is very simple: upload an image and click [Generate] to create; no prompts are needed.
Here are a few small details to note:
If you have the paid version, remember to turn on the [Upscale] option to generate HD videos, with the widest dimension reaching about 1800 pixels. Turn on the [Remove watermark] option to eliminate the watermark.
Do not upload images that are too large; images over 4k may get stuck and fail to generate. In this case, you need to manually cancel the task and upload a smaller image for generation.
The videos are stored by default in the Gen-2 folder. If there are too many videos in the folder, you can rename it and create a new [Gen-2] folder, and videos will automatically be stored in the new folder. The folder can be downloaded in bulk, which is very convenient.
Generation Patterns
Here are some patterns I have discovered; they may not be 100% accurate, and discussions are welcome.
Refreshing the same image multiple times generally does not change the major generation direction. For example, if the first generation is a rotation, refreshing several times will still be a rotation, just with different details.
Using high-quality and complex images increases the chances of achieving better results; often, the generation pattern simplifies the existing image.
Currently, it excels at generating abstract effects, such as abstract lines, animations, etc. The results for generating realistic characters are poor, while the results for IP are quite good.
Some videos may appear blurry during generation, while others may not. The probabilities of IP and abstract images not being blurry are relatively high.
There are many directions for video generation, including rotation, zooming in, dynamic IP characters, blending, etc. Some images may only have slight dynamic effects.
The video generation effect of Gen-1 is currently not very good, with low clarity and average results. Unless there is a specific need, it is not recommended for use.
Conclusion
If you want to discuss more details, you can also join the Stable Diffusion Alchemy Pavilion AI video task force to exchange secrets of the alchemy path with fellow practitioners, such as discussing the selection of herbs and control of fire, evaluation schemes for pill formation, etc. The path of immortality is long, let’s walk together, haha!
