Continuing the story with this project. I'm diving deeper and deeper into Google Flow, and I thought it would be good practice to make a video for this museum, especially since last week, all generations for Veo3 Fast were free if you have an Ultra subscription, so now you don't have to worry as much about spending all your credits and having to buy additional generations. It is worth mentioning that different features in Flow are available through different models, so it is not possible to make all parts of the video in Veo3 Fast; for many things, the service switches to Veo2 (which is not unlimited). In general, such a video takes about 4-5 thousand credits if you don't skimp and sometimes switch from Fast models to Quality. If you do skimp, you can easily get by with 1 thousand. Moreover, the difference between Fast and Quality lies in the models' structure, but in practice it is insignificant, and Fast generation can easily give better results simply because it knows or tries some other concept at the right moment. In total, you have 12,500 credits in the Ultra subscription (there was an announcement that they would be doubled, but this did not happen in my account; maybe they will update it from the next billing month).

Sound, captions, and editing in Adobe Premier Pro. Veo3 can generate videos with sound right away, but the audio won't be consistent from clip to clip, so you have to work with the sound separately to glue it together. I used free libraries from YouTube.

I thought out the plot of the generations in advance and made many attempts for each fragment to achieve the desired result, but in some places the AI model came up with interesting concepts that I left in. For example, the projection with the wolf turned out that way by accident. I didn't describe it in the prompt, I just asked for a video with a modern exposition about the northern lights and the Finnish forest.

By the way, my second workshop on the PA Academy platform starts this weekend. Grasshopper, Rhino.Inside, Revit, and a bit of AI. We'll be modeling this building from the video. #education
https://paacademy.com/course/bim-rhinoinside-for-advanced-tower-design

one more sample, this with Flux.Kontext

I tried making a few videos in Sora2, and here are the pros:
- The model itself is obviously cool; you can immediately see how well it handles physics and character movement.
- It's free for now.
- The clip length is 10 seconds, which is very good.

But there are also plenty of cons:
- Not a lot of settings and features on the site; in fact, you can only choose the format (horizontal or vertical video) and upload the first frame (except for people, they are not allowed).
- There is no option to extend or stitch clips together, meaning you can't make anything long, only simple short clips, but not a full-fledged video project. You also can't upload the last frame to make transitions, and you can't take the last frame from one clip to use it as the first frame for the next one. There are no additional modes, such as when you upload images of characters and scenes instead of the first frame, and describe the action itself only with text. Google has all of this.
- Watermarks make it impossible to use the service for work.
- There is no built-in Upscale.
- It works slowly and only produces one video at a time. While Sora2 makes one video, Google Flow can make three generations of four videos (12!) and you can pick the best one, which is important because the probability of successful generation on the first try is not very high, and you want to be able to choose between several successful ones.

Conclusion: Google Flow is more interesting in every way except for the model itself, which feels slightly better in OpenAI, but the difference is not significant enough to compensate for the lack of functionality for work. #ai

the best Upscale workflow i used so far. it based on Flux.1 dev and requires a strong graphic card, but the result is amazing. #comfyui

here is workflow itself