Crypto Tries AI Generation

[Crypto]

My David's Day story...

[Crypto]

My Heather's Hike story...

[Crypto]

My Gail's Gallop story...

[Viridian]

Nice! I actually used David's Day as a story concept in 2024 when I first started animating. My result is on my DA page:
https://www.deviantart.com/viridianqs/a ... 1085084000

That's something I want to revisit and remake using my current workflow. Gail's Gallop was harder to put together as a sequence at the time. I believe I could faithfully recreate the scenes.

I do like how the way you envisioned your stories is exactly how I saw them. I am now working on a new attempt.

I'm guessing you're in the honeymoon phase of "Holy shit, I can make cinematic quality clips". We've all been there

In regards to your queries on output, particularly in the texture of the mud/quicksand, there are two vital things to consider:

Use Image to Video (or Frames to Video, for Google)
Google's Flow platform is its actual video generation platform that uses VEO, Gemini is the chat interface which doesn't allow for as many options. Virtually every AI model will not know how to animate mud or quicksand when prompted, so all the good outputs you see start with a source image, which we would typically generate first using whichever image model we prefer. I still use a workflow that uses SDXL that has a specific blend of photorealism that I like, but newer output models like Google's Nano Banana are going to generate exactly what you want with easier prompting. This gets even easier with photomanipulation skills. In short, without a visual source, it makes water.

The caveat, especially working with mud textures, is that it needs to "trick" the AI to recognising it as mud. The text prompt goes a long way (I like to use words like "sticky" and "clay" to avoid muddy water). Different AIs will read a mud visual differently. Most will very strongly lean towards water. In some models, I simply cannot use anything that has even a glisten of moisture since it will automatically turn the video into water. Other models are very responsive to even a wet mud visual.

Something like this is far more likely to work in most AI, though... see below.

jlue.jpg

Animation model
Most AI animation models are really good at specific things, very good at generally most others, but suck in other areas. A "better" AI model doesn't necessarily mean getting the ideal result. Again, we have a very specific requirement: the quicksand needs to be... quicksand. Which can come in different flavours depending on the context, and the AI doesn't contextualise that well. Hence, water.

Some AI models are trained on more sources that have the specific interaction we look for in a QS scene. Others will try or have to be coaxed with less consistency. Even the good ones have their own problems.

In short, my ranking of the AI models I've used specifically for a QS scene is: WAN-2.5 > Kling 2.6 > VEO 3.1

This is my result with WAN.

julie.mp4

WAN seems to have the most training on QS scenes and it virtually creates our preferred style. It has good audio, but bad voice lines and has an almost unusable level of artefacting.

Kling is the leader in its dynamic physical interaction, but its movement can feel floaty and sluggish. It's recently vastly improved its audio and voice. It's QS interaction is fairly good - it really does need the right visual source. I previously used it a lot more, but I was fighting to get good results consistently. I use Kling for anything that is not specifically a struggle scene, but it can do QS scenes moderately well and is very good at maintaining cohesion.

VEO 3.1 has fallen off, in my opinion. Its edge was audio integration and voice, but now other models have implemented it. Its animation quality is lagging behind. It's still very good, and the voice lines are still very good. It's QS ability is bad. It can barely keep a buried character from not turning it into water. My biggest issue with VEO is that it doesn't switch off its creativity. When you prompt it for a voice line, the character actually acts out the actions to a frustrating degree (like, "I hope there's no quicksand" and then they flop like fish into a puddle). Its launch was also horrendously buggy, and the Google AI plans are _very_ expensive for a service that isn't up half the time.

The last issue is censorship. Our videos tend to push it a bit, but most AI models struggle with filtering prompts and outputs. This is why I've not bothered with Sora 2 (expensive, very high censorship), VEO (high censorship), and Kling has struggled a lot too, though it is more willing to accept suggestive outputs. WAN is almost unlimited, especially when used locally, though the 2.5 and 2.6 models are only commercial use, so it depends on the host site. Higgsfield has almost no restrictions, so if you see QS sexual AI videos, it's most likely WAN-2.5 on Higgsfield. Other models have filters embedded in their algorithms.

Hope that provides food for thought. Happy to answer questions. Hopefully I can push out the David's Day video out soon.

[Viridian]

And the David's Day film project is done!

Link my DA page here:
https://www.deviantart.com/viridianqs/a ... 1279502281

5-minute video, file size is way too large to upload here. I think this is a great sample of what you can do with AI now. I honestly think this is a faithful recreation of your story. David's Day and Gail's Gallop still remain my favourites.