VEO 3 Quicksand Tests

[Viridian]

With VEO 3 being launched and marketed as one of the best AI video tools out there, I gave it a shot. The main selling point is its crisp quality, prompt adherence and - most of all - its automatically generated sound AND speech. This is a composite of test videos I made using the mud/QS theme.

The first video is ENTIRELY generated on VEO 3 (no source image). The others used image to video. I generated the prompts using ChatGPT following prompt guidelines, including specifying dialogue.

Pros:
- Outstanding prompt adherence
- Dialogue is almost perfect word-for-word
- Full emotional expression
- Very good automatic lip sync
- Excellent sound effects and timing

Cons:
- Extremely expensive for an 8s video
- Very sensitive to NSFW - virtually no cleavage/boobs allowed
- Virtually no mud interaction - has the water problem
- Generally poor dynamic motion - good set piece scenes, but poor movement - Kling 2.1 absolutely destroys this

The biggest hit to VEO 3 is its cost. It's main interface is Google Flow, which is accessed through Google's Ultra AI subscription, which is behind a $200/3 month paywall, which has limited credits.

I actually accessed it by signing up to OpenArt AI, which has combined image and video gen models, including Hailuo, Wan, Veo and Kling. At current sale price - $56/month, or $28/month if you paid annual plan upfront - you get 24000 credits, with options to purchase extra credit packs per month. It's a lower entry point than Google Ultra, but generations are still high cost.

A Veo 3 generation of 8s with full sound and voice is 2500 credits. Assuming the higher monthly cost, that's basically $6 for an 8s video. That is likely not usable. Without sound, it is 2000 credits, or about $4.60.

In contrast, Kling 2.1 Master is 1000 credits for 10s ($2.34), and Kling 2.1 Pro 10s is 600 credits ($1.40).

As I paid for an annual OpenArt subscription, which is 50% off for each month, these values are halved, which is substantially better value, and currently Kling gens have a special discount. I'd be paying about 0.70 for a 10s Kling 2.1 Pro generation on OpenArt, and about $0.50 on KlingAI on the Premier Annual plan - though with the regular flash sales you can recharge fairly efficiently.

My verdict? VEO3 is nice, but it gets a Fail grade for QS. It's also too pricey for casual hobby use.

[Duncan Edwards]

Thank you for the review. I very nearly did the exact same thing with OpenArt AI a few days ago and decided to wait after playing with it a bit. Right now it does look like KlingAI is the place to be and I've got a Premier subscription there. I'll keep tinkering with that until the next best thing happens. It won't be long.

Addendum of sorts - The biggest show stopper for me was the extreme adherence to almost preschool "G" rated limitations. KlingAI can be coaxed into a little cheescake.

[Viridian]

VERY huge update dropped yesterday.

VEO 3 is still in Beta and had numerous problems including - very annoyingly - inserting gibberish subtitles, and generally costing a kidney to generate. Last month they released the VEO 3 Fast model, which is a turbo-quality output - 1/10th the cost and much faster, but with lower quality. Not really a problem if you're making primarily static scenes, though it only ran Text to Video. It's good for one-off clips but useless for sequences using the same setting and character.

Yesterday, just as I was beginning my workflow for a new project, VEO 3 update. It largely fixed the issues with not generating sound and making subtitles (though it's not perfect).

Most importantly, it now allows Image to Video. This is the biggest game changer in AI animation. I've been able to integrate VEO 3 Fast into my workflow. The single best selling point is its audio control, including near perfect lip sync and expression. It cuts out a big part of my workflow, as I don't need to record and sync voice lines, and creates a superior output.

There are still a few downsides:
- Extremely sensitive censor - won't do anything that hints at sexual content (e.g. cleavage)
- Tends to trip the "prominent person" filter too much - this is meant to deter deepfakes of celebrities, but often a random face resembles a celebrity in the system.
- Heavily coated mud is also blocked - I'm guessing that it is blocking out fetish content
- Accents are kind of random. Since you can't assign a specific voice to a character, you kind of have to accept what it gives you
- Mediocre dynamic movement: Kling 2.1 still destroys VEO for interaction. I prefer to use Kling when I need to do sinking sequences or less sensitive NSFW visuals

I find there are workarounds if you have the eye for editing. Things like zooming in for facial closeups to use VEO 3's superior dialogue outputs and using the whole frame for Kling animation actually suits the flow of a TV-style interaction.

In short, VEO 3 has catapulted to being my preferred model. Originally I used Kling 2.1 as the main source and VEO 3 to make the filler scenes with lively audio, but since Image > Video is a thing, I've flipped the roles: VEO 3 as my main model for voice lines and environment, Kling 2.1 for dynamic movement.

Here is a taste of the future.

[Duncan Edwards]

Thanks for the preview. This is really getting to be amazing stuff. What this really says to me is the next gen of Ve0X or Kling 3 or whatever is going to be the one that has it all together. My biggest complaint with many of them, and as you have discovered with Ve0, the absolutely frantic efforts to avoid anything "dirty". I will always be more open to anything that doesn't have a seizure at the very suggestion of breasts.

[MDLambert]

I never had much luck using Veo 3

[Viridian]

I never had much luck using Veo 3

I've been dumping a lot of time (and no small amount of funds) into working with VEO 3. It's been instrumental as part of my storytelling process. However, it's pretty hopeless at making anything actually move, so sinking is out of the question. In fact, VEO is so opposed to the peril that it _continually_ pushes partially submerged characters back up when you generate using source images.