- This topic has 21 replies, 5 voices, and was last updated 1 day, 16 hours ago by
.
-
Posts
-
-
Spent a chunk of my Sunday trying out video generation with audio. Qwen is too restricted. Grok is a little better but hard to control. Pollo has its issues but it’s easiest to control and has few restrictions. I’ve put together a 43 second video (from 4-5 second clips) with speech all the way through. It’s clunky (editing-wise)but I think it’s the best demonstration I’ve seen yet of what will be possible for us very soon.
The big question is, doesn’t anyone here want to see (and hear!) it?
-
Yes 🙂
-
Sorry for the delay. Had a bit of a messed up evening and didn’t get anything done.
-
As I’ve (temporarily) given up on Pollo (trying to rebuild some credits), I had a mess around again with Grok imagine and Qwen. They’re both pretty well locked down when it comes to nudity, so I went down a different route (but still on topic):
-
And being the Luddite that I am, that’s all AI? Bloody hell! Amazing.
-
All completely AI creations. Very detailed prompts, cherry-picking the results (I picked the best 20 videos out of 56 created), and an element of luck combined to generate these.
Unlike the previous couple of videos I worked on, these clips were created by Grok and Qwen, mostly via my mobile in spare moments. As I’ve spent a lot of time in hospital waiting rooms lately, it seemed like a good use of time when I don’t have my laptop available.
-
Tell me about it! If I added up all the endless hours I’ve spent waiting in hospital corridors and hospital waiting rooms since 2009, it’d run into weeks.
-
I have to feel like I’m achieving something otherwise I’d go mad. After spending 10 days in hospital in November, I’ve been a lot less able to let time just slip by.
-
-
-
The image quality is also excellent and the lip sync, though not 100% perfect, is as good as most videos on the web.
-
Speech synchronisation is still a little wonky, and it varies from video to video. Some of it might be manually correctable with intensive video editing but I’m not going there.
Another advantage of these short clips is that the lack of consistency with voices isn’t a problem. If I create two video clips with the same characters in them, the AIs cannot be relied upon to be consistent with the voices produced.
Grok seems much more reliable on lip-sync than Qwen, but models are constantly being upgraded so that can change quickly.
-
-
I’m trying to get a straight answer on whether Grok Spicy does or doesn’t allow NSFW content. Even their website says both.
-
And I certainly won’t be getting a subscription if the UK government is threatening to block X from the country.
-
I’m beginning to feel pressured to get a VPN.
-
-
-
Please post a comment if you’d like to see more of these.
-
I enjoy them. But would like to see them take their clothes off.
-
I’ll need to pay £100-200 per month to do these videos with proper uncensored nudity. We need a lot more people to become NE patrons to make that viable.
I’ve got an NSFW one on the way but any nudity is fleeting or minor and snuck through the moderation. If I directly prompt nudity, it gets reliably rejected.
-
-
Sometimes it feels like a video generator is trying to be infuriating. I tried a clip of a London cabbie telling a nudity-related joke:
- The first go was fine except the joke was too long and the last words were cut off. I reworded it and tried again.
- Go 2 was fine except, for some weird reason the London cab was bouncing all the way through as if someone heavy was repeatedly getting into the back. Added an instruction to keep the cab still.
- Go 3 – saw the back door slide open and then partially shut. That made no sense so I added an instruction to keep the doors shut. It was distracting and the cabbie’s performance wasn’t great.
- Go 4 – the driver’s door opened at the start and the cabbie was (unlike the 1st 4) facing the wrong way. Stipulated which way the cabbie should look and tried again.
- Go 5 – the cab moved perfectly, no doors opened, the cabbie looked at the camera and spoke total gibberish. No idea what he said.
- Go 6 – the cab rocked a little too much but not as badly as 2. Also, the driver’s door opened slightly, but the cabbie’s vocal performance was perfect. I try really hard not to spend too many credits on a single clip so I’m calling it a day there.
I feel that video could easily have been perfect on attempt 2. I’m sure a lot of people think you just tell it what you want and you get a perfect result first time. As I’m trying for more precise output, it’s taking a lot more work (and credits) to get there.
-
I’ve tried using Pollo to make video and to be honest, it’s pretty rubbish. I only have a free account and log in frequently to pick up free credits, but I’m getting to the stage of giving up now.
-
It all depends on what AI model you are using on Pollo. They have a huge range of them. Most are content-restricted and expensive but some of the cheaper ones are pretty good. The complication is that things keep varying so what works one month may not give the same results or offer the same options a month later. Wan 2.5 on Pollo used to work very well, but I haven’t used it in a few weeks. Pollo 2.5 was total cr*p last time I used it – bizarre distortion, random nonsense and completely inconsistent results. I’m hoping they’ve improved that.
I like the models which allow you to set a start frame and an end frame – that usually gives a LOT more control. You often need to turn the inference up too, and most of the models don’t give you control over that. I’ve found wording is SO important and you often need to specify when something is not doing anything or a person is not speaking just as much as the things or people who are active. Qwen isn’t bad and Grok is very good, but the nudity moderation is pretty strict on both of them (and about to get stricter, I imagine).
Of course, the addition of audio has complicated the entire field, but I’m not giving up yet. I swear all of them have good days and bad days too. Sometimes I can’t get a result on Qwen, but do on Grok or vice versa. Sometimes I end up saving a prompt and using it a day or two later when the AI is in a better mood.
I didn’t once have such a bad day getting videos out of Qwen that I used the prompt “A woman smiling at the camera” with no start frame image and the result got moderated.
-
It’s so annoying when you get the resulting clip almost perfect on the first or second go but then you’re still trying to get rid of messed-up speech or weird animation (stuff I can’t edit away) after another 6 attempts.
-
-
- You must be logged in to reply to this topic.