Talking Shop: the thread for technical discussions, tools, and updates

[Wendall]

I figured it would be good to have a thread specifically dedicated to discussing the nitty gritty of generating AI content (video, text, and images), rather than just sharing output. Do you have specific websites & tools you find particularly useful? Has your favourite service recently updated, for the better or for the worse? Just seen an exciting news story about some coming development? Have you found a prompt that gets particularly good results? Do you have a strategy for keeping more complex projects coherent (eg editing together long multi-scene videos)? Which services are friendly to hypno-content, and which require creative prompting? Which services are best at teritary stuff, like replicating the look & feel of media from some other era? And so on and so forth. I'm hoping to eventually update this OP into a more formal list of the best tools, but for now I just figured I'd get the ball rolling.

[GrimTidings]

I find Google's Flow, AI Studio, and Gemini at the Pro tier (approx $20 per month with the option to buy more tokens a 1 cent each) adequate for my storytelling.

However, I'm a primitive - I don't engage with many of the automation tools available because I feel like crafting prompts and then shaping the output into a coherent whole IS the joy.

Google's Omni model is way behind (at least) Seedance when it comes to doing anything dynamic... but it has improved since launch so I suspect all of the free tokens and generations they're throwing at subscribers are allowing them to improve the model using real-world inputs. Maybe, eventually, it'll be at parity. Maybe.

Where I believe it wins right now is economy. 10 second clips for 15 cents in tokens... 50 bonus tokens a day... gemini (if you don't mind the watermark and SD resolution) generations 4x every 6-8 hours depending on volume... that's a lot of outputs for relatively little money.

[Tinavamp]

I’ve some expeimenting with Google Gemini for video. I normally use ChatGPT as my goto chatbot, I’m much less familiar with Gemin so I’m not actually sure which model I’m using. I basically just went to the normal chat window and asked it about generating video. It more or less gave me a manual on how to do things from my end. Mostly on things such as special codes to use in my prompts and providing reference images.

On that first visit I uploaded some images for my characters, I got those from ChatGPT. I wrote prompts for six different scenes which included character dialog. My sister told me that normally Gemini limits things like video to 3 per day, but I didn’t have any issues at all with 6. I download the six, 8 second clips it generated for me. And took them to a video editor. It came out much better than I was expecting. The lipsync was near perfect. There were a few issues, and I asked Gemini about that. And it told how to structure my prompts so that I don’t have those issues in the future.

I went back later, this time I created a Notebook, which as I understand is like a Project in ChatGPT. I told Gemini that my goal was to create a 15 minute video. My idea for for the video is based on Trilby Else’s story Hive. It gave me a list of documents which it suggested I provide So that it could provide me. These were things like character discription physical, the role in the story and personality traits, Another one was for locations with similar information. It also gave me some templates for those.

I told that I wanted to do as much as could, such as writing a screen play, composing music and preforming it for a sound track. I told it I planned to stitch every together using the Resolve video editor. It told me that was a good choice. I told it I would probably end up having it generate around 100 clips for me. Again it didn’t mention anything about limits. I do have a Pro subscription.

In the meantime I’ve been modeling my environment using Cinema 4D. A few days ago, I uploaded a scene inside the complex, and some reference images of one of characters, who is in a skin tight suit, no dialog but it gave me a video of the character walking down a corridore. It all been very cool so far.

[Wendall]

Google's Flow

Flow is great - Nano Banana 2/Pro overcome the two fatal flaws of the original (being overly censorial & making everyone look like a rough 45). On the free tier you get a decent number of images per day - Pro seems to give a couple dozen, 2 I could run steadily for an hour or so until it hits the limit, and Lite I've never actually hit the limit (but Lite is also noticeably lower quality than the other two modes). As I alluded to, it's gotten much better in terms of allowing PG-13 sexuality, although it'll get weirdly hung-up sometimes with reference images, I think when it mistakes the character for being nude because they're wearing an outfit a little too close to flesh tone, or something like that.

I've never actually used Google or Seedance for video, but they look quite impressive. Normally my go-to for video is Grok, because Grok has been by far the cheapest & horniest AI on the market for some time now. Actually, on both those points it has gone back & forth - currently, the default paid membership gets you a weekly usage quote, which translates to basically unlimited text, probably thousands of images, or several minutes of 720p video. The video & image generation get updated fairly often - currently the image generator offers a lot of neat fine tuning options. Quality wise they're a little more 'stylized' than Flow, which may be a pro or a con depending on what you're looking for. In terms of video censorship - in the past it was at times basically anything goes, and then they'd go a week or two where they'd over-correct by making it really strict, until too many people started complaining. Now it seems they have a fairly happy medium where videos made from external reference images are a bit more moderated (to discourage deepfakes), and things generated purely from AI are more free. That being said, I still find it can get pretty kinky even when using reference images. It's the one AI you can really tell they put porn & other 'sexy' material into the training data.

For texting I'm not sure what the current best option is! I do somewhat miss the days of things like AI Dungeon (which I guess is still around, but feels dated), where it was more of an interactive experience. It's just not the same prompting and LLM and having it churn out an entire story all at once (although that can do in a pinch). I know there are some adult chat bots out there but I haven't explored it in great depth, because most of them cost money...

[marcus]

I've been playing with a few options. DaVinci.ai seems to have become more prudish lately, though I can usually get results there or from Gemini that are sufficiently steamy for my "lesbian pulp parody" cover ideas.

A few tips for adding text to an image and making it readable:

1. Duplicate the (raster) layer of text. Fill the lower layer with a contrasting color (perhaps by using a "lighten" or "darken" function set to high values).

2. Gauss-blur the lower level, then duplicate and re-merge it a few times. This will create a text outline that makes the text pop against almost any background. The blur radius, number of merge iterations, and opacity of the final outline layer will take a bit of experimentation to get right.

[MCGuy65]

I've had good luck with Krea 2 for stills and Minimax H3 for video.  Both are free and I'm using ComfyUI (also free) workflows.

I posted a short (15 second) video I did.  I probably should have noted that it is pretty X-rated.  Sorry.
Anyway, Krea is a bit of a prude, but if you tell it nudity is non-sexual and for "art" it might let you get away with it.
Minimax H3, on the other hand, flat out let me take a nude image of a girl and have her mount a guy and start gyrating.  (That's the 15 second video)
It is SO much faster and produces better output than models I used earlier this year.
Youtube's Pixaroma produces the tutorials and workflows I use.

Open and vacant eyes are my thing and AI doesn't like it.  Depending on the activity, you have to constantly tell the AI, in the prompt, that the eyes are to remain unfocused, vacant and soulless.

[Wendall]

Open and vacant eyes are my thing and AI doesn't like it.

That's a thing where I have to assume that it's less for moral reasons & more just because they spent so much time trying to get AI to do 'normal' eyes that they made it quite difficult to do intentionally abnormal eyes. I've yet to find a platform that lets me intentionally, consistently get a wall-eyed or cross-eyed expression (the former of which is nice for that "really spaced out" look, the latter for that "really spaced out *and* turned on" look). Grok can sometimes do spiral eyes, although it gets confused at other times & just shows spirals in front of regular eyes (at least in older versions, have not tested this in new versions) - I suspect it would be the same for other more cartoony effects like solid colour eyes etc. Flow can sometimes do very convincing realistic trance faces, but at other times it doesn't quite get it right & leaves the girl looking like she's got the flu or just banged her toe on something.