Stable diffusion thread

[doogyhatts]

side track abit... is there a free good text to speech program or app

ElevenLabs has free monthly credits.

For Japanese anime, use VoiceVox.

 

[jadedRomeo]

Only tried the Dev version on their HF demo. Didn't try the Pro version.
Prompt adherence is good for people characters but couldn't get a cinematic artistic style.
Lack of diversity and doesn't work well with monsters.

The good anatomy from Flux won't matter because YT now has been sufficiently trained to disengage recommendation of AI-girls.
So all the generated images from Flux that have scantily clad, bikini, lingerie, sexy poses, cannot be monetised on YT.
Even fully clothed young female characters or young children can be implicated as well.

Old woman and old man still works though.

I think I'm not that well versed in SD yet, but you're spot on.
I feel Flux has more "freedom" than Midjourney, while still retaining its ease of use.

 

[jadedRomeo]

 

[jadedRomeo]

 

[BlackWing1977]

 

[x1243x]

anyone tried running nf4 flux on comfy? not bad. much faster

 

[doogyhatts]

I now use two Kling accounts. Who else here is doing the same?

 

[jadedRomeo]

I now use two Kling accounts. Who else here is doing the same?

Why would u need 2 accounts?

 

[doogyhatts]

Why would u need 2 accounts?

It is one china account + one international account.
I calculated the total amount of points and realised that I need two accounts in order to make daily videos.
Each long-form video has a duration of about 1 min 20 secs, so I need about sixteen 5-sec clips each day.

The goal is to follow the strategy used by the Zythi Studio channel.
That channel is doing multiple videos each day while I can only do about two videos every three days.

 

[doogyhatts]

Looks like KlingAI is back with correct facial animation. No more distortions unlike yesterday.

 

[doogyhatts]

anyone tried running nf4 flux on comfy? not bad. much faster

I managed to get the schnell-nf4 version working and added some inpainting nodes to it.
Still very slow and couldn't get any suitable results.

 

[x1243x]

I managed to get the schnell-nf4 version working and added some inpainting nodes to it.
Still very slow and couldn't get any suitable results.

hmm.. I was using it on comfyui but since it's available on forge now, went back to forge.. it's not bad on forge.. 80s on average

 

[doogyhatts]

hmm.. I was using it on comfyui but since it's available on forge now, went back to forge.. it's not bad on forge.. 80s on average

For some reason, I couldn't reinstall forge on Stability Matrix. Kept getting errors.
Looks like I will have to use ComfyUI to do the inpainting instead.

 

[x1243x]

For some reason, I couldn't reinstall forge on Stability Matrix. Kept getting errors.
Looks like I will have to use ComfyUI to do the inpainting instead.

hmm ok.. I don't like to do inpainting on comfyui..

 

[doogyhatts]

Dreamina is ByteDance's answer to KuaiShou's Kolors & Kling.

I am unsure about the quality differences at the moment.
Dreamina might be closer to Pika, Gen2, Haiper.

Some observations for video model:
1) It costs 30 credits to generate one video clip of 3 seconds.
Daily credits is 150. So five videos each day for default mode (non-relaxed).
Cheapest tier gives 1260 credits per month.
Video clip is generated at 8 frames per second, so have to use FlowFrames to do interpolation.

2) Relaxed generations is only available for premium users, but I am not sure if it is only restricted to the most expensive tier.
Enabling this mode allows a cost reduction of 80%, but incur a delay of 8-24 hours.

3) Without relaxed generations, Dreamina can generate a total of 192 clips each month, with each video clip at 3-seconds.
When comparing with cheapest tier for Kling, it can generate a total of 264 clips each month, with each video clip at 5-seconds.

4) If Dreamina's cheapest tier can also do relaxed generations, then it can generate 960 clips each month.
Extending each clip to 6-seconds, would probably yield only 480 clips each month, assuming the cost is the same.

I will stick with Kling.

 

[doogyhatts]

Hotshot is another video generation site.
The quality looks similar to Kling.
The free tier is only 2 generations per day, and max 60 generations each month.
And it also allows for commercial usage, even though you didn't pay anything for using the free tier.

 

[doogyhatts]

Just upgraded the text encoder to GGUF for PixArt-Sigma.
For the Q8 model, it takes only 8 seconds to load, as compared to the bf16 one which takes 5 minutes.
https://github.com/city96/ComfyUI-GGUF

 

[doogyhatts]

Midjourney web platform is now open.
25 free generations before end of this week. Each generation is 4 images.

 

[x1243x]

anyone using forge now? getting weird results today. was generating fine. suddenly image is garbage

 

[doogyhatts]

Level4 has released his Flux.1D derivative model, FL4X on civitai.
It is 20gb in size.