Stable diffusion thread

[zheng]

This is good. Izit forever free?

not free i guess need subscription fee

 

[kgluong]

https://www.pcgamer.com/games/card-games/champions-tcg-ai-artist/

Card game developer says it paid an 'AI artist' $90,000 to generate card art because 'no one comes close to the quality he delivers'​

holy f

Well it looks like prompt command artist get paid for their prompt skills instead of their art skills.

 

[zheng]

Well it looks like prompt command artist get paid for their prompt skills instead of their art skills.

that one guy is an experienced artist himself. ai just did most his work. He still manually do digital touchup and edit.

 

[doogyhatts]

This is good. Izit forever free?

SA2 has a personal license which gets you 20 free music tracks each month.

However, you can use Suno.ai for other genres of music (eg mandopop, k-pop, j-pop, english, instrumental, etc).
50 credits daily will get you 10 free songs each day.

 

[doogyhatts]

After further testing, I found out that the Multi-Diffusion upscaler in ComfyUI can help to improve the quality of the animation frames.
This is done before face fixing.

An update to the flickering issue after further testing.

Previously, I combined face-fixing using ADetailer with image upscaling using Ultimate SD-Upscale in A1111/Forge.
That is incorrect as it produces flickering results on the character's forehead and can introduce image errors across the animation frames.

The solution is to split up the process into two parts. Upscale the frames first before doing face fixing on them.

The ComfyUI version of Ultimate SD upscale does not appear to work on a list of animation frames, as it ended up in an infinite loop.
So this part I used Forge instead.

The incorrect version:

Spoiler

The correct version:

Spoiler

 

[doogyhatts]

Raw audio2video output from AniPortrait HF demo.
Not bad, but need to do some resampling and fix the teeth.

 

[zheng]

Raw audio2video output from AniPortrait HF demo.
Not bad, but need to do some resampling and fix the teeth.

too creepy already

 

[PikaPika33]

next time no more arts liao

 

[doogyhatts]

Raw audio2video output from AniPortrait HF demo.
Not bad, but need to do some resampling and fix the teeth.

Ok so I attempted to fix the teeth using ADetailer.
This also results in severe flickering and changed the lips, which is not suitable for adding back the audio as the lip-sync is lost.

Next, I used FlowFrames to interpolate for additional frames and slow down the animation.
This will make it more useful for ASMR type of videos.

 

[doogyhatts]

This PAG node can be used to generate correct looking images before SD3 arrives.

 

[doogyhatts]

SD3 will be available soon in the membership plan.

 

[doogyhatts]

 

[zheng]

anywhere can test already?

 

[doogyhatts]

anywhere can test already?

SAI uses a pay for credits system. Not sure how it links with Colab.
https://platform.stability.ai/pricing

Perhaps, wait a few more weeks to get the SD3 weights model.

 

[doogyhatts]

 

[zheng]

https://stable-diffusion-art.com/stable-diffusion-3-api/


Hands are still problematic, unfortunately.

 

[doogyhatts]

I haven't tried PixArt yet. It uses T5 encodings instead of Clip.

 

[AZE]

https://stable-diffusion-art.com/stable-diffusion-3-api/


Hands are still problematic, unfortunately.

Hands/palms/fingers are almost always problematic because of the nature of our physical biomechanics.
1st there ish left/right hand, then there ish the palm side and fist side, then there ish frontal view and back view relative to the body, then there ish also a 1st/2nd/3rd/4th person's perspective. Trying to compress 3D space accurately into 2D space ish difficult.
Not only do that complicate the training, they also complicate the tagging and inference prompting.
Natural language prompts are also problematic since different ppl uses different ways and means of expression, and hab different capabilities of expression and there ish significant redundancy and irregularity in the use of languages.
Maybe if someone designs another layer of AI that estimates or guess or predicts or directs the person's expression capabilities or style through some sort of prompt Q&A tests, and remaps that customisation into LLM layer, things might be better, but then that complicates inputs.

 

[doogyhatts]

https://stable-diffusion-art.com/stable-diffusion-3-api/
Hands are still problematic, unfortunately.

Lykon says the API service is using an older model, while he is using the up-to-date one.

 

[zheng]

generated from llama3 on whatsapp