Transp. Chan

Dreaming about using AI to make anime since 2014. Author of Live3D/CoNR/AttNR.

[Transp. Chan @transpchan]

it's like using text to generate scene graphs, one frame per line, containing a bunch of bouding boxes and openpose (if needed)

[Transp. Chan @transpchan]

Is there any paper using llm to do video layout generation?

[niji・journey ✨ @nijijourney]

Are you ready? 🌈 #nijijourney

[Transp. Chan @transpchan]

Is there any paper that uses a token to represent position instead of positional embedding?

[Transp. Chan @transpchan]

Is there any paper doing image translation with pretrianed rectified flow models? Something like instructNoise2Noise connecting two domians, maybe?

[Yanhong Zeng @zengyh1900]

WOW!!! A GAME FROM NIJI 😍

[まっくす @minux302]

こんな感じでラフから線画化するAI作ってます。ご興味ある方いたらリプいただけたら変換します！（大ラフほど荒いものにはまだ対応できていないです。今鋭意学習中）

[Transp. Chan @transpchan]

@DavidSHolz @tekbog https://t.co/NoihyCCF5i

Maybe this would help...

[Transp. Chan @transpchan]

People are hyped up at openai #sora. Rumor has it that synthetic data is the key to its success.

Fun fact: ALL anime are synthesized data becase they never "really" existed. People have trained neural network using them for a decade.

[Transp. Chan @transpchan]

Then people write paper about its "emerging" halting problem, LOL

[Transp. Chan @transpchan]

A question that has puzzled me since the days of masked language models. GPTs that can detour insertions to the end of sequence, but is there any LM which predicts the insertion position of evey new token (or even deletion)? Maybe something like diffusion+autoregressive LM?

[Transp. Chan @transpchan]

@C_Sobilac @ddvd233 Join the discord and look at the chat history :)

[[email protected] @ddvd233]

二次元浓度过高（
#IJCAI

[Transp. Chan @transpchan]

@Yokohara_h I am (still) working on it. :)

[Transp. Chan @transpchan]

@yuliangxiu PhDs can reject paper without reading them. GPT has to read every paper word by word carefully

[Transp. Chan @transpchan]

@dragonmeteor I think the method is heavily based on eg3d from nvlabs. If you fit it with your own datasets, I think it will work similarly.

Also, I wonder how did they get the CoNR dataset. I didn't even have a chance to release the datasets before I was kicked out LoL.

[Transp. Chan @transpchan]

@EMostaque @nerdymomocat @OpenAI any chance you give me some GPUs to build a system that turns manga into anime?
https://t.co/eIsGNKxxGY

[Transp. Chan @transpchan]

A new Live3Dv2 demo turning AI generated character design sheets (4pics) into a continuous animation video.

This may also run in real-time is you have a powerful GPU (as opposed to stable diffusion based methods which requires multiple steps)

https://t.co/sgKpnng5GP

[Transp. Chan @transpchan]

@8co28 Wow! That's awesome!!

Can this technology also be used to fix other flickering stable diffusion videos by simply giving it character sheets instead of LoRA finetuning?
Also, can we use arbitrary MMD motion file and untextured models as inputs? This may save the texturing time.