Pinned Tweet
The core idea started from a time when @GlitterPixely helped me try creating some Mecha designs with MJ.
In the end we ended up with a stone Golem guardian for the witch.
After that I thought "what if the witch was a character in Glitter Pixely's typical cute style"... and here she is.
I had planned to create this video for a long time.
But Seedance was too expensive for my current situation, and the cheaper AIs weren't strong enough in the action parts, so I temporarily put it aside.
And finally thanks to @itsPolloAI letting me into their CPP program, I was able to complete it.
Thank you. #PolloAI
🤝 Paid partnership · 🤖 Made with AI
The final images and video turned out reasonably decent, considering I only reran the 3D environment generation three times.
Besides that, the environment ended up looking slightly different from the reference image. And generating the video took over 2 hours, not even counting the time spent creating the 3D environment.
Maybe this weekend, if Claude still has some credits left, I should give it a try and compare the results. But with ChatGPT Terra on the Plus plan, everything is just... exhausting.🫠
That's exactly why I say these stupid AI haters have no brains.
Even if they're too lazy to use their own brains, they could just click the "Grok" button in the corner and have it explain everything to them, instead of making complete idiots of themselves in public, right!?
Do they hate AI so much that they can't even bring themselves to click the "Grok" button and would rather expose their own stupidity?
To make things clear, I've been regularly sharing my prompts and how I create things with AI for a long time.
And obviously, if I share something publicly, it's because I want people to use it.
If I shared something but didn't want anyone to touch it, I'd be just as stupid as those AI haters.
I have absolutely no problem with anyone using my prompts. But the question is, how do they use them?
If someone uses my prompts to create something new and incorporate it into their own work, I welcome that, whether they tag me to give credit afterward or not. That doesn't really matter to me.
But if someone just takes my entire prompt unchanged and throws it into an AI, that's a completely different story.
It's like when someone leaves a bowl of candy outside their house on Halloween.
Even if there's no sign telling kids how much they can take, a well-behaved kid will obviously only take their fair share.
But those who just grab the entire bowl and stuff it into their bags? Get lost.
Replying to @CuriousOne_01
Your issue is that you think using genAI is totally cool and based but draw the line at prompts being copied because you're a smarmy internet hoplite who has no consistent ideals or values and just wants to feel like the special bestest boi.
This world is so adorable!!! 😍😍😍
Inspired by @awwbhi2 and @zzzzshawn I made some small levels featuring effects (tweaked) from @patch_gg
Of course it's running in @threejs because it's all 3D!
Should I turn this into a whole game?
Let's give Terra another try and see if it can actually produce something decent this time...
🤖 Made with AI
I used open-source code from @AmirMushich to create a game with a 3D pixel art style.
And open-source code from @KanaWorks_AI for the game's sound effects and music.
And the result is this chill little fishing game, made with Opus 5.5.
You can try it out just by clicking the link:
nxtr1609.github.io/lakeside/
I've included links to both creators' open-source GitHub repositories in the replies below. If you're interested, be sure to check out their GitHub pages! 👇
Curious 1 retweeted
noteのメンバーシップ記事で、この部屋のモデリング作業記録前半を公開しました。まずは5回分のイテレーション記録。
AIエージェントによる3Dモデリング実験 - 001
note.com/kogu_dev/n/n8c61ea2…
次は後半を記事にします。それから動画と、作業記録からTerra用の指示の最適化についても。
Curious 1 retweeted
#ElevenLabsAmbassador
ずっと見ていられる、ドット絵の神社。
真っ白な世界にちょこんと浮かぶ、小さな箱庭です。影はほんのり青紫、日なたはあたたかな黄色。桜が揺れる境内で、うさぎたちが蹴鞠をしています。
ボタンひとつで遊べる『ドット庭』、開発中です⛩️
声・効果音・BGMはすべて @ElevenLabs で生成。
@ElevenCreative
#ElevenCreativeAmbassador
🤝 Paid partnership · 🤖 Made with AI
When I share my prompts publicly, I hope people will use AI to analyze and learn from them, to figure out how I set up my camera angles, effects, and so on. But I really dislike and strongly disapprove of people simply taking the original creator's entire prompt and throwing it into AI like this.
I think creators who share their prompts should add a short line instructing the AI to display their @username in the center of the video, starting at the 5-second mark and remaining there until the end, like a watermark.
That way, if these lazy copycats just copy and paste your prompts to generate videos, they'll end up with a video featuring your username in huge letters right in the middle.
Free advertising for you!
Curious 1 retweeted
🎃ハロウィンナイト、ありがとうございました!
24時間で328回の夜を、みなさんと過ごしました🐰
賞金キャンペーンの結果を発表します🏆✨
🏆最高スコア賞(5,000円)
エリちゃんだよ @naoyuki_okada さん
🍬3,852(ツバキでプレイ)
✨最も選ばれしHERO賞(5,000円)
ルル(88回プレイ)
クリエイター Curious 1 さん @CuriousOne_01
おめでとうございます!🎉
夜明けまで生きのびた割合 52%(スコア記録のある194回中101回)
ゾンビになって誰かの夜に現れた 155体
みんなで集めたキャンディ 298,400個以上
SPECIAL HEROを生んでくれたクリエイターのみなさん、遊んでくれたみなさん、本当にありがとう!
これからもよろしくおねがいしますね🐰
I was blown away by what another creator had managed to make using Terra. And when they shared how they did it, they explained that they had Terra repeatedly refine and improve the same project through 10 iterations to achieve that level of quality.
So I wanted to give it a try myself. I gave Terra a screenshot from a game to analyze and asked it to create a 3D model based on it.
First iteration: about 7 minutes.
Second iteration: about 7 minutes.
Third iteration: about 20 minutes.
Fourth iteration... OUT OF CREDITS (5h cooldown). LOLLLL
And if you're a ChatGPT Plus user who's ever run out of credits in the middle of a task, you already know what happens next.
Unlike Claude Pro, where you can simply ask it to continue once your credits reset, I have no idea why ChatGPT Plus starts rereading everything and doing the whole task all over again after hitting the credit limit. This makes it almost impossible to finish what you were working on.
And yep, after my credits reset, I asked it to continue, and I've now been waiting for over 2 hours. It's STILL running. LOLLLL
ChatGPT might be good, but only if you're on the $100 plan or higher. The $20 plan is practically useless. If your budget is only $20, you're much better off using Claude Pro. It's 10 times better than ChatGPT Pluss, and the workflow is so much smoother.
This is the post by the creator who ran Terra through 10 iterations to create their work.
Seriously, if nobody told me it was AI generated and I just took a quick glance at it, I'd think it was a real room filmed with a handheld phone.
If you're on a higher-tier ChatGPT plan, you might want to try their method with Terra. 👍✨
But if you're on a lower-tier ChatGPT plan like me... please don't. 😅💀
nitter.gq/koguGameDev/status/211…
この動画にたくさんの反応をもらい、疑う意見もあったため、まとめて答えます。
これは今AIエージェントで実現可能なことの基本的な実験でしかなく、最高品質でもなければ、当然フェイクでもありません。
私の過去の投稿を見ればAI使ったフェイクを忌み嫌ってるのはわかると思います。この動画も部屋のシーンのモデリングも、一切フェイクはありません。結果に対する私の感想も「予想通りTerraはなかなか良い」であって、最高にリアルだとか革命だとか思っていません。
部屋のシーンは次の条件で作成しました。
・デスクトップ版のChatGPTで、CodexからGPT 6 .1 Terraを使う
・人間はBlenderを一切操作しない。見もしない
・簡単な指示から開始して、10回の改善イテレーションを行う
・画像も3Dモデルも、リファレンスとなるデータは渡さない
動画の方は、部屋のモデリングが10回の改善を終えた時点から、カメラワークとレンダリング、ポストエフェクトについての指示を与えて5回ほどリテイクしたものです。リテイクのほとんどはカメラワークで、しつこく丹念に指示しました。
—-
繰り返しますが、全ての工程で私自身は一度もBlenderを起動していません。目標を与え、結果への指摘を行い、改善の繰り返しを重ねただけです。
そもそも10回目でも動画でも、個々のオブジェクトのモデリング品質は酷いものです。拡大すればすぐわかりますが、プリミティブそのままや実物に程遠い物だらけ。服なんかクローゼットの中も窓際も破綻しています。
それでも全体の画像や動画がそれなりのリアリティを感じさせるのは、オブジェクトの選択と物量、そして総合的な印象を構成する配置とモチーフの選択、何よりBlenderの性能によるものです。それらはいずれも指示次第で大きく結果が変わる部分です。
Terraだけなんて信じられないと疑う声もありましたが、実際にTerraのみを使ってこの結果が得られています。そして私にはその点の驚きはありません。
すでに9月、Astraで以下の実験を行なってます。
nitter.gq/kogugamedev/status/209…
さらにそれ以前から、ビジョン有するモデルを使ったエージェントの3Dモデリング性能はテストしてますし、AIの 3Dに関するテストは延々続けています。
今回の動画もそうしたテストの一環でありおまけです。Terraはコーディングタスクに向かない一方、Computer useなど視覚を求められるタスクには強いというベンチマークを見て、これならAstraに匹敵する3Dモデリングも可能かもしれないと考えました。
あくまで実験であり、本来の目的は10回のイテレーション中に取らせた作業記録です。それを元に、どのような指示が作業効率を上げ、どのような操作はコストに見合わないのか、分析して次の実験を行います。
フェイクは無いし、最高品質でもない。全てを事細かに指示もしていないし、全部丸投げでもない。これはそうした実験の一つに過ぎません。


























