設営完了しました！
DT16でお待ちしてます！
DT16でお待ちしてます！
July 12, 2025 at 3:29 AM
設営完了しました！
DT16でお待ちしてます！
DT16でお待ちしてます！
当日だーーーーー！イェーーー
こちらのギラギラテナさんポスターが目印です✨✨
スペースDT16でお待ちしてます👿👼
こちらのギラギラテナさんポスターが目印です✨✨
スペースDT16でお待ちしてます👿👼
July 11, 2025 at 11:56 PM
当日だーーーーー！イェーーー
こちらのギラギラテナさんポスターが目印です✨✨
スペースDT16でお待ちしてます👿👼
こちらのギラギラテナさんポスターが目印です✨✨
スペースDT16でお待ちしてます👿👼
I tested DT16 with many IQA metrics and found out DT16 has severe noisy edges and blocking even on the high-quality mode. There's no metric that can detect both; the DT16's architecture is too slow to continue. So I have to restart the project from scratch.
September 28, 2026 at 6:08 AM
I tested DT16 with many IQA metrics and found out DT16 has severe noisy edges and blocking even on the high-quality mode. There's no metric that can detect both; the DT16's architecture is too slow to continue. So I have to restart the project from scratch.
DTの新しいアルバム制作が始まりましたね。マイアンがご機嫌そうで何よりです！
www.facebook.com/100044162677...
www.facebook.com/100044162677...
James Labrie
DT16 is officially underway!
#dt16 #dreamtheater #newalbum
www.facebook.com
February 10, 2024 at 12:43 AM
DTの新しいアルバム制作が始まりましたね。マイアンがご機嫌そうで何よりです！
www.facebook.com/100044162677...
www.facebook.com/100044162677...
7月ムプチ場所出ておりました！
DT16です！
アジクロとクロアジの境界線だよ
進捗ヤベェですが持ってけるよう頑張ります！
🇬🇧🏴グオメロケ地巡り、テナベス鑑賞、ロンドンコミコンシーンさん謁見盛りだくさんの一冊になりそうです〜
DT16です！
アジクロとクロアジの境界線だよ
進捗ヤベェですが持ってけるよう頑張ります！
🇬🇧🏴グオメロケ地巡り、テナベス鑑賞、ロンドンコミコンシーンさん謁見盛りだくさんの一冊になりそうです〜
June 6, 2025 at 11:27 AM
7月ムプチ場所出ておりました！
DT16です！
アジクロとクロアジの境界線だよ
進捗ヤベェですが持ってけるよう頑張ります！
🇬🇧🏴グオメロケ地巡り、テナベス鑑賞、ロンドンコミコンシーンさん謁見盛りだくさんの一冊になりそうです〜
DT16です！
アジクロとクロアジの境界線だよ
進捗ヤベェですが持ってけるよう頑張ります！
🇬🇧🏴グオメロケ地巡り、テナベス鑑賞、ロンドンコミコンシーンさん謁見盛りだくさんの一冊になりそうです〜
These Madden ratings/rankings all feel pretty right 😒
Cam Ward- 72 (QB30)
Kobie Turner- 85 (DT16)
Edgerrin Cooper- 79 (LB32)
Puka Nacua- 88 (WR13)
Brian Thomas Jr- 84 (WR28)
Moro Ojomo- 68 (DTIstoppedcounting)
Travon Walker- 78 (Edge47)
Garrett Williams- 80 (CB40)
Trey Hendrickson- 92 (Edge7)
Cam Ward- 72 (QB30)
Kobie Turner- 85 (DT16)
Edgerrin Cooper- 79 (LB32)
Puka Nacua- 88 (WR13)
Brian Thomas Jr- 84 (WR28)
Moro Ojomo- 68 (DTIstoppedcounting)
Travon Walker- 78 (Edge47)
Garrett Williams- 80 (CB40)
Trey Hendrickson- 92 (Edge7)
August 8, 2025 at 1:52 PM
These Madden ratings/rankings all feel pretty right 😒
Cam Ward- 72 (QB30)
Kobie Turner- 85 (DT16)
Edgerrin Cooper- 79 (LB32)
Puka Nacua- 88 (WR13)
Brian Thomas Jr- 84 (WR28)
Moro Ojomo- 68 (DTIstoppedcounting)
Travon Walker- 78 (Edge47)
Garrett Williams- 80 (CB40)
Trey Hendrickson- 92 (Edge7)
Cam Ward- 72 (QB30)
Kobie Turner- 85 (DT16)
Edgerrin Cooper- 79 (LB32)
Puka Nacua- 88 (WR13)
Brian Thomas Jr- 84 (WR28)
Moro Ojomo- 68 (DTIstoppedcounting)
Travon Walker- 78 (Edge47)
Garrett Williams- 80 (CB40)
Trey Hendrickson- 92 (Edge7)
I think the lossy photos have fully solved, unless you care about that single pixel of sand. Lossy screenshots are still hard for DT16 to compress. So that's what I want to start with.
September 28, 2026 at 6:22 AM
I think the lossy photos have fully solved, unless you care about that single pixel of sand. Lossy screenshots are still hard for DT16 to compress. So that's what I want to start with.
The DT16's encoder is heavier than x265, but the decoder is 2-3x heavier than JPG, still in the same class. But its real competitor are JXL and AVIF.
September 26, 2026 at 11:55 PM
The DT16's encoder is heavier than x265, but the decoder is 2-3x heavier than JPG, still in the same class. But its real competitor are JXL and AVIF.
My agent told me the DT16 architecture has reached its limit. But I have to get the image smaller by 1%. I can spend more time figuring out how to improve the ratio; it's getting very hard now.
September 27, 2026 at 11:58 AM
My agent told me the DT16 architecture has reached its limit. But I have to get the image smaller by 1%. I can spend more time figuring out how to improve the ratio; it's getting very hard now.
This puts me in a hard place where I've got the stuff out or improved the byte ratio. I've to explain why I fell asleep while working on DT16. I need a cuddle from my imaginary girlfriend to ease my worries.
My agent told me the DT16 architecture has reached its limit. But I have to get the image smaller by 1%. I can spend more time figuring out how to improve the ratio; it's getting very hard now.
September 27, 2026 at 1:17 PM
This puts me in a hard place where I've got the stuff out or improved the byte ratio. I've to explain why I fell asleep while working on DT16. I need a cuddle from my imaginary girlfriend to ease my worries.
This drum set looks like if a regular kit started mutating like Tetsuo at the end of Akira.
February 7, 2024 at 8:48 PM
This drum set looks like if a regular kit started mutating like Tetsuo at the end of Akira.
DT16 is encoding kodim01 at 0.007 MP/s with strength reduction on Mac Mini M4, it was 0.0011 MP/s on multi-thread only. Now it takes 57 seconds instead of 6 minutes to encode kodim01. it's fast compared to PAQ8 0.004 MP/s, but PAQ8 was on single-thread. But DT16 decoding speed is 15 MP/s now.
September 25, 2026 at 4:54 PM
DT16 is encoding kodim01 at 0.007 MP/s with strength reduction on Mac Mini M4, it was 0.0011 MP/s on multi-thread only. Now it takes 57 seconds instead of 6 minutes to encode kodim01. it's fast compared to PAQ8 0.004 MP/s, but PAQ8 was on single-thread. But DT16 decoding speed is 15 MP/s now.
DT16 should be game-changing for games because it cuts down the image size by 15% better compared to JXL in some images. And games usually compress their images in high-quality mode, and only one encode. DT16 does it better when the image is noisy rather than flat gradient.
September 24, 2026 at 7:05 PM
DT16 should be game-changing for games because it cuts down the image size by 15% better compared to JXL in some images. And games usually compress their images in high-quality mode, and only one encode. DT16 does it better when the image is noisy rather than flat gradient.
I admit that the DT16 Rust codebase wasn't optimized at all. Zero SIMD and parallelism on the decoder, and the encoder does use parallelism. The decoder is 1.1k LOC, and JXL was 12k LOC without modular mode. It was hard to improve DT16’s compress ratio because the encoder is very slow.
September 24, 2026 at 7:43 PM
I admit that the DT16 Rust codebase wasn't optimized at all. Zero SIMD and parallelism on the decoder, and the encoder does use parallelism. The decoder is 1.1k LOC, and JXL was 12k LOC without modular mode. It was hard to improve DT16’s compress ratio because the encoder is very slow.
ヒットルアー
・阿修羅EXDR
・ハイカットDR
・エリー95MD
・マッドペッパーマグナム
・ラパラDT16
・スメルトヘッド+スターリングシャッド
・ドロップショットミノー
・HP3Dワッキー
・ドライブスティックspec2
・阿修羅EXDR
・ハイカットDR
・エリー95MD
・マッドペッパーマグナム
・ラパラDT16
・スメルトヘッド+スターリングシャッド
・ドロップショットミノー
・HP3Dワッキー
・ドライブスティックspec2
December 7, 2024 at 5:04 AM
ヒットルアー
・阿修羅EXDR
・ハイカットDR
・エリー95MD
・マッドペッパーマグナム
・ラパラDT16
・スメルトヘッド+スターリングシャッド
・ドロップショットミノー
・HP3Dワッキー
・ドライブスティックspec2
・阿修羅EXDR
・ハイカットDR
・エリー95MD
・マッドペッパーマグナム
・ラパラDT16
・スメルトヘッド+スターリングシャッド
・ドロップショットミノー
・HP3Dワッキー
・ドライブスティックspec2
The DCT of DT16 was using O(n³) f64 matrix multiply, and the range coder was calculating 128-bit division per decoded symbol and binary search for cumfreq. The color-conversion loop is in scalar f64. So the improvement headroom for DT16 is a lot. The code was written in the worst way possible.
September 24, 2026 at 8:11 PM
The DCT of DT16 was using O(n³) f64 matrix multiply, and the range coder was calculating 128-bit division per decoded symbol and binary search for cumfreq. The color-conversion loop is in scalar f64. So the improvement headroom for DT16 is a lot. The code was written in the worst way possible.
2014年8月。淡路の麺や六三六の特製ラーメン。ここはだいたい iPhone で撮った写真が多いけれど、この写真は α77 に DT16-50mm F2.8 で撮ったもの。
May 25, 2024 at 4:17 AM
2014年8月。淡路の麺や六三六の特製ラーメン。ここはだいたい iPhone で撮った写真が多いけれど、この写真は α77 に DT16-50mm F2.8 で撮ったもの。
91% of the bytes DT16 spends on an image is the luminance channel; the chroma only takes up 9%. I'm saying wow, this is the toughest lossy compression that I have to optimize. Because others have to use external data like a LSTM weight to compress enwik8, it's the same for image but with diffusers.
September 26, 2026 at 10:57 AM
91% of the bytes DT16 spends on an image is the luminance channel; the chroma only takes up 9%. I'm saying wow, this is the toughest lossy compression that I have to optimize. Because others have to use external data like a LSTM weight to compress enwik8, it's the same for image but with diffusers.
There was a bug where my LLM lied in the context summary about a new DT16 color space. There weren’t any tool calls creating the color space. It convinced itself the code was lost and spent 1 hour searching for that folder. I woke up and got messages of missing files from the agent. I'm so terrified
September 26, 2026 at 10:03 PM
There was a bug where my LLM lied in the context summary about a new DT16 color space. There weren’t any tool calls creating the color space. It convinced itself the code was lost and spent 1 hour searching for that folder. I woke up and got messages of missing files from the agent. I'm so terrified
I have to weaken the MDSI HaarPSI metric to just MDSI because it takes 22.7% of time. And switch out zopfli (14.3%) for something else. zzg11 range coder encode (13.5%). They are the ones that make DT16's speed in the same category as PAQ8.
September 25, 2026 at 5:32 PM
I have to weaken the MDSI HaarPSI metric to just MDSI because it takes 22.7% of time. And switch out zopfli (14.3%) for something else. zzg11 range coder encode (13.5%). They are the ones that make DT16's speed in the same category as PAQ8.
The DT16 was written in Rust and called from Python, which is less atrocious to call the function, and easy to speed it up in Rust. I can ask an LLM to speed the Rust code up and go back to sleep for two hours instead of sitting there. Or in the past, writing the Rust code by hand, which is hard.
I'm seeing a weird design choice in those image formats. Like the PNG using 5 predictors instead of avg2. And AV1 using DST and rotation. Non-square sizes DCT blocks on JXL.
Thank you so much for LLMs. I developed an image codec called DT16. It was able to beat JXL 6/11 of the time on "web" quality mode, and 10/11 "high" mode. It only loses to JXL. It wasn't able to figure out how to compress low-frequency details efficiently. I wasn't able to beat JXL in lossless mode.
September 26, 2026 at 9:36 AM
The DT16 was written in Rust and called from Python, which is less atrocious to call the function, and easy to speed it up in Rust. I can ask an LLM to speed the Rust code up and go back to sleep for two hours instead of sitting there. Or in the past, writing the Rust code by hand, which is hard.
DT16 decodes kodim01 at 6.4 MP/s on single-thread; simulated 10 threads put this at 15-20 MP/s on Mac Mini M4. Compared to 15 MP/s of JXL single-thread. Running JXL on 10 threads doesn't improve the decoding speed.
September 24, 2026 at 7:39 PM
DT16 decodes kodim01 at 6.4 MP/s on single-thread; simulated 10 threads put this at 15-20 MP/s on Mac Mini M4. Compared to 15 MP/s of JXL single-thread. Running JXL on 10 threads doesn't improve the decoding speed.
When the LLM was developing DT16, there were a lot of rage-inducing and punishing issues. It decided to init a git repo and leave it uncommitted. Or random tool calls that took 4 hours to finish. If I pressed abort out of exasperation. Thanks, I'm back to the beginning waiting for the script again.
September 25, 2026 at 6:52 PM
When the LLM was developing DT16, there were a lot of rage-inducing and punishing issues. It decided to init a git repo and leave it uncommitted. Or random tool calls that took 4 hours to finish. If I pressed abort out of exasperation. Thanks, I'm back to the beginning waiting for the script again.
It was extraordinarily difficult in lossless mode because the transformed residual is nothing but noise before feeding it into a range coder. But at the lossy part, JXL is wildly inefficient by speed reasons. DT16 takes 6 minutes to compress kodim01 on a Mac Mini M4, which is a powerful computer.
September 24, 2026 at 6:47 PM
It was extraordinarily difficult in lossless mode because the transformed residual is nothing but noise before feeding it into a range coder. But at the lossy part, JXL is wildly inefficient by speed reasons. DT16 takes 6 minutes to compress kodim01 on a Mac Mini M4, which is a powerful computer.
Thank you so much for LLMs. I developed an image codec called DT16. It was able to beat JXL 6/11 of the time on "web" quality mode, and 10/11 "high" mode. It only loses to JXL. It wasn't able to figure out how to compress low-frequency details efficiently. I wasn't able to beat JXL in lossless mode.
September 24, 2026 at 6:21 PM
Thank you so much for LLMs. I developed an image codec called DT16. It was able to beat JXL 6/11 of the time on "web" quality mode, and 10/11 "high" mode. It only loses to JXL. It wasn't able to figure out how to compress low-frequency details efficiently. I wasn't able to beat JXL in lossless mode.