LazyKernel

@komena1725 The sweater is coming off 👁️👁️

久々の通常コスチューム

@SlipperyGem If the dataset is ready (from previous trainings on flux, qwen, anima etc) and you have moderate size server behind you, it is possible. Also heard of new style lora training method that apparently super fast.

An unfathomably large and comprehensive set of style loras for Krea 2. This is the type of stuff that I could get lost in... HOWEVER, this may be slop, and, indeed how can over 1.5k loras dumped in one go be good? Reddit link for context. huggingface.co/ilkerzgi/fal-K…

@gamewikionline Are they adding sakiri date? No? I am disappointed still

皆さん、今日の生放送で 何が一番嬉しかったですか？🤔 海外掲示板を見ていると、 アプデ内容そのものよりも Sランクキャラチケット配布 が一番嬉しいという声がかなり多いようです👀 「ずっと欲しかったキャラをようやく取れる」 「誰を交換するか悩む」 「3パッチでSキャラ3体はかなり太っ腹」 「ダフォディールが入手できる」 「凸るべきか未入手キャラを取るべきか」 など、かなり盛り上がっている模様。 皆さんは今回の配布で、 誰を交換する予定ですか？ #NTE

@DailyShinkuNTE Thats normal for hotta. After all it is a matter of time until all S rank weapons become trash absolute piece of shit.

THEY CASUALLY GIVING FREE S RANK IN ABSOLUTELY RANDOM PATCH WHAT THE HELLL??????????????????

@marlene4719 Why would that matter to him? He literally owns the world.

I’m beginning to believe he really doesn’t get it that he’s a laughingstock worldwide.

雏鸟组

効果のほどは不明wだけど、Krea 2で使える無検閲Text Encoder！ huggingface.co/ahmed22xa/Huih…

@MuhdA34015 @NBinted And it works every time

@NBinted washed content creator resort to extremism when their niche fandom doesn't like them anymore

This guy used to collab with Ironmouse and MoistCr1TiKaL

The game sexualizes the loli characters. You're not the target audience.

@dgle3370 😭😭😭

#NTE #NevernessToEverness Birthday🎂

表情が良き 千夏 / Sunna #zzzero #ゼンゼロ

@Ben_L_Tracey @PrasantGopinat @SqSehrish I actually went backward, noticed the difference between 39 and 19 is 20, and the difference between 19 and 9 is 10, and 9 and 4 is 5. So the difference is doubled every time so the difference between 39 and the next number should be 40 if we follow the same rule.

@PrasantGopinat @SqSehrish Interesting. I got the same number in a different way. 79. Starting with 5, doubling, then adding to last number. 4+5=9 9+10=19 19+20=39 39+40=79.

Only 1% of people can solve this sequence correctly Can you? 💀

@SqSehrish 79?

@ailker @fal @krea_ai What are you training with to get this speed??? Even when I tried the 6000 on runpod it was not nearly as fast.

🎨 I trained 999 Style LoRAs with the @fal Krea 2 Trainer. Every single image in this video is a different LoRA. That's 999 unique Style LoRAs, each trained for just 100 steps at 0.00035 LR, taking an average of ~30 seconds to train. As I said before, @krea_ai 2 continues to prove that it's an incredible model for open-source Style LoRAs. Next week I'll be sharing even more LoRAs, along with a few surprises.

@Imamatdin_S @ApoStructura Thats regulation problem not AI. It is crazy to me how the government will allow water bills to go up because the local region are selling it to a company.

@ApoStructura i dont think these comparisons prove anything cz the water being used is getting drained locally and people seeing the result are locals who get affected. not international charts. It doesn't matter it uses 0.5% globally if it takes 60% of local availability

Putting AI water use in context

待って今日レクイエムちゃん誕生日じゃん！？！？！？悠長におはよーとか言ってる場合じゃない！みんな祝え！祝うんだ！！！！！🍅🍅🍅🍅 #NTE #NTEネバエバ #レクイエム

@023PerfectDark @NexusSites @AyakaMods What is the mod?

1. BTW @NexusSites deleted my NTE mods. So I can't upload my mods. The files needed to prevent you from being banned in-game are on a site they don't like. I said "look up @AyakaMods urself" and linked Google for a reason.

@LockedInLucas Why is the video like 6 frames per second?

People aren’t ready for the GTA 6 Those who are saying GTA 6 looks like GTA 5 Look again. Ongoing road construction

@Kuberuntung Ok, addict

@kohya_tech I had to use block swap at 16 to not run out of memory. Will try the other ones when I get to it thanks. But it is mostly because I have a 180 dataset and alot of repeat. Will probably need to tune it down.

@LazyKernel2025 もしGPUメモリが不足しているようなら、--fp8_base --fp8_scaledと、--blocks_to_swap、--block_swap_h2d_only、--use_pinned_memory_for_block_swapなどを組み合わせてみてください。

Krea 2のLoRA学習結果。わりと短時間で学習できるみたい（このLoRAだとRTX 6000 Pro Blackwell Max-Qで15分くらい）。 1枚目LoRAなし、2枚目あり、3枚目プロンプト応答をみたやつ（眼鏡外す）、4枚目画風を変えてみたやつ。