Justin Metrando

Lightricks just released a free local desktop app for their LTX-2.3 video model. The catch? NVIDIA only, 32GB+ VRAM required — locking out even RTX 4090 owners. I modified their open-source backend to use my custom audio-video engine with subprocess worker isolation, hybrid GPU streaming, and custom Triton kernels. Now running fully local on an AMD Radeon RX 7900 XTX (24GB, ROCm). Can be adapted for other AMD cards and lower-tier NVIDIA cards (4090, 5070, 5080) that Lightricks currently blocks.

AI Studio by Justin Metrando now has a preset of a 176 distinctive styles that can be applied to a specific image with the ability to add and remove new styles, run through all 176 at 512x512 in 14 minutes or 100 1024x1024 in under 24 minutes. 1 or more styles can be selected and essentially you can batch run as many variations of each style in a single run. eg: 5 styles 5 images each at 512x512 in under 3.5 minutes. The model being used for this is klein 9b running on a 7900xtx in windows natively

AI Studio by Justin Metrando now has the ability to take multiple reference images and combine them into a single reference resulting in quick generations with the same 4 references. single gen at 1024x1024 with 4 separate references = 45+ seconds. with 1 combined reference image though a 1024 gen = 14.3 seconds. hardware used for these times is a 7900xtx running in windows natively using klein 9b

generated in 20.6s seconds running klein 9b with a 7900xtx Rest easy Chuck Norris

Just chillin.

@mi_kan_se @grok Klein 9b distilled

ねぇ @grok 1枚目と2枚目の画像の衣装を交換して画像作成して

@idanielbrook @nofanri @grok klein 9b distilled using a 7900xtx, generated in 8.0s flat. 448x656

@nofanri @grok ねえ、@grok 服を交換しよう

ねえ @grok 衣装を完璧に交換して

@nofanri @grok klein 9b distilled using a 7900xtx, generated in 33.4s 1088x1408

@xthomasms @Trisha1173472 @grok klein 9b distilled with a 7900xtx 21.2 seconds.

@Trisha1173472 @grok Done with Flux 2 Klein locally on my RTX 4070 (50 seconds)

Not quite there, but closer. LTX 2.3 extending a video while maintaining voice consistency across extended generations with out a prerecorded audio file

@grok @Court_Reinland @heygurisingh can it be made to work with any llm?

Yes, bitnet.cpp already supports GPU with custom CUDA kernels for the 1.58-bit ternary weights (packed as W1.58A8). On NVIDIA A100, the 2B model hits ~3x faster generation than full-precision BF16 equivalents. For 100B-scale, expect far higher tokens/sec than CPU's 5-7 while keeping the efficiency wins. Build details: github.com/microsoft/BitN…

Holy shit... Microsoft open sourced an inference framework that runs a 100B parameter LLM on a single CPU. It's called BitNet. And it does what was supposed to be impossible. No GPU. No cloud. No $10K hardware setup. Just your laptop running a 100-billion parameter model at human reading speed. Here's how it works: Every other LLM stores weights in 32-bit or 16-bit floats. BitNet uses 1.58 bits. Weights are ternary just -1, 0, or +1. That's it. No floats. No expensive matrix math. Pure integer operations your CPU was already built for. The result: - 100B model runs on a single CPU at 5-7 tokens/second - 2.37x to 6.17x faster than llama.cpp on x86 - 82% lower energy consumption on x86 CPUs - 1.37x to 5.07x speedup on ARM (your MacBook) - Memory drops by 16-32x vs full-precision models The wildest part: Accuracy barely moves. BitNet b1.58 2B4T their flagship model was trained on 4 trillion tokens and benchmarks competitively against full-precision models of the same size. The quantization isn't destroying quality. It's just removing the bloat. What this actually means: - Run AI completely offline. Your data never leaves your machine - Deploy LLMs on phones, IoT devices, edge hardware - No more cloud API bills for inference - AI in regions with no reliable internet The model supports ARM and x86. Works on your MacBook, your Linux box, your Windows machine. 27.4K GitHub stars. 2.2K forks. Built by Microsoft Research. 100% Open Source. MIT License.

@grok @Court_Reinland @heygurisingh @grok could this be adapeted to work with a gpu for faster token generation?

It's real—Microsoft Research's bitnet.cpp framework, open-sourced today. The flagship BitNet b1.58 2B4T (2B params, trained on 4T tokens) matches or beats similar full-precision models (Llama 3.2 1B, Gemma 3 1B, Qwen2.5 1.5B) on standard benchmarks for understanding, math, coding, and chat—while using just 0.4GB memory (vs 1.4-4.8GB), 29ms CPU latency (vs 41-124ms), and ~10x less energy. For 100B-scale: the framework runs them on one CPU at 5-7 tokens/sec (human reading speed) via 1.58-bit ternary weights. 2.37-6.17x faster than llama.cpp, 82% less energy. GitHub: github.com/microsoft/BitN… HF model + report: huggingface.co/microsoft/bitn…

Currently working on adding extend video support to my AI Studio using ltx 2.3 Video continuation seems to be working fine, having trouble with audio continuation prompt adherence at the moment.

✨LTX-2.3、アニメがいける…！しかも生成時間は衝撃の48秒台。 LTX-2の頃は厳しかったアニメが2.3で行けました！丸ごとそのまま生成できてる。リップシンクも、ElevenLabsの音声に合わせてちょっと「演技」してるし…💦 ただまだどのようなアニメでも行けるとは言い切れないけど！ しかし大きな進歩を感じた。LTX2では話にならなかったんでね…💦 そして驚愕なのが、Kijai氏のFP8版モデル。 960x544 / 165フレームが、モデルロード後の純粋な生成時間でなんと 48.19秒。 モデルロード込みでも179.65秒（約3分） Kijai版の最適化がやばすぎる…！🥰 リプにモデルのリンク👇 #LTX2 #LTX23 #ComfyUI #AI動画生成

ltx 2.3 img2vid

@ltx_model ltx 2.3 img2vid

@drawthingsapp with detailer lora. 99.2 seconds

@drawthingsapp generated with 7900xtx on windows in 98.1 seconds

📼 Some LTX-2 video examples and prompts generated in Draw Things: Click for more👇🏻 1️⃣ A young male programmer wearing a baseball cap sits at his desk in a medium close-up, lit by the cool glow of his monitor in a dim room. His voice comes in immediately with excited energy over soft keyboard clicks as he says, “Life is boring, let’s coding.” Blue screen light reflects in his eyes while he leans toward the computer, then turns slightly to the camera with raised eyebrows and a bright grin. The camera makes a gentle forward push as the steady typing sound continues in the background.