linker @keylinker

Very exciting work with my amazing collaborators @Hope7Happiness, @Lyy_iiis, Kangyang Zhou, Linrui Ma, and Kaiming He! All code, models, and full training recipe are open-sourced. Blog Post: peppaking8.github.io/#/post/minit2i Code: github.com/PeppaKing8/min… (JAX), github.com/Hope7Happiness… (PyTorch). 🤗 Models: huggingface.co/MiniT2I

. github.com/Pierrot-vision… 문의/질문등은 언제든 환영합니다. 많은 관심 부탁드립니다. 감사합니다.

아직 미완성이지만, 먼저 피에로 코드와 베타 버전 0.8b 학습 모델을 먼저 공개합니다. 계속 업데이트할 계획이며, 완성하는대로 다양한 실험 모델들을 계속 공개할 예정입니다. 자세한 내용(소개/취지/코드/알고리즘등)들은 링크한 레퍼지토리의 README에 상세하게 내용을 써놓았습니다

저 혼자 솔로로 진행하고 있는 이미지 생성 모델 프로젝트 '피에로' 를 소개하고자 합니다. 홀로 개인 비용과 스크래치부터 학습한 모델입니다. 최소한 비용을 학습하기 위해 다양한 알고리즘들을 마구잡이(?)로 적용했습니다.

[주둥이코딩-DL] 논문 리뷰 (계속 업데이트 될 예정) github.com/Pierrot-vision…

Zero training needed, Image-to-LoRA(i2L) V2 is now live 🚀! As a much improved version, the updated i2L can turn one or more reference images into style LoRA in one single forward pass, oblivating the need for explicit training to customize image styles. 🤖 Model Collection: modelscope.ai/collections/Di… 🎠 Studio: modelscope.ai/studios/DiffSy… 📄 Paper: modelscope.ai/papers/2606.13… 🌟 V2 Upgrades: - Compatibility: applicable to various base models including Z-Image, Klein-4B and Hidream-O1. - Speed: Direct style LoRA weights prediction = Zero training. - Style fidelity: High fidelity style transfer, no semantic leakge. - Control: explicit weights unlock asymmetric CFG, multi-reference style fusion, and controllable-generation modules. - Architecture: image encoder + learnable LoRA queries + compressed decoding heads that generate the adapted matrices . It's open. It's out. Go build something cool. 🛠️✨

keyboard_arrow_left Previous keyboard_arrow_right Next

The Z-Image team publiced their latest paper: High-Fidelity Two-Step Image Generation via Teacher-Aligned End-to-End Distillation Z-image turbo++ 2steps distillation model.

keyboard_arrow_left Previous keyboard_arrow_right Next

DiP: new pixel-space model, no VAE needed. - Crisp textures without compression artifacts. - 10x faster than prior pixel models - 1.79 FID on ImageNet huggingface.co/zhen-nan/DiP

Andrej Karpathy가 2시간 동안 자기가 실제로 AI를 쓰는 화면을 그대로 보여줬습니다 (OpenAI 공동 창립자, Tesla AI책임자였던) 핵심은 놀라울 정도로 단순합니다 말로 일을 맡기고, 결과를 보고, 다시 한 문장으로 고칩니다 AI 시대의 차이는 프롬프트 암기가 아니라 “일을 맡기는 감각”에서 갈립니다 핵심만 한국어 자막으로 1시간 정도 압축했습니다 AI를 제대로 쓰고 싶은 사람은 꼭 보세요

Chrome @0xchromium

2 weeks ago

Andrej Karpathy spent 2h showing how he actually uses AI day to day he's a co-founder of OpenAI and led AI at Tesla, so when he shows how he works, it’s worth watching and the whole session is just him telling the machine what he wants in simple terms, like he's briefing a

130 1K 11K 1.8M 30K

Alibaba released Qwen-Image-Flash Few-step distillation goes beyond objectives. Data composition, teacher guidance, and task mixture unlock fast, high-quality text-to-image and instruction-guided editing.

Bytedance drops an interesting paper, Representation Forcing. Understanding and generation meet in a single representation space, learned end to end, with no VAE in between. Looks like VAE will not be necessary for future models. yuqingwang1029.github.io/Representation…

#CVPR2026 paper: It's Never Too Late: Noise Optimization for Collapse Recovery in Trained Diffusion Models Text-to-image models often collapse to near-identical samples. Our fix: optimize the noise. Start from pink 🩷, not white noise. 🔗akoepke.github.io/divgen/index.h… 1/6

We are thrilled to introduce our new work, Representation Forcing for Bottleneck-Free Unified Multimodal Models! What lies between language and pixels? Representation. We propose to ground visual representation in the decoder of a unified model. The decoder learns to predict it from text, just as the encoder reads it from images. Understanding and generation then meet in a single representation space, learned end to end, with no frozen VAE in between. We call this Representation Forcing: the encoder's representations force the decoder to learn the same visual structure, and the decoder's predictions, in turn, force the pixels to follow it. Paper: arxiv.org/abs/2605.31604 Project page: yuqingwang1029.github.io/Representation… Huggingface: huggingface.co/papers/2605.31… Feel free to chat with @yuqingwang1029 at #CVPR2026

keyboard_arrow_left Previous keyboard_arrow_right Next

What if you could speed up AI image generation by 22x without retraining from scratch? Researchers from Zhejiang University and the University of Adelaide introduce FlashAR. They add a lightweight vertical prediction head to existing autoregressive models, enabling parallel two-way next-token prediction. This preserves the original training objective while dynamically combining horizontal and vertical predictions. Result: up to 22.9x speedup for 512x512 image generation, using just 0.05% of the original training data. FlashAR: Efficient Post-Training Acceleration for Autoregressive Image Generation Project: lxazjk.github.io/FlashAR/ Paper: arxiv.org/abs/2605.09430 Code: github.com/lxazjk/Emu3.5-… Our report: mp.weixin.qq.com/s/k6wry8wpyzOe… 📬 #PapersAccepted by Jiqizhixin

昨日に引き続きブログ公開 Day 2です。 KV cacheの基礎について、「Key, ValueだけcacheしてQueryをcacheしないのはなぜか？」という問いに答える形式でやさしく解説を行いました。 MLエンジニアのための本質から理解するLLM推論 KV cache編｜Kazuki Fujii zenn.dev/kaz20/articles…

keyboard_arrow_left Previous keyboard_arrow_right Next

Kazuki Fujii @kazukifujii

3 weeks ago

普段利用しているNVIDIA NGCをGPU Clusterで利用するときの環境構築方法についてTips記事を執筆しました。 Tips: Containerを利用したDL分散学習Libraryの開発環境｜Kazuki Fujii zenn.dev/kaz20/articles…

keyboard_arrow_left Previous keyboard_arrow_right Next

0 10 60 44K 37

やべぇ、画像生成のデータセットが出た huggingface.co/datasets/jaspe…

JLT: Clean-Latent Prediction in Latent Diffusion Transformers. This is an initial version; we will continue to update it: Paper: arxiv.org/abs/2605.27102 Page: akatsuki-neo.github.io/JLT Code: github.com/akatsuki-neo/J… Model: huggingface.co/dawn-neo/JLT

The latent-vs-pixel debate misses the point. GPT Image 2 shows what users notice: pixel-level fidelity. Latent models show what scales: compact semantic structure. We connect them by replacing VAE/RAE decoders with a Pixel Diffusion Decoder. Code and Model available: research.nvidia.com/labs/sil/proje… 🧵(1/N)

PiD + Flux Klein 9B BASE this is 205x zoomed in. the amount of detail in the skin on a BASE gguf model with NO LoRAs is actually unbelieveable. Flux is great but does struggle with skin textures. pixel space is where it belongs. civitai.com/models/2651727…

keyboard_arrow_left Previous keyboard_arrow_right Next