ZODA

Show, Don’t Tell. 🧠 Can a text-to-image model reason during generation without turning every intermediate thought into text? This is the question behind LatentMorph, an ICML 2026 work our team contributed to on implicit latent reasoning for image generation. Most reasoning-augmented T2I methods rely on explicit thoughts: intermediate reasoning is decoded into text, then re-encoded back into the generation process. This creates a bottleneck: • visual information can be lost • decode–encode loops add latency • fixed-step reasoning does not match the dynamics of visual creation LatentMorph keeps reasoning in continuous latent space instead. ✨ It monitors the autoregressive generation stream, invokes reasoning only when needed, and injects latent guidance back into later image token prediction. Core components: 🧩 Condenser compresses intermediate states into compact visual memory 🔁 Translator maps latent thoughts into generation-compatible guidance 🎛️ Shaper injects control signals into the generator’s KV cache ⚡ RL-trained Invoker adaptively decides when reasoning should be triggered On Janus-Pro, LatentMorph reports: 📈 +16.0% on GenEval 📈 +25.3% on T2I-CompBench 📈 +15.6% / +11.3% over explicit reason-while-generation on WISE and IPV-Txt ⏱️ 44.3% less inference time 🪙 51.0% lower token consumption 🧠 71.8% alignment with human reasoning-invocation preferences The takeaway: for visual generation, reasoning does not always need to be verbalized. LatentMorph shows that latent thoughts can support more adaptive, efficient, and cognitively aligned text-to-image generation. arXiv: 2602.02227 #ICML2026 #GenerativeAI #ComputerVision #TextToImage #MultimodalAI #AIResearch

⚙️ The policy ScalingAR turns confidence into inference-time control: • Adaptive Termination Gate prunes low-confidence trajectories early • Guidance Scheduler adjusts guidance strength to balance semantic alignment, visual quality, and diversity The broader shift: visual generation is moving from output-level quality to process-level control.

ScalingAR has been accepted to ICML 2026 🎉 Our team contributed to this work on test-time scaling for NTP-based autoregressive image generation. TTS has become a powerful inference-time strategy for LLMs. But token-by-token image generation is a different problem. Partial text can still be meaningful. A truncated image token stream is often just an unusable artifact. ScalingAR introduces the first TTS framework specifically designed for NTP-based AR image generation. The key idea: use token entropy as an intrinsic confidence signal — without frequent partial decoding or external reward models. Core components 🖼️ • Dual-channel confidence profiling • Early termination of low-confidence trajectories • Dynamic guidance scheduling across generation stages Results 📊 • +12.5% on GenEval • +15.2% on TIIF-Bench • 62.0% less visual token consumption • 26.0% mitigation of performance drop on challenging prompts arXiv: 2509.26376 A few notes below 🧵 #ICML2026 #GenerativeAI #ComputerVision #AutoregressiveModels

🧠 The signal ScalingAR builds a Dual-Channel Confidence Profile: • Intrinsic Channel: token-level confidence + local spatial stability • Conditional Channel: text-conditioning utilization One checks whether generation is stable. The other checks whether the model is still using the prompt.

🔍 The gap Existing visual TTS methods often rely on partial decoding and external reward models. For NTP-based AR image generation, this is a poor fit: intermediate visual outputs are incomplete and unstable, while image quality depends on full-sequence coherence.

Register: luma.com/1yrpn8vn

ICML week in Seoul should not be only about papers. It should also be about meeting the people building what comes next. We’re hosting ZODA Happy Hour × D-Robotics Innovators Night @ ICML ✨ A high-signal evening with researchers, builders, investors, and open-source AI contributors. Highlights: 🔬 Researcher talks on Quantum AI, World Models, and Embodied Intelligence 🎙️ Cross-border panel on open-source AI ecosystems and commercialization 🌃 High-rise views over Seoul’s night skyline 🍸 Signature cocktails + curated gourmet bites 🎁 Lucky draw: D-Robotics RDK S100 Dev Kit + MaaS Compute Credits worth ¥30,000 Co-hosted by ZODA × MolarData × D-Robotics Supported by ZhenFund × ComNergy Tech No rigid seating. Just fluid ideas and real conversations. 🗓️ July 9, 2026 ⏰ 7:00 PM – 9:00 PM 📍 Seoul Register: luma.com/1yrpn8vn See you at ICML. ✨ #icml #ICML #ICML2026

AlignVid has been accepted to ICML 2026. 🎉 Our team contributed to this work on semantic fidelity in text-guided image-to-video generation. A common failure in I2V: the model preserves the reference image well, but misses the edit specified by the prompt. We identify this as visual dominance — high-fidelity image tokens dominate attention and underweight text-driven semantic edits, leading to semantic negligence. AlignVid is a training-free attention modulation method. No fine-tuning. No input degradation. No external masks. It uses: ⚙️ ASM: rescales Q/K representations to sharpen attention and reduce entropy, helping prompt-relevant semantic tokens compete with dense visual priors. 🔧 GS: applies modulation selectively across foreground-sensitive transformer blocks and early denoising steps, improving prompt adherence while limiting visual-quality degradation. We also introduce OmitI2V, a benchmark with 367 image-text pairs covering modification, addition, and deletion. On Wan2.1, AlignVid improves semantic alignment accuracy: • Modification: 72.35 → 77.20 • Addition: 71.75 → 79.54 • Deletion: 63.13 → 69.47 Runtime overhead stays negligible: • FramePack: +0.09% • FramePack F1: +0.03% • Wan2.1: +0.01% arXiv: 2512.01334 #ICML2026 #GenerativeAI #ComputerVision #ImageToVideo #VideoGeneration

Across these three works, we see a clear shift: visual generation is no longer only about output quality. It is also about process control: when to stop, when to reason, when to follow the prompt, and how to do it efficiently. More detailed breakdowns coming soon from ZODA ✨

🧠 Show, Don’t Tell Problem: Explicit reasoning in T2I often requires decoding intermediate thoughts into text, causing information loss, extra latency, and a mismatch with how humans create. Method: Continuous latent-space reasoning via four lightweight modules, plus an RL-trained invoker for adaptive reasoning timing. Results: +16.0% on GenEval, +25.3% on T2I-CompBench, 44.3% inference-time reduction, and 71.8% alignment with human intuition on reasoning invocation. arXiv: 2602.02227

Excited to share that our team contributed to 3 papers accepted to ICML 2026 🎉 🎬 AlignVid: Taming Visual Dominance in Text-guided Image-to-Video Generation via Training-Free Attention Modulation 🖼️ ScalingAR: Scaling Confidence for Autoregressive Image Generation 🧠 Show,Don’t Tell: Morphing Latent Reasoning into Image Generation Together, these works explore a shared direction: making visual generation models not only produce better outputs, but also better control the generation process. arXiv links + short breakdowns below ↓ #ICML2026 #ICML

@hongming731 It would be great if I could subscribe to specific categories—for instance, focusing only on trending AI topics, or narrowing it down even further to just open-source-related results.

🔥Open-source AI does not stop at open models. Models are getting stronger. But the real question is becoming harder: --What data was the model built on? --How was it evaluated? --Can the benchmark be trusted? --Who defines the standard? 🫂That is why we are building ZODA. ZODA is an international open-source community for AI data and model evaluation.🤜🤛 We start with four core initiatives: 🧬ZODA Benchmark 👥ZODA Leaderboard 🏡ZODA House 🫀ZODA Challenge Open-source AI needs trusted data, harder benchmarks, real evaluation, and a global builder community. Fueling AI, Together.🤟🤝 #ZODA #opensourcedata #opensourcecommunity

@TheTriceRozay If anyone likes and follows me, I’ll wish you all the best in life every single day from now on.😽