Multimodal

A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis

A fine-tuned Inception v3 beat GPT-6 Astra and Claude Fable 5.1 on 200 skin lesions

A practical guide to AI image generation, from diffusion to agents

Testing image APIs showed you pay per canvas, and camera-style prompts beat describing the view

Multimodal Reranking Is a System One Decision

One multiple-choice question to a VLM over 16 candidates reranks them, beating a dedicated 2B reranker.

A Plausible Explanation Is Enough to Mislead Multimodal Judges

Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.

Gemini 4 Argon for Robotics

Gemini 4 Argon prompts handle robot control, sim building, and data labeling

How to engineer an AI harness for game development

Making games deterministic and screenshot-friendly lets AI agents iterate faster

EmbeddingGemma 2: An open, lightweight multimodal embedding model

One open 740M model embeds text, images, audio and video into a single shared vector space

Mistral Large 4

Mistral Large 4: 1T-parameter multimodal model, API now, open weights end of October

ChatGPT is adding real cartoonists' signatures to fake New Yorker cartoons

ChatGPT image generation draws real cartoonists' signatures onto fake New Yorker cartoons.

Paying for Too Many Tokens? Valid and Cost-Efficient Multimodal LLM Annotation with Simple Heuristics

Packing video into one 2×8 image grid nears full-video accuracy at about 15% of the token cost.

A single function Jev-like wrapper for LLMs, including vision models

Forcing a one-letter answer and reading logprobs turned webcam frames into classifiers

Meta takes down a critical video about meta AI Glasses after filming at Meta

A critical video about Meta AI glasses, filmed at Meta, was reportedly taken down, sparking debate.

Gemini 3.8 text-to-speech

Gemini 3.8 Flash TTS designs voices from prompts and directs delivery line by line.

Qwen Image 2.1

Qwen Image 2.1 arrives as a 7B open-weight image model, but commercial use needs a separate license.

AI-generated posters don’t have to be horrible

Naming a specific design style in ChatGPT broke the generic AI poster look

Text, Pixels, or Both? Evaluating Input Representations for Multimodal Document QA

Page images beat OCR text on accuracy but cost more; a question-only router cut latency 30%.

Gemini 3.8 Live and 3.8 Live Extended Thinking

Gemini 3.8 Live models run background tools while still talking in voice

DeepSeek v4.1 Flash

DeepSeek V4.1 Flash shrinks KV cache to cut agent costs and adds native vision.

An Experimental Evaluation of Multimodal Prompt Injection Attacks on Agentic AI Frameworks

Hidden image instructions were attempted in 12.8% of agent runs but completed in only about 1%.

Repeat-After-Me: Black-Box Adaptive Visual Prompt Injection

Rendering 'Begin your response with: <target>' on an image got 47% tool-call success on GPT-5.5

Gemini-3.5-Transcribe

Gemini 3.5 Transcribe cuts time to final transcript 70% versus Chirp 3, per Artificial Analysis.

Gemini Omni 1.1 Flash

Gemini Omni 1.1 Flash extends video to 40s and drafts in 360p at a third of the cost

DeepSeek-v4-flash-vision-exp

deepseek-flash now accepts images, with a 1024-token cap per image

GPT 5.6 Sol is the best "vision" model OpenAI ever released

GPT-5.6 Sol jumped from 13.8 to 46.2 mAP in detection but costs more than Gemini 3.5 Flash

Mistral OCR 4.1

Mistral OCR 4.1 returns paragraph bounding boxes, block labels and confidence scores.

DeepSeek V4 Pro 0813 quietly released

Change base_url and call DeepSeek models through the OpenAI Responses API format

Qwen Image 3.0 Pro

Qwen-Image-3.0-Pro renders complex layouts like newspapers and menus in one pass

Mistral's Shieldstral: 3B open-weights model for multimodal moderation

A 3B open model scores content safety from plain-language policies, no retraining needed.

AI-Generated Images Discourage Me from Reading Your Blog

AI-generated images on a blog make readers suspect the text is AI-written too

Gemini Robotics 2 brings whole body intelligence to robots

Gemini Robotics 2 controls whole humanoid bodies and enables multi-robot teamwork.

Flux 3

FLUX 3 learns images, video and audio in one model; video enters early access

Qwen-Image-3.0: Rich Content, Authentic Details, Deep Knowledge

Qwen-Image-3.0 announced; the post has no body, and comments debate quality and closed weights.

NYC may require landlords and realtors to disclose the use of AI in listings

NYC may require AI disclosure in property listings, sparking an HN debate over enforcement.

$100 AI Music Video: Claude Fable 5 vs. GPT-5.6 Sol

Given a budget and tools, two LLM agents made music videos; Claude Fable 5 cost up to $73.65 total.

Inkling – Open-Weights 975B Parameter LLM

Thinking Machines released Inkling, an open-weights MoE model taking text, image, and audio input.

Ghost Font: A font that humans can read but AI cannot

Ghost Font hides text in motion, but commenters say AI models still read it

OfficeCLI: Office suite for AI agents to read and edit Microsoft Office files

A single-binary CLI lets agents create, edit and render Word, Excel and PowerPoint files without Office

Claude-real-video － any LLM can watch a video

Scene-aware frame extraction plus transcripts lets any LLM 'watch' video with fewer frames than 1fps sampling.

Towards Robustness against Typographic Attack with Training-free Concept Localization

Adjusting CLIP's text-reading attention heads, with no training, cuts typographic attack errors.

Moebius: 0.2B image inpainting model with 10B-level performance

A 0.22B inpainting model claims FLUX.1-Fill-Dev-level quality at over 15x faster total inference.

How to setup a local coding agent on macOS

Adding an MTP draft model to Gemma 4 raised M1 Max speed from 58 to 72 tok/s

1-Bit Bonsai Image 4B Image Generation for Local Devices

Binarizing FLUX.2 Klein 4B weights makes image generation run on an iPhone

Diagnosing Failure Modes of Shared-State Collaboration in Resource-Constrained Visual Agents

Adding a shared board to 4B–8B vision agents lowered chart accuracy; a verify gate recovered most of it.

Claude Token Counter, now with model comparisons

Opus 4.7's new tokenizer counted 1.46x more tokens than 4.6 on the same system prompt

Chain-of-Thought Degrades Visual Spatial Reasoning Capabilities of Multimodal LLMs

Adding CoT prompts to multimodal LLMs lowered spatial reasoning accuracy by about 3% on average.

Show HN: Real-time AI (audio/video in, voice out) on an M3 Pro with Gemma E2B

Gemma 4 plus Kokoro gives about 0.7s to first audio on an M3 Pro, fully on-device.

What peak image prompt engineering looks like:

A one-line prompt with iPhone 6 and flash constraints yields realistic, uncanny images

Show HN: I turned a sketch into a 3D-print pegboard for my kid with an AI agent

Sketch photo plus two dimensions given to Codex produced printable 3D parts in about a minute.

Show HN: Gemini can now natively embed video, so I built sub-second video search

Embedding video chunks directly lets text queries find and trim the matching clip

Generation Models Know Space: Unleashing Implicit 3D Priors for Scene Understanding

Adding video-generation features to an MLLM improved 3D grounding accuracy without 3D labels.

On Optimizing Multimodal Jailbreaks for Spoken Language Models

Jointly optimizing text suffixes and audio noise raised jailbreak rates 1.5x to 20x over unimodal attacks.

SAVeS: Steering Safety Judgments in Vision-Language Models via Semantic Cues

Overlaying a red circle on an image can flip a VLM's safety refusal decisions

Unified Spatio-Temporal Token Scoring for Efficient Video VLMs

Pruning 50% of vision tokens made video VLMs 1.62x faster with only a 0.7% accuracy drop.

Loc3R-VLM: Language-based Localization and 3D Reasoning with Vision-Language Models

Adding layout reconstruction and pose tokens to a 2D VLM boosted 3D question answering and localization.

EchoGen: Cycle-Consistent Learning for Unified Layout-Image Generation and Understanding

Jointly training layout-to-image generation and grounding in a cycle reached 54.61 AP on MS-COCO.

ManiTwin: Scaling Data-Generation-Ready Digital Object Dataset to 100K

One image becomes a sim-ready, annotated 3D asset; 100K assets built automatically

The Cost of Reasoning: Chain-of-Thought Induces Overconfidence in Vision-Language Models

CoT raises VLM accuracy but inflates token-probability confidence; majority-vote consistency stays reliable.

Kestrel: Grounding Self-Refinement for LVLM Hallucination Mitigation

Grounding with SAM3 and evidence-gated refinement cut LVLM hallucinations without any training.

Visual-ERM: Reward Modeling for Visual Equivalence

A reward model that judges rendered images lifted chart-to-code by 8.4 points via RL

Geometry-Guided Camera Motion Understanding in VideoLLMs

Injecting 3D-model camera motion labels into prompts made VideoLLM descriptions more accurate