TL;DR An AI that writes code to generate images or video can run that code successfully while still failing to meet what the visuals actually need to look like. This paper proposes a way to measure that "program-to-visual" gap. Title: MaLiang-Harness: A Programmable Path to Image and Video Generation URL: Points 🖼️ A stateful framework that inspects and revises generated output while preserving state across edits 🔧 A traceable generation process links each code change to its rendered visual outcome 🎨 Unifies Canvas, SVG, Scene2d, and Three.js backends under one shared protocol 📊 On 50 text-to-image tasks, GPT-6-Astra hits 100% generation success and 96.0% full quality compliance 🎬 On 13 text-to-video tasks, GPT-6-Astra again reaches 100% success and 76.9% full quality compliance ⚠️ DeepSeek-class models manage only 12-40% on images and 0% on video 🔍 Two models with identical general-capability scores diverge sharply, 44% vs 88% on drawing quality pass rate The point that really lands: a general benchmark score alone doesn't tell you how good a model actually is at this. #MultimodalAI# #ImageGeneration#

Ask for "a lizard wearing sunglasses" and you get either no sunglasses or a distorted lizard. Google just gave the control problem behind image generation a mathematical answer. Title: How Diffusion Controller unifies and simplifies AI image generation URL: 📝 Overview Diffusion Controller reframes image generation as a continuous control problem rather than a sequence of isolated steps. It keeps the base model frozen and adds a lightweight side network to steer its behavior. ❓ Problem Solved Inference-time guidance and heavy fine-tuning methods like LoRA sit in two disconnected camps with no unified mathematical principle, leaving engineers to guess their way through balancing preference alignment against image quality. 💡 Method & Approach The side network observes intermediate denoising states and injects tiny steering corrections without touching the model's internal weights. Two implementations — a policy-gradient (PPO) approach and a reward-weighted loss approach — both include guardrails against quality degradation. This design even lets you control closed-source models you can't otherwise access. 📊 Results On Stable Diffusion v1.4, Diffusion Controller outperformed LoRA across both SFT and RWL training tracks despite touching fewer layers, hit a 90% win rate over baseline in its fully unlocked version, and delivered the best subjective quality in human evaluation on complex prompts. #ImageGeneration# #DiffusionModels#

🎨 A new unified visual AI where "seeing" and "creating" get trained to reinforce each other in one shared model. Title: SenseNova-U1.5: Towards Native Unified Visual Intelligence URL: 🧩 Overview OpenSenseNova's 8B Mixture-of-Transformers model drops external vision encoders and VAEs entirely, processing raw pixels directly while handling image understanding, reasoning, generation, and editing in one model. ⚙️ Problem it solves Most unified models route understanding and generation through separate latent spaces, causing consistency loss. This work aims to truly unify them on one shared representation. 🛠 Methodology Visual tokens are projected onto a 2D feature field and reconstructed with a spatial decoder that exchanges info across neighboring 32×32 regions. Four specialized experts (aesthetics, OCR, editing, infographics) are trained separately with RL, then merged into one model via On-Policy Distillation. 📊 Results GenEval overall hits 0.92, ImgEdit reaches 4.59 (up from 3.90 in U1), and BizGenEval jumps from 51.4% to 72.2% — outperforming its predecessor and competitors across understanding, generation, and editing alike. #UnifiedMultimodalAI# #ImageGeneration#

🎨 A text-to-image model trained entirely from scratch with a fully open recipe, closing in on top-tier model performance. Title: LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes URL: 🖼️ Overview LLaDA-Image trains a 6B-parameter Diffusion Transformer from scratch and pairs it with a dLLM-based vision-language understanding module in one unified generation-and-editing model. A TwinFlow-distilled Turbo version also ships, cutting inference down to just 2-4 sampling steps. 🧩 Problem Solved High-quality image generation has traditionally demanded massive paired image-text datasets, while inconsistent caption quality and the difficulty of unifying generation and editing in a single model remained open challenges. 🛠️ Methodology & Proposed Approach The recipe centers on image-only pretraining: over 90% of the 220M training samples are image-only, with real images making up 98% of generation-stage data and over 70% even during supervised fine-tuning — a deliberately "real-data-dominant" strategy. A parameter-free RMSNorm stabilizes long training runs, and an editing architecture conditions on reference images through both a semantic path and a pixel path. 📊 Use Cases / Experimental Results LLaDA-Image scores 53.53 on the English Qwen-Image-Bench track and 53.38 on Chinese, beating the leading open-source Z-Image Turbo (51.66 / 52.71). It also tops the long-text-rendering CVTG-2K benchmark with 0.875 word accuracy. On the flip side, GenEval's counting sub-task sits at just 0.53, a clearly flagged area for improvement. #ImageGeneration# #DiffusionModels#

Image Generation Quality Mode is now available on the xAI API. This model has already powered the generation of over 300 million images on Grok. It brings higher realism, stronger text rendering, and better creative control for business professionals.

Image generation just landed on the xAI API. Pretty cool stuff—developers can now build some wild visuals. Have at it! - API Console: - Developer Docs:

Qwen-Image-2.1-Turbo is here, bringing image generation and editing down to just 8 denoising steps. 🤖 ⚡ The accelerated checkpoint keeps the same 7B visual architecture while using a ready-to-run 8-step sampling schedule. 🎨 Generate or edit images at the recommended 2048 resolution, with presets covering square, portrait, landscape, and widescreen formats. 🧠 CFG=1 and prefix KV caching reuse text and reference-image context across denoising steps for more efficient inference. 🛠️ Loads directly with QwenImage21Pipeline in Diffusers; the recommended scheduler configuration is bundled with the checkpoint. 📜 Qwen Research License Agreement.

Your image prompting just got absolutely cooked. It’s called Mono-Color, and it gives Hermes a full design system for creating one- and two-ink editorial visuals. Think risograph posters, duotone covers, halftone photography, zines, packaging, and bold social graphics. The interesting part is that this isn’t a new image model. It’s an official optional Hermes skill that sits on top of Hermes image generation and handles the design decisions for you: ink setup, exact palette, composition, negative space, typography, print texture, and even a quality check after generation. So instead of engineering some monster prompt, you can give Hermes a reference image and say something as simple as: “Turn this into a cobalt one-ink editorial poster.” Hermes builds the recipe, generates it, checks the result, and can retry if it breaks the one/two-ink rules. To get it, install the skill with: hermes skills install official/creative/mono-color