TLDR Design 2026-07-20

Figma now binds colors, type, and spacing to existing file variables when code-backed screens are brought onto the canvas, rather than importing them as hardcoded values. This applies to Make preview copies, code-to-canvas generation via the Figma MCP server, and live web pages captured through the Chrome extension. Frames also arrive with auto layout more often, letting edits resize and reflow so teams can refine designs instead of rebuilding them.

OpenAI has updated its redesigned ChatGPT app for Mac after widespread criticism of its confusing interface, restoring easy access to chat history and Projects. The app now features a clear toggle between Chat and Work, while a sidebar menu lets users switch between ChatGPT and Codex, making the standard chat experience much easier to find again. Although the update fixes the biggest usability complaint, the app still relies on a larger Electron-based framework and has a more complex interface than before, leading some to question OpenAI's focus on integrating new AI tools over maintaining a polished Mac experience.

Google Vids now lets users create personalized AI avatars from a selfie and voice recording, tied to their Google account and invisibly watermarked with SynthID. The update also brings the Gemini Omni model to Vids, enabling video creation from prompts and reference images, background/lighting fixes, and step-by-step edits without restarting. These additions expand Vids from a workplace presentation tool into a broader video platform, positioning it against avatar competitors like HeyGen and Synthesia. Access is restricted to users 18+ in select regions.

AI struggles with design because, unlike code, design has no automatic way to verify if an output is correct, forcing reliance on slow, subjective human judgment. This subjectivity pulls AI toward generic, "average" outputs, worsened by scarce design training data and models generating interfaces blindly, without truly seeing what they create. While vision-based feedback loops and better models may narrow some gaps, the subjective nature of taste remains a structural limit, meaning humans will still be essential for genuine design judgment and originality.

Meta briefly introduced a feature in its new Muse Image AI tool that allowed users to generate and edit images using photos from any public Instagram account without prior consent, with the feature enabled by default. After widespread criticism from creative professionals, unions, talent agencies, and privacy advocates, the company removed the feature within three days, acknowledging that it had "missed the mark." The incident highlights a broader concern about AI companies launching opt-out features that treat publicly shared content as available for AI use unless users explicitly disable it. It also underscores the importance of consent, creator rights, and vigilance, as platforms continue to push AI capabilities that can affect photographers, designers, illustrators, and other creative professionals.

Paste a public domain, and Branda pulls the logo, colors, style, and products, then generates 4–6 ready-to-ship ads: three for the company and up to three for its products.

Create 3D models, characters, game assets, and animations with V2Fun's AI 3D model generator.

Miora reads your canvas, remembers your style, and iterates with you. Not a one-and-done tool — a creative partner that gets smarter every session.

Design systems have adopted software engineering practices, but agents generating code against these systems still can't be verified for compliance without dedicated evaluation suites. Effective evals combine mechanical checks—validating props, tokens, and component usage against the actual system—with LLM-based judgment on pattern selection and edge-case handling, run repeatedly since underlying models and rules keep shifting. While a few teams like Meta's Astryx and Intuit have begun building these regression suites into their pipelines, running such evals in CI remains an uncommon practice rather than a standard expectation.

Dropdown lists are often overused because they save space and enforce valid input, but they also hide options, increase interaction cost, create accessibility issues, and frequently force users into inaccurate choices. In many cases, radio buttons (for a few options), comboboxes (for many options), or text inputs (for familiar values like dates or ages) provide a faster and more usable experience. Dropdowns are most effective only in a narrow set of situations: when there are roughly 5–10 predefined options, the field is secondary to the user's main task, or it needs to remain compact as part of a larger group of related controls. Rather than treating dropdowns as the default, designers should choose the input control that minimizes user effort while matching the task and the nature of the data.

Blender 5.2 LTS launches with major new features, including a Geometry Nodes-based physics system for cloth and hair simulations powered by an XPBD solver.

This roundup features 20+ fonts suited for subtitles and subheadings, spanning styles from geometric sans serifs to slab serifs, handwritten, and pixelated designs.

New New Typography challenges traditional typographic principles through historical references, contemporary experiments, and contributions from nearly 100 designers, encouraging readers to rethink what typography can become.