
Canva's Craft Plans the Poster Before It Paints It
Canva's first in-house image model writes a composition plan, then renders it. The template-trained data is the real advantage, and the missing reference image support is the real limit.
Category
327 articles

Canva's first in-house image model writes a composition plan, then renders it. The template-trained data is the real advantage, and the missing reference image support is the real limit.

Fal post-trained FLUX 3 for speed and got a 1139 Elo claim out of it. The vote count is preliminary, the confidence interval overlaps the leader, and the 75% discount runs three weeks.

Two convictions in Hangzhou and Changsha show that a fifteen-second generated video plus tampered camera data was enough to clear face verification and reset a payment password.

A ninety-second generated film for the Double Ninth Festival turned old photographs into paired memory scenes. The format is about to become a product category, and the marketing is already sloppy.

Analysts, a bull, a bear, a trader, a risk team, a portfolio manager. The org chart is the architecture, and the README's caveat about nondeterminism is the most important line in it.

No code, no build step, no benchmark. Four principles in a single CLAUDE.md, and more stars than most capability releases this year. That gap is the story.

Six phases, three verdicts, and one rule that matters: the agent that checks a finding is never the agent that found it. Cloudflare also admits a single run finds about half.

Inter everywhere, purple gradients, cards inside cards. A design skill for coding agents ships 24 commands and 62 deterministic detector rules that run in CI without a single token.

A Weibo topic claiming the AI version of The Legend of the Condor Heroes looks better than the live-action remakes is not a claim about acting. It's a claim about production standards.

Point a coding agent at a folder, type /graphify, get a graph with edges labelled extracted, inferred, or ambiguous. The tagging is the interesting part, not the token savings.

Black Forest Labs' 5.6B open-weights model claims to run on 8 GB. Adding up the component files gives you a different number, and the gap explains a lot about local image generation.

A generated video of a monkey rescuing a dog from floodwater reached Weibo's tech chart in October. The sequence that followed — moved, suspicious, verified — is now the default.