Generated video's next test is physics, not how good it looks

The problem: pretty video that does not make sense

Generative video models are increasingly used as visual world simulators, yet their primary strength lies in rendering plausible pixels rather than reasoning about how a world should evolve, according to the EvoWorld preprint. Existing diffusion-based video models lack an explicit mechanism for maintaining world state and reasoning over event-conditioned consequences, often producing visually plausible but behaviourally static or causally inconsistent outputs. That is the gap a creator runs into in practice: a clip can look fine frame by frame and still fall apart as a scene, because something that left the frame stopped existing, or something that should have changed did not. m:…0731

The same problem shows up in how video is graded. PhyProbe, a preprint, argues that existing fine-tuned evaluators for video generation degrade sharply in no-correspondence and generated-generated settings, which are the cases where there is no clean reference video to compare against, or where both videos being compared are generated. If the grader only works when a real video is on hand, it is of limited use for the work creators actually do, which is generating new footage and judging whether it holds together. The field has been grading looks; it now needs to grade behaviour. m:…0826

EvoWorld: keep the world alive off screen

EvoWorld's answer is to split the job in two. It is a brain-renderer collaboration framework in which a vision-language model serves as the world brain, while a segment-autoregressive video renderer realises state transitions visually. The vision-language model observes generated segments to maintain an evolving account of the world, and reasons about plausible state transitions. In other words, one part of the system remembers what is happening, and the other part draws it. m:…0731

That split is what lets EvoWorld handle the hard cases. Within a bidirectional autoregressive paradigm, EvoWorld supports event-triggered world evolution, subject and scene revisitation, and out-of-sight dynamics rendering through model reasoning. An event can trigger a change, a subject can leave and come back, and something happening off screen can keep evolving rather than freezing. The authors also introduced EvoWorld-Data and EvoWorld-Bench for training and evaluating event realisation, causal reactions and revisit consistency. Experiments show that EvoWorld substantially improves event-conditioned causal evolution over recent models while preserving visual quality and consistency, according to the preprint. m:…0731

The scores, and what they actually show

Video-Coder's reported gains are large on the benchmarks that measure this. The authors report that Video-Coder improved Minimax-H3 by approximately 7 points on PhyGenBench and 24 points on VBench-2.0, and they have open-sourced the data, models and code. The researchers also curated two datasets for training, Video-Coder-SFT-6k and Video-Coder-RL-2k. These are preprint results, so they should be read as the authors' own report rather than an independent verdict, but the direction is clear: building the scene in code, then rewarding the result as video, moves the physics scores, not just the look. m:…0732

PhyProbe is the grader that would catch whether those gains are real. It uses a frozen spatio-temporal encoder and a lightweight scoring head, and it achieves strong correlation with human judgments, with scores anchored to a stable 0 to 1 scale. It outperforms prior methods on pairwise benchmarks for physical consistency in generated videos, particularly in no-correspondence and generated-generated settings. For a creator, the practical point is that a 0 to 1 physical-consistency score, if it holds up, is easier to use than a vibe check: a clip either obeys the world or it does not, and the score should say so. m:…0826

The image side of the same problem is more embarrassing. A preprint introducing the T2I-NSI benchmark finds that text-to-image models exhibit low negation sensitivity, with open-weights models scoring no more than 0.038, and that inverse interpretation of negation is essentially absent, at 0.00%. The primary failure mode is under-suppression, where the negated concept is rendered despite the prompt. Models do not usually draw the opposite of what you asked for; they simply draw the forbidden thing anyway. For anyone prompting an image with a negative, that is a reliability gap, not a quirk. m:…0836

What creators should take from Seedream 5.5

ByteDance's Seedream 5.5 sits on the other side of the same trade-off: reliability in professional production rather than raw novelty. ByteDance introduced Seedream 5.5 as a next-generation image creation model, and the company says it supports up to 100 rounds of successive editing while preserving unedited regions at the pixel level. It also generates native RGBA assets with an alpha channel for full and partial transparency, which matters for design workflows where an image has to drop into a layout rather than sit on its own. m:…2905

The text numbers are the ones creators will feel first. ByteDance says Seedream 5.5's severe text corruption rate is 0.4% and its minor text corruption rate is 1.7%, and that text rendering improves significantly, including complex character shapes and densely arranged small text. The company also says Seedream 5.5 outperforms GPT Image 2.5 Sunburst and Nano Banana 2.1 in text structure and realism, and that usability rates were more than twice those of the previous version in film and television and in office productivity. Those comparisons come from ByteDance itself, so they should be read as the company's own report, and a Pro version of the Seedream 5.5 series will follow. m:…2905