In short — LLMs have gotten good enough to hold context and keep track of a protagonist across tens of thousands of tokens, yet multimodal generation still struggles with character consistency between cuts, rough transitions, and awkward physics. Both problems grow from the same root: attention that preserves identity across a long context. And since 2025, AI agents have begun designing and discovering neural architectures themselves (ASI-Arch, AlphaEvolve). This piece traces, from public research, whether "AI designing AI" is reaching multimodal's consistency problem.
1. The question — text keeps its protagonist, so why can't video?
Text generation crossed a threshold. Across a document of tens of thousands of tokens, an LLM holds on to a character's name, personality, and speech patterns from the opening to the end, and can take writing you already have and revise it in context. What makes that possible is attention — a mechanism that computes relationships even between distant tokens so the model never loses track of who this sentence is about.
Multimodal works on the same principle. Keeping "the person in this frame is the same person from three seconds ago" is exactly the same kind of problem as keeping a protagonist in text. It just runs at far higher dimension (pixels × frames), and the failures are far more visible.
As of 2026, multimodal generation still wrestles with four weaknesses.
- Character consistency across cuts and clips — face, hair, wardrobe, and build shift subtly when the scene changes
- Smooth transitions — stitch clips together and motion, lighting, and camera jump (flicker, jump cuts)
- Physical and anatomical plausibility — fingers, joints, and object interactions break down
- Holding a long narrative — past ten seconds and into the minutes, story and character arc fall apart
The research literature is clear about the bottleneck. Most mainstream backbones are 2D-native, so producing temporally consistent video means bolting on a separate temporal attention module or smoothing stage; without it you get flicker and artifacts. (Zylos Research)
In other words, keeping the protagonist in multimodal is not magic — it is an architecture design problem. And that is exactly where the second trend meets it.
2. AI started designing AI — ASI-Arch and AlphaEvolve
One of the biggest shifts of 2025 is that neural architectures began to be discovered by AI agents rather than people. Where traditional NAS (Neural Architecture Search) only picked optima inside a search space a human had defined, the newest systems invent structures nobody had thought of.
ASI-Arch — "the AlphaGo moment for architecture discovery"
Released in July 2025, ASI-Arch is an autonomous research system that produces neural architectures without human intervention. It casts LLMs as agents split into three roles — Researcher (hypothesis), Engineer (implementation), Analyst (analysis) — and runs hypothesis → implementation → training → validation on its own.
- Roughly 20,000 GPU-hours spent on 1,773 autonomous experiments
- 106 new SOTA linear-attention architectures discovered
- They outperform human-designed baselines
- Above all, discovery scales linearly with compute — double the GPU-hours and you find twice as many strong models
The paper is titled "AlphaGo Moment for Model Architecture Discovery." It is a declaration that the bottleneck of human research capacity, which only grows linearly, can be broken by autonomous discovery that scales with compute. (arXiv 2507.18074)
That linearity is the most important line in this article. Until now the pace of architecture improvement was tied to the number of researchers. Tie it to compute invested instead, and the improvement cycle becomes something you can buy forward with money and GPUs.
AlphaEvolve — a coding agent that evolves the algorithm itself
Announced by Google DeepMind in May 2025, AlphaEvolve is a Gemini-powered evolutionary coding agent. It iteratively evolves entire programs rather than single functions, verifying performance with automatic evaluators.
- Discovered an algorithm computing 4×4 complex matrix multiplication in fewer operations, breaking a record that had stood since 1969
- The algorithms it found were deployed in Google's data centers, chip design, and AI training pipelines
What the two cases share is clear. Hand the search–implement–evaluate loop to an agent, and structures people never found come out in proportion to the compute you pour in. And the fact that what ASI-Arch found happens to be attention-family architectures connects directly to this article's subject.
3. So is it being used for multimodal? — what the survey shows
The conclusion first: no large-scale, direct case of autonomously discovering a multimodal generation backbone has surfaced yet. Autonomous architecture discovery in 2025–2026 concentrates on linear attention in language models. But three separate signals converge hard.
① What is being discovered is exactly the part responsible for consistency
What ASI-Arch found were linear-attention architectures. And what governs character consistency in video is precisely temporal attention and long-context handling. Research points to a Long-Context Video Transformer — holding plot and character arc across ten to twenty minutes — as the next step. The area autonomous discovery is currently good at (efficient attention) and the area multimodal is thirstiest for (long-context consistency) are the same place. (Long Context Tuning for Video Generation, arXiv 2503.10589)
② Multimodal itself is being redefined as a reasoning agent
Next-generation models are heading toward combining multimodal composition — weaving text, image, and video freely — with the character of an active agent that reasons about complex user intent. Research now frames video generation as a paradigm of multimodal reasoning. Agentic judgment is entering the generation pipeline itself. (Thinking with Video, arXiv 2511.04570)
③ Combining NAS with generative models already exists as a methodology
Work using generative models for architecture search — graph-diffusion-based NAS, for instance — has accumulated. The tools are ready; what remains is applying the methodology to multimodal generation backbones at scale. (Multi-conditioned Graph Diffusion for NAS, arXiv 2403.06020)
To sum up — "AI agents autonomously designing multimodal backbones" is still early, but the materials, the motivation, and the method are all in place. Extending the ASI-Arch approach to multimodal looks like a matter of time and compute rather than a technical leap.
4. Meanwhile, consistency is being solved at the product level
While autonomous architecture discovery matures, commercial models have pushed consistency to a workable level through engineering. This is what a creator can actually use today.
| Model | Reference images | Character consistency | Strength |
|---|---|---|---|
| Nano Banana Pro | Up to 14 | Up to 5 people in one scene without fine-tuning | Storyboards and campaigns where the same characters recur across many images |
| Seedream 4.5 | Up to 10 | No separate figure stated | Multi-image editing that preserves subject detail when you supply source images |
The split between them is clear. For making a character who does not exist recur across many cuts, Nano Banana Pro is the fit; for editing while preserving source material you already have, Seedream is. (fal, Nano Banana vs Seedream)
There is also one shared technique independent of model choice: fixing the seed to hold a character across multiple clips. It has effectively become the standard.
As of early 2026, leading video models generate native-resolution clips of 8 to 20 seconds with synchronized audio, plausible physics, and characters consistent between cuts. Not perfect — but compared with a year ago, "keeping the protagonist" has already reached a point you can work with. (WaveSpeed)
5. The XBRUSH view — what this means for creators
XBRUSH is an AI creative platform that handles image, video, and audio in one workspace. Translated into working terms, the trends above come down to three things.
1. Keeping the protagonist is now a workflow, not a feature
Putting a brand model, a mascot, or a product consistently across many cuts is achievable through reference images and seed management. You can shoot an entire campaign with the same face, not just a single ad.
XBRUSH AI Studio's Cinema implements exactly this flow. Pick a persona you have registered, upload reference images alongside it, write one line of an idea, and an ad video featuring that character is generated.
The results differ completely in setting and situation, yet the person appearing in them is the same. A concert stage, a city street, an indoor close-up — all carried by one persona.
2. Architectural progress arrives as consistency that takes less handling
What a person currently manages through seeds and references moves toward something the model maintains on its own as long-context attention matures. Fewer variables for the creator to track.
3. AI designing AI changes the pace of tooling itself
If discovery scales linearly with compute, multimodal quality improvement scales with compute invested rather than the old number of researchers. For creators that means shorter improvement cycles — the experience of "this didn't work six months ago and now it does" repeats more often.
In one line: the attention that keeps a protagonist in text and the attention that keeps a character in video are the same problem, and AI agents have started finding that structure themselves. Multimodal consistency is not a "someday" problem — it is scaling right now.
Frequently Asked Questions
What does it mean for an AI agent to design a neural network directly?
It goes beyond traditional NAS, which picks from candidates a human defined, to LLM-based agents that hypothesize, implement, train, and validate new architectures themselves. The flagship case is ASI-Arch, released in July 2025, which used 20,000 GPU-hours and 1,773 autonomous experiments to discover 106 new SOTA linear-attention architectures that outperform human-designed baselines.
Is this approach being used for image and video generation yet?
Autonomous discovery in 2025–2026 concentrates on efficient attention in language models, and large-scale autonomous design of multimodal generation backbones is still early. That said, the discovery target (attention structures) is the same component that governs video consistency, multimodal is evolving toward agents and reasoning, and methodologies combining NAS with generative models already exist — so the extension looks like a matter of time and compute.
Why is character consistency in video still fundamentally hard?
Because mainstream backbones are 2D-native. Producing temporally consistent video requires a separate temporal attention module; without it you get flicker and artifacts. It is essentially the same "maintaining identity across a long context" problem as keeping a protagonist in long-form text.
How do I keep a character consistent right now?
Fixing the seed and working with reference images is the standard. Nano Banana Pro accepts up to 14 reference images and holds consistency for up to 5 people in one scene without fine-tuning, while Seedream 4.5 takes up to 10 references and excels at editing that preserves subject detail from source images. In an integrated workspace like XBRUSH you can register a persona and use it together with references, managing the whole process inside the image, video, and audio production flow.
How do AlphaEvolve and ASI-Arch differ?
AlphaEvolve (Google DeepMind, May 2025) is a general-purpose coding agent that evolves whole algorithms and programs; applied to mathematical and infrastructure problems such as matrix-multiplication optimization, its results were deployed in Google's data centers, chips, and training pipelines. ASI-Arch is a system specialized in neural architecture discovery specifically. Both share the same principle: hand the search-implement-evaluate loop to an agent and discoveries arrive in proportion to compute.
References
- ASI-Arch: AlphaGo Moment for Model Architecture Discovery — arXiv 2507.18074
- AlphaEvolve — Google DeepMind blog
- Video generation and consistency bottlenecks — Zylos Research (2026-02), WaveSpeed 2026 guide
- Long-context video — Long Context Tuning for Video Generation, arXiv 2503.10589
- Multimodal reasoning and composition — Thinking with Video, arXiv 2511.04570
- NAS × generative models — Multi-conditioned Graph Diffusion for NAS, arXiv 2403.06020
- Product-level consistency comparison — fal, Nano Banana vs Seedream
This article is a trend survey based on public research and market material as of August 2026. The figures cited were reconfirmed against their original sources at publication, but this field moves quickly — please check each source again before relying on it.