Solaris: Towards Interfaces That Are Generated, Not Coded
Yuval Alaluf, Omri Avrahami, Guy Bukchin Leshem, Michal Geyer, Kfir Goldberg, Elad Richardson, Diego Alarcón, Alejandro Alvarez, et al.
cs.CV · cs.AI
Quick verdict
Solaris generates interactive UIs frame-by-frame via a world model; compelling vision and a sizable user study, but thin technical disclosure and no code.
Summary
Solaris is Runway's 'interface world model': instead of implementing UIs in code, it autoregressively generates the interface frame by frame, conditioning on mouse and keyboard input. Built on the Gen-4.5 video model, it targets real-time interaction through three stages — autoregressive frame generation, few-step distillation of the denoising process, and training on its own outputs for long-session stability — while a separate LLM interprets user intent and emits per-state prompts, separating reasoning from rendering. The paper argues that coded interfaces are a lossy intermediate representation, supporting this with a screenshot-reconstruction experiment: multimodal LLMs (Claude Fable 5, Gemini, GPT-4o) rebuild 30 interfaces of increasing visual complexity, with fidelity degrading for image-heavy and natural-image content as measured by SSIM and DINOv3 similarity (values shown only in figures). A user study with 250 participants and roughly 7,500 pairwise judgments against a Claude Opus 5-coded interface found Solaris preferred 62% vs 25% (13% ties) for instruction following and 72% vs 21% (7% ties) for natural behavior. Limitations, candidly stated: unstable generated text, trust and grounding concerns, long-session coherence, and accessibility support. No architecture, training-data, latency, or cost details are given, and no code or data is released — the paper reads as a polished corporate tech report rather than a reproducible scientific contribution.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Translating interfaces through code is a lossy intermediate representation, demonstrated by degrading SSIM/DINOv3 fidelity as multimodal LLMs reconstruct 30 increasingly complex screenshots (§4.1).
- An autoregressive video world model conditioned on mouse/keyboard input, few-step distilled and trained on its own outputs, renders interactive 720p interfaces in real time with session-long coherence (§3).
- Users prefer Solaris over a Claude Opus 5-coded interface — 62% vs 25% for instruction following and 72% vs 21% for natural behavior — from 250 participants and ~7,500 pairwise judgments (§4.2).
- Pairing an LLM intent-interpreter with the visual world model decouples reasoning from rendering, enabling open-ended interaction semantics not predefined by developers (§2–3).
Strengths
Crisp paradigm articulation
§1–2 clearly separates Solaris from reproduction-oriented world models (NeuralOS, ViMo, CUWM): there is no reference implementation, so behaviors are open-ended; Figure 3 (the same drag interpreted as wrist-zoom vs hand-measurement) makes the claim concrete.
Sizable pairwise user study
§4.2 uses 250 participants, 30 interactions, and ~7,500 judgments on two questions; the margins (62% vs 25% instruction-following; 72% vs 21% naturalness) are large enough that sampling noise is an implausible explanation even without reported tests.
Candid, specific limitations section
Two complementary metrics for the translation-loss argument
Weaknesses
No technical disclosure of the system
Quantitative results exist only as figures
Single proprietary comparison baseline
Irreproducible evaluation assets
Thin citations and absent related work
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Solaris (arXiv 2609.00776, Runway) proposes an "interface world model": rather than implementing UIs in code, a video-generation model renders the interface autoregressively, frame by frame, conditioned on mouse/keyboard input, while a companion LLM interprets intent and emits per-state prompts. The paper makes four central claims: (1) code is a lossy intermediate representation for interfaces (§1, §4.1); (2) few-step distillation plus training on the model's own outputs makes an autoregressive video model fast and coherent enough for real-time 720p interaction (§3); (3) users prefer the generated interface over a Claude Opus 5-coded interface (§4.2); (4) separating reasoning (LLM) from rendering (world model) enables open-ended behaviors not defined in advance (Figure 3's same-drag/two-outcomes demo). This is a corporate tech report in paper form: no appendix, no architecture or training details, and a reference list where 3 of 11 entries are company blog posts.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Solaris?
Solaris is a research system from Runway described as an 'interface world model'. Instead of running an app built from code, it generates the visual interface frame by frame in real time, treating mouse clicks, drags, and typing as conditioning input, while a separate language model decides what your actions should mean in the scene.
How does Solaris compare to normal coded interfaces?
In the paper's user study, 250 participants made about 7,500 pairwise comparisons against an interface coded by Claude Opus 5. Solaris was preferred in 62% vs 25% of comparisons (13% ties) on following the requested interaction, and 72% vs 21% (7% ties) on behaving naturally within the scene.
Can I download Solaris or its code?
No. The paper provides no code, model weights, evaluation data, or public demo; the authors state only that Runway is 'working with key partners to launch Solaris publicly'.
What are Solaris's main limitations?
The authors list four: generated text is still unstable; grounding in verified content is needed so the system does not produce convincing wrong answers; long-session visual and semantic coherence remains unsolved; and generated interfaces do not yet integrate with accessibility tools such as screen readers.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- All quantitative results (Figure 10 fidelity plots, Figure 11 study proportions) exist only as graphics; exact SSIM/DINOv3 values are not recoverable from the extracted text.
- Figure contents (demo screenshots, Figures 1–8 and 12–14) cannot be assessed from text extraction; qualitative judgments about visual quality rest on the authors' captions alone.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.