Lucy — Decart's real-time image-to-video DiT (lucy.decart.ai)

TL;DR

Lucy is Decart’s image-to-video product landing page. Lucy is a diffusion-transformer (DiT) model family — sized at 5B and 14B — that generates a 5-second 720p video from a single still image plus a text prompt, with the 5B variant claimed to produce a full 5-second clip in under 5 seconds end-to-end (faster than real-time). The page itself is a marketing/product surface (no method writeup); identity-bearing details come from Decart’s API platform and ecosystem coverage (WaveSpeedAI, ImagineArt, eachlabs). It is filed primarily as a market datapoint: how fast a production image-to-video service can be served given that no published paper accompanies the launch.

Key claims

Lucy is built on a Diffusion Transformer architecture and is positioned by Decart’s ecosystem partners as part of “Decart’s Lucy model family” that “pushes the boundaries of video generation speed” [Decart platform docs; WaveSpeedAI launch post].
The 5B variant generates a full 5-second video in under 5 seconds end-to-end, i.e. greater-than-real-time generation [WaveSpeedAI launch post].
Lucy supports both text-to-video and image-to-video — the same model accepts a text description or an image as initial frame [Decart platform model card].
Output is 720p MP4 at 16:9 or 9:16 aspect ratios, 5-second clips [ImagineArt docs].
Decart’s ecosystem materials attribute the speed to “custom low-level engineering” and “hardware-level optimization for maximum throughput” rather than to a specific algorithmic distillation recipe [WaveSpeedAI launch post; eachlabs model card].

Method

The lucy.decart.ai page is a product surface (image upload + prompt field + generate button) and does not document the method. From adjacent Decart materials: Lucy is a DiT, sized 5B and 14B in the publicly named variants, accepting a still image as the first frame plus a natural-language motion prompt and emitting a 720p MP4 of ~5 seconds. Decart’s claimed differentiator is systems-level — “custom low-level engineering” at the kernel / scheduler / batching layer — rather than a disclosed algorithmic novelty. There is no companion paper, technical report, or open-weights release at filing time. Whether the speed comes from few-step distillation, a custom attention kernel stack, FP4/NVFP4 execution, an autoregressive frame-by-frame rewiring, or some combination is not stated.

Results

No quantitative benchmark numbers (VBench, ArtificialAnalysis ranks, FVD, motion-quality scores) are disclosed on the lucy.decart.ai page or in the partner pages. The only quantitative claim is the wall-clock-to-output budget: 5-second video in <5 seconds for the 5B variant. Comparable production-class baselines on the wiki: Seedance 2.0 (Seedance 2.0: Advancing Video Generation for World Complexity) and Kling at the quality end; LTX-2 (LTX-2: Efficient Joint Audio-Visual Foundation Model) and the FastVideo / LTX-2.3 stack (FastVideo: Create a 5s 1080p Video in 4.5s on a Single GPU) at the latency end (which hits ~4.55 s for 5-second 1080p TI2AV on a single B200 via system-level optimization on an undistilled model).

Why it’s interesting

Pablo’s question — “is it just distillation or how do they achieve such a fast i2v model” — sits at exactly the open boundary the wiki has been mapping. The current evidence on that boundary: pure system-level serving recipes can already hit ~real-time at 1080p on an unmodified LTX-2.3 (FastVideo: Create a 5s 1080p Video in 4.5s on a Single GPU), which means Lucy’s <5s-for-5s claim does not require distillation on its own merits — a B200-class kernel stack plus end-to-end graph fusion is enough. The orthogonal lever is distillation: the wiki’s Diffusion Distillation thread documents 1–4 NFE students for video DiTs at Wan2.1 14B class (TMD, DiagDistill, RMD, Causal Forcing, Causal Forcing++) with VBench-Total parity to undistilled teachers. Decart’s own prior public artifact, MirageLSD (MirageLSD: The First Live-Stream Diffusion AI Video Model), is a live-stream diffusion model — i.e. they have already shipped an autoregressive / streaming video stack, so an AR few-step distillation path (Self-Forcing → Causal Forcing → DiagDistill family) is at least as plausible as a pure systems story. Without a tech report, both routes remain consistent with the launch claims; the filing is mainly a marker that Decart’s image-to-video product is now public and that the speed-quality frontier for i2v has compressed again.