Lucy 2.5 Real-Time AI Video Editing: Decart's 40ms Live Stream Revolution

TL;DR

A third-party (Viddi AI / community-author “ResterChed”) writeup of Decart’s Lucy 2.5 release — a real-time video editing model that transforms an incoming live stream frame-by-frame at sub-40 ms latency and 30 FPS, positioned as the editing successor to Decart’s earlier MirageLSD. The post claims Lucy 2.5 is built on Decart’s “DOS” inference stack — MXFP8/NVFP4 quantization, dynamic sparse attention, and deep kernel fusion — plus a training-time trick called Self-Anchoring that periodically snapshots the model’s own output as a new reference to prevent identity drift over long streams. Ships with eight named edit modes (character replacement, virtual try-on, object add/remove/replace, background replacement, style transfer, VFX) at $0.02/sec real-time /$ 0.04/sec offline via an API-only pay-as-you-go pricing model. The piece is a product-launch community writeup, not a tech report, and mixes marketing framing with a small number of specific technical claims worth logging.

Key claims

Lucy 2.5 runs frame-by-frame video editing at under 40 ms latency, 30 FPS, at 720p, via Decart’s proprietary DOS inference stack [§How 40ms Works, table].
The stack combines MXFP8/NVFP4 quantization (down to 4-bit inference precision to cut memory bandwidth), dynamic sparse attention (compute attention only for regions that change, skip static-background recomputation), and deep kernel fusion (multi-operator merging into single CUDA kernels) [§How 40ms Works, table].
The 40 ms budget is claimed to be “roughly 4× faster than comparable systems,” though the comparison set is unspecified [§How 40ms Works].
Self-Anchoring addresses the temporal-drift failure mode (eye colour shift, texture blur, subtle facial-structure change over long streams): every few dozen frames the model takes a snapshot of its own output and uses it as a new reference frame; because the anchor comes from the model’s own output distribution, it “doesn’t diverge over time” and edits remain stable over minutes, including when subjects leave and re-enter frame [§Self-Anchoring].
Eight named edit modes ship at launch: character replacement, virtual try-on, object addition, object replacement, object removal, background replacement, style transfer, VFX effects; each mode takes combined text-prompt + reference-image control [§8 Edit Modes, table].
Pricing: $0.02/sec ($ 72/hour) for real-time 720p; $0.04/sec ($ 144/hour) for offline 720p; API-only, no local GPU required, WebRTC + SDK integration, Android and iOS SDKs [§Pricing].
Positioning vs Runway Gen-3/4, Pika, Sora: Lucy 2.5 is the only listed option operating in the “live real-time” regime (<40 ms, 30 FPS, unlimited continuous length), while all three competitors are batch-offline with seconds-to-minutes latency and 5–60s clip limits [§Lucy 2.5 vs Mainstream, table].

Method

The post is a product-launch community writeup, not a technical report — there is no architecture diagram, no ablation, no benchmark table beyond the FPS/latency headline. What the post does state, mechanically:

Inference recipe (DOS stack). MXFP8/NVFP4 quantization brings weights and activations to 4-bit for reduced memory bandwidth pressure. Dynamic sparse attention is described as skipping attention computation for regions that do not change frame-to-frame — a temporal-coherence sparsity trick, distinct from the token-pruning or Radial-Attention style spatial sparsity common in the Diffusion serving optimization thread. Deep kernel fusion merges the operator graph into single CUDA kernels to reduce memory read/write latency — same family as MegaKernels/ThunderKittens.
Training-time drift control (Self-Anchoring). Every ~few-dozen frames during rollout, the model snapshots its current output and injects it back as an updated reference frame for subsequent generation. The post frames this as a training/inference-time recipe rather than an architectural addition; the claim is that because the anchor is drawn from the model’s own output distribution (not a fixed initial reference), it is “in-distribution” and therefore doesn’t induce the distribution shift that a fixed clean reference does. The post explicitly ties this to the observed identity-drift failure mode of long live edits.
Editing interface. Channel-wise source conditioning is not confirmed here (the post doesn’t get into architecture), but the eight-mode editor + text-prompt + reference-image control matches the API surface of Lucy Edit 1.1 Dev (Lucy Edit 1.1 Dev: Open-Weight Instruction-Guided Video Editing on Wan2.2 5B) and EgoEdit’s category (EgoEdit: Dataset, Real-Time Streaming Model, and Benchmark for Egocentric Video Editing).
Deployment. API-only through Decart’s platform; WebRTC streaming for live use; mobile SDKs on Android and iOS; OBS-compatible for streaming production workflows.

The post is silent on model size, base architecture (whether Wan-derived like Lucy Edit 1.1 Dev, or an evolution of MirageLSD’s frame-by-frame diffusion), training data, distillation recipe (few-step? consistency? shortcut?), and any quantitative comparison to prior work.

Results

Headline throughput: 30 FPS at 720p, sub-40 ms per-frame latency, unlimited continuous stream length. No benchmark scores (VBench, editing benchmarks, human eval) are reported.
Comparative table vs Runway Gen-3/Gen-4, Pika, and Sora is a categorical (regime) comparison, not a quality benchmark — the paper takes for granted that offline tools “still offer superior generation quality and creative control” and that Lucy 2.5’s win is the real-time regime, not raw quality.
Application-level claims: VTubers get identity-stable avatars including through leave-and-return, live commerce gets background/wardrobe swaps without green screen, ad creative gets infinite real-time variations of a base video.

Why it’s interesting

Lucy 2.5 is Decart’s second product-surface release on the live-stream video editing regime, and this community writeup is the first public description of two new named techniques that don’t appear in Decart’s earlier MirageLSD post (MirageLSD: The First Live-Stream Diffusion AI Video Model):

Self-Anchoring is a new named point in the Autoregressive Video Generation drift-fix taxonomy. The wiki’s concept page catalogues at least four training-time injection sites for the drift-suppression idea — MALT’s noise-augmented memory (MALT Diffusion: Memory-Augmented Latent Transformers for Any-Length Video Generation), Context Forcing’s Error-Recycling Fine-Tuning (Context Forcing: Consistent Autoregressive Video Generation with Long Context), LongVie 2’s degradation-aware training (LongVie 2: Multimodal Controllable Ultra-Long Video World Model), and FlowAct-R1’s generated-GT memory sampling (FlowAct-R1: Towards Interactive Humanoid Video Generation) — plus one inference-time recipe (FlowAct-R1 Memory Refinement). Self-Anchoring as described here sounds closest to a periodic in-distribution reference-refresh — neither a training-time perturbation nor a re-noise pass, but a rolling replacement of the fixed initial reference by fresh model output. This is a fifth injection site worth watching. The post’s claim that the anchor “comes from the model’s own output distribution” and therefore doesn’t diverge is a strong statement — it’s the operational form of the argument Causal Forcing (Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation) makes about distributional matching between train and rollout, applied to reference conditioning instead of context conditioning.
Dynamic sparse attention “for regions that change” is a temporal-coherence sparsity primitive. The Diffusion serving optimization and IO-Aware Kernel Design threads catalogue static and semantic-aware attention sparsity (Radial Attention, Sparse VideoGen2, MonarchRT) but the specific “only recompute attention where the frame changed” formulation isn’t in the current filed set. If real, this is a distinct third axis alongside token-level sparsity and static attention masks.

The honest caveat: this is a community-author blog on Hugging Face’s Community Article surface, not a first-party technical report from Decart. Every headline number here — the 40 ms budget, the 4× comparison, Self-Anchoring’s stability, the eight edit modes — needs corroboration from Decart’s own materials before being treated as ground truth. The post also mixes commercial framing (pricing table, use-case pitches for VTubers/commerce/ad-creative) with the technical claims, and the “Lucy 2.5 vs mainstream tools” table is a category comparison, not a benchmark. Treat this filing as a marker that Lucy 2.5 exists and that the DOS/Self-Anchoring nomenclature is now in circulation, not as an endorsement of the specifics — Decart’s own tech-blog post, if one lands, should be filed separately and preferred as the canonical source.