Lucy Edit 1.1 Dev: Open-Weight Instruction-Guided Video Editing on Wan2.2 5B

TL;DR

Lucy Edit 1.1 Dev is Decart’s open-weight ~5B-parameter instruction-guided video editing model, built on the Wan2.2 5B DiT + high-compression VAE stack. Given a source video and a free-text instruction, it performs clothing/accessory changes, character replacements, object insertions, color changes, and global scene/style transformations while preserving the original motion and composition. Decart pitches it as the first open-source instruction-guided video editing model, with a Diffusers LucyEditPipeline already merged and a technical report linked. Version 1.1 over the original release improves boundary blending, scene-wide (“global edit”) reliability, and general-object insertion (occlusion, shadows).

Key claims

Lucy Edit 1.1 Dev is the “first open-source instruction-guided video editing model” — text-only prompt, no per-edit masks or finetuning required for common edits [§Lucy Edit 1.1 Dev (5B)].
Built on the Wan2.2 5B architecture, inheriting its high-compression VAE + DiT stack so existing Wan-2.2 scripts/workflows adapt directly [§Lucy Edit 1.1 Dev (5B)].
Six edit categories are explicitly enumerated: clothing/accessory changes, character/creature replacement, object replacement, color changes, object addition (wearable + handheld + scene objects), and global background/style transformation [§Supported Edit Types].
v1.1 improvements over the initial 1.0 release: higher fidelity + prompt adherence + identity preservation; more reliable scene-wide (“global”) edits; cleaner cross-frame color/lighting blending at edit boundaries; better spatial grounding, occlusion handling, and shadows for newly-added general objects [§✨ What’s new in 1.1].
Recommended inference window is 81 frames at 480×832, conditioned on 20–30-word descriptive prompts using trigger verbs Change / Add / Replace / Transform to [§Inference, §Prompting Guidelines].
Diffusers PR #12340 introducing LucyEditPipeline was merged on 2025-09-16; initial 1.1 weights and reference code shipped 2025-09-18 [§🔥 Latest News].
No quantitative benchmark numbers are reported in the model card — claims of “more robust” edits, “significantly improved” object insertion, etc. are qualitative; a separate Technical Paper is linked but its content was not fetched at filing time [§Lucy Edit 1.1 Dev (5B)].

Method

Lucy Edit 1.1 Dev is a Wan2.2-5B DiT fine-tuned for instruction-guided V2V editing. At inference, the Diffusers LucyEditPipeline takes a source video (loaded as 81 frames at 480×832), a positive text prompt describing the desired edit, an optional negative prompt, and a CFG guidance_scale=5.0; the pipeline returns an edited 81-frame video at 24 fps. Mechanically the model conditions the Wan2.2 DiT denoiser on (a) the source-video latents and (b) the instruction text, and learns to produce target-video latents that obey the instruction while preserving the source’s motion field and spatial composition — the model card’s central claim is that motion is “preserved perfectly.” The training recipe is not described in the model card; the linked technical report (Lucy_Edit__High_Fidelity_Text_Guided_Video_Editing.pdf) is the place to look. The card emphasizes prompt-engineering guidance: 20–30-word enriched prompts work best, and four trigger verbs (Change / Add / Replace / Transform to) signal the intended edit category to the model.

Results

Results in the model card are qualitative demo grids — “Turn the man into an alien,” “Turn the woman into Harley Quinn,” “Make it snowy,” “Change the apron and blouse to a clown costume.” No FID/CLIP/edit-fidelity numbers, no comparison table to closed-source editors (Runway Aleph, Pika Modify) or to open in-context recipes (UNIC: Unified In-Context Video Editing, OmniTransfer: All-in-one Framework for Spatio-temporal Video Transfer). The card self-reports reliability tiers per edit category: clothing changes are “best performance”; character/object replacement and object addition are “strong / significantly improved”; color changes are “mixed reliability” with sometimes-exaggerated outputs; global transformations are “much improved” but extreme styles may still alter identity. Roadmap items still pending at the time of the card: local-inference ComfyUI nodes, LoRA / fine-tuning scripts (remote ComfyUI custom nodes, public Batch API, Diffusers pipeline, and the technical report are marked done).

Why it’s interesting

This is the first filed evidence of an open-weight, instruction-only video editor at the Wan2.2 scale — the closest filed analogues are closed (Kling’s UNIC: Unified In-Context Video Editing which targets the same edit taxonomy but via in-context multimodal conditioning rather than pure text instructions, and Alibaba’s Wan-Animate: Unified Character Animation and Replacement with Holistic Replication which is character-animation-specific). The Wan2.2 5B base means the broader open-source toolchain (DiffSynth, ComfyUI, Wan-2.x LoRAs) transfers more or less for free, which is a non-trivial ecosystem effect — contrast with MirageLSD: The First Live-Stream Diffusion AI Video Model, Decart’s previous releaseable artifact, which was closed and frame-by-frame rather than batch-edit. It also lands in the Decart product narrative as the second public open-weight artifact alongside MirageLSD’s productized v2v stack — same company, two different bets on the V2V axis (real-time stylization vs. batch instruction editing).