InfinityEdit: Infinite Video Editing with a Lightweight Edit-Ignition Adapter
Yunze Tong, Mushui Liu, Canyu Zhao, Shiyi Zhang, Didi Zhu, Peng Zhang, Wanggui He, Jinlong Liu, Ying Chen, Hao Jiang, Pipei Huang, Bo Zheng
cs.CV
2026-08-21
A frozen 14B Helios plus a light adapter does unbounded stream edits. On 200 clips x 3 rounds, camera motion is 0.7654 and VLM faithfulness 3.83, std 0.023.
Instruction-based video editors assume a finished clip. The output matches the source length and lines up frame by frame. Restyling a live game or adding a camera move to a shot that is still running does not fit that contract, because future frames have not arrived yet.
Zhejiang University and Alibaba name the setting infinite video editing. Given a preceding segment and an edit request, the model must emit the next segment in time, continuing the stream and applying the edit. Requests keep arriving; each output becomes history for the next round. Three constraints follow. The edit has to land. Unaffected subjects and motion have to continue across the boundary, not be redrawn timestamp by timestamp. Errors feed into the next round, so quality cannot decay as edits stack.
Switching the scene prompt is not enough. Streaming generators use anchor frames and multi-scale history that pull later chunks back toward the unedited look, so the edit gets washed out.
The backbone is frozen 14B Helios-Distilled, an autoregressive video diffuser that emits chunks with a hierarchical history memory. Trainable pieces are adapter blocks inserted after each transformer layer, with zero-initialized output projections so training starts as identity.
Each adapter updates only the current denoising chunk, in three stages:
Training is flow matching. History frames are corrupted with some probability, matching the dirty self-outputs seen at inference. Noise levels are drawn from a Gaussian mixture centered on the discrete σ values used when an edit is applied, instead of Helios's resolution-tied pyramid. A two-phase curriculum first covers those σ values uniformly, then shifts mass and per-frame loss toward low noise and later frames in the chunk.
Supervision is synthetic triplets. Sources come from UltraVideo. Gemini 3 Flash turns abstract edit types into concrete instructions. The last source frame is processed by edit type: appearance edits go through Qwen-Image-Edit, camera-move edits keep the frame, then Wan2.2-I2V-A14B continues forward. Twenty raters filter on alignment, continuity, plausibility, and quality, each 1 to 4.
Inference is ignite-then-continue. The adapter runs only on the chunk where a request arrives. Later chunks are the frozen backbone. The anchor frame is reset to the first edited frame, the scene prompt is replaced by the current instruction, and a sliding history window keeps memory bounded.
No existing benchmark fits, because in-place metrics need frame correspondence that this task does not have. They build an OOD set: 200 UltraVideo clips held out from training, three sequential scene-grounded instructions per clip from Gemini-3-Flash, 15 edit types in four groups (entity change, style, camera, motion transfer), four chunks per segment. Baselines fall in three families: prompt-swapped Helios, in-place editors (Lucy-Edit, SANA-Streaming), and prompt-switching generators (Anchor-Forcing, Infinity-RoPE) that do not take the source video. Default output is 384×640 at 16 fps, trained on 32 H20 GPUs.
| Method | Camera motion | Motion smoothness | Temporal flickering |
| Helios backbone | 0.5494 | 0.9869 | 0.9641 |
| SANA-Streaming | 0.5432 | 0.9858 | 0.9631 |
| InfinityEdit | 0.7654 | 0.9833 | 0.9660 |
Camera motion is about 0.22 above the next-best source-conditioned method. Gemini-3.5-Flash scores 1 to 5:
| Method | Faithfulness | Quality | Preservation | Coherence |
| SANA-Streaming | 3.303 | 3.093 | 3.060 | 3.030 |
| Anchor-Forcing | 2.545 | 2.620 | 1.830 | 3.480 |
| InfinityEdit | 3.828 | 3.765 | 3.815 | 3.840 |
Faithfulness across three rounds is 3.860 / 3.805 / 3.820, std 0.023. Source-conditioned baselines drop 0.02 to 0.04 in aesthetic score after the first round; InfinityEdit stays flat. Lengthening each segment past 1,000 frames still follows instructions, and attributes from earlier edits persist.
Open-ended streams need "continue and change," not "rewrite this clip." Freezing a long-video generator and lighting an adapter only at edit boundaries is cheaper than full fine-tuning and keeps the streaming prior. The numbers say prompt swap is too weak, in-place editors freeze the timeline, and prompt-switching generators drop source identity.
Training targets and eval instructions both come from Gemini; target video comes from Wan and Qwen-Image-Edit. The adapter may be learning those synthesizers' edit distribution. The VLM judge is also Gemini-family, so faithfulness scores can be circular. There is no commercial-editor comparison, and 384×640 is not a shipping resolution. The paper notes language-only control and abrupt switches at edit boundaries. The adapter is wired to Helios's history interface; a new backbone means a rebuild. Parameter count and end-to-end latency are not reported.