Temporal Fixer

Image edits, video fixes.
A self-supervised, editor-agnostic temporal fixer that turns any per-frame image editor into a temporally-consistent, streaming video editor — with zero paired edit data and a single generation step.

---

The idea

A causal video model should not be asked to edit; it should be asked to fix.

State-of-the-art image editors (e.g. Qwen-Image-Edit) are excellent per frame but, run frame-by-frame on a video, produce a flickery draft — each frame lands in a slightly different look. Instead of training yet another dedicated, paired-data video editor, we keep the powerful image editor frozen and add a lightweight causal video fixer that removes the flicker while preserving the edit. The fixer is trained only on clean video (pseudo-style ⊗ per-frame degradation), so it never needs a single paired (source, edited) video.

_{SOURCE · PER-FRAME EDIT (independent frames → flicker) · TEMPORAL FIXER (ours). In motion, the middle panel jitters frame-to-frame while ours stays temporally stable.}

Headline results

Measured on OpenVE-Bench (spatially-aligned style edits) with the flow-warped flicker metric (warp-error; lower = more temporally consistent; natural-video floor ≈ 1.05):

	flow-warped flicker ↓	edit preserved (CLIP) ↑
Per-frame image editor (independent seeds)	48.4	—
Ours (co-adapted editor + fixer)	6.13	0.959

End-to-end flicker cut by 87% (48.4 → 6.13), single generation step, edit appearance retained, no paired edit data. The fixer also beats a naive optical-flow smoother on both flicker and edit-preservation.

The result generalizes across a second independent benchmark and different edit types:

Benchmark	flicker reduction	VBench motion-smoothness	VBench subject-consistency
OpenVE-Bench (global style)	−87%	0.838 → 0.979	0.670 → 0.918
FiVE-Bench (color + material)	−68%	0.863 → 0.966	0.704 → 0.841

Head-to-head vs a dedicated streaming editor (LiveEdit)

Same 48 videos, official VBench protocol. LiveEdit is a strong, end-to-end, edit-supervised streaming video editor; ours is editor-agnostic and uses no paired edit data.

VBench dim	Ours	LiveEdit	note
dynamic_degree (motion)	0.375	0.25	ours preserves more motion
motion_smoothness	0.979	0.993	close
subject_consistency	0.918	0.983	LiveEdit leads
background_consistency	0.929	0.966	LiveEdit leads
imaging_quality	0.565	0.714	LiveEdit leads (bounded by the frozen editor's per-frame re-render)
aesthetic_quality	0.535	0.580	close

Honest takeaway. Our system is temporally near-SOTA (motion-smoothness within noise of LiveEdit) and uniquely motion-preserving — dedicated editors damp motion to look smooth (DD 0.25), whereas ours keeps the source dynamics (DD 0.375). The remaining gap is per-frame imaging quality, which is fundamentally capped by the frozen image editor's re-render. Closing that gap is the focus of our ongoing work (a distribution-matching training objective, below).

More comparisons

_{Sketch style — per-frame line art flickers heavily; the fixer stabilizes it.}

_{Pop-art style.}

_{Watercolor style.}

_{GIFs are shown at reduced resolution/frame-rate; full-resolution MP4s available on request.}

Method at a glance

1. Editor (frozen). A per-frame image editor with a stabilized sampling protocol (shared seed across frames) produces a draft. Fixing the sampler alone removes most of the "flicker" — much of it is a sampling artifact, not an editing artifact.
2. Fixer (learned, causal, single-step). A Wan-2.1 causal DiT, conditioned by channel-concat on [draft | source], is trained self-supervised on clean video only. At inference it maps the draft to a temporally-consistent output in one Euler step — a streaming, editor-agnostic module.

The two stages are co-adapted: the fixer is trained to match the stabilized editor's statistics. That co-adaptation is what breaks the ceiling that single-sided fixes could not.

Ongoing work

We are testing whether the per-frame imaging-quality ceiling is an artifact of the regression (MSE) training objective rather than the fixer mechanism or model scale. Replacing regression with a self-supervised distribution-matching (DMD) objective — pulling the fixer's output onto the real clean-video manifold while anchoring to the draft — is designed to synthesize sharp, temporally-coherent detail without paired data. Results forthcoming.

Status

Research preview. Numbers above are from held-out benchmark evaluations (OpenVE-Bench, FiVE-Bench, VBench). Code release is in progress.

_{Comparison videos are model outputs on public benchmark clips, shown at reduced resolution/looped for illustration.}