WanPE: Towards Cinematic Prompt Enhancement for Modern Text-to-Video Generation

What it claims

Beyond the headline model, the page pins down the numbers and the training recipe. WanPE scales across 4B–397B variants, and the 397B improves human preference over raw prompts by 10.66–18.84 points at 5–15s and 50.86 points at 30s. It leads all evaluated commercial offerings at 5–15s and stays competitive with Seedance 2.5 at 30s. The claimed advantage grows with duration — the payoff is on long (30s) generation, not short clips.

The summary conflates the benchmark size: WanPEval is 249 curated requests (stratified by duration and granularity), evaluated via ~11K blind pairwise video assessments — the 11K is the assessment count, not the request set.

How it works

The core is reverse pair construction rather than forward prompt rewriting. The stated rationale: video generators are trained on video-grounded captions, so a forward enhancer's synthetic rewriting distribution creates a train/inference mismatch (shown as a t-SNE distribution gap). To fix this, they caption real videos to get the cinematic condition y, then reconstruct a user request x from it — so targets sit on the real caption distribution p_vg.

flowchart LR
  V[Real video clip] --> C[Caption to cinematic condition y]
  C --> R[Reconstruct request x from y]
  R --> S[SFT pairs x,y]
  S --> G[SC-GRPO: 9-dim reward]
  G --> P[Prompt enhancer]

Concrete pipeline details the summary omits:

• ~1.05M clips, ≤30s, 10 content categories, multi-stage quality/motion filtering.
• Request reconstruction uses gpt-5.4 with few-shot prompting from a pool of 2K human-written requests.
• SC-GRPO dataset ~15K prompts (broad coverage + flagged challenging cases). Reward comes from Qwen3.7-Max, scoring [0,100] across nine dimensions: style, subjects, actions, dialogue, sound, camera, lighting, spatial relations, scene — checking omissions, altered/weakened content, wrong subject–attribute or speaker–dialogue bindings, and cross-shot conflicts.

Evidence

Ablations (as stated): reverse construction outperforms forward rewriting, and SC-GRPO improves semantic consistency by 18.6–23.3 points across model scales. The authors also state WanPE transfers across video generators after format adaptation, though numbers for that transfer are not given in the fetched text.

Limits

• The fetched text cuts off mid–Section 2.4 (expert video evaluation); full result tables, per-scale numbers, and the transfer results are not visible here.
• Downstream results are measured on Wan3.0's generator; cross-generator claims rest on "format adaptation" without shown figures.
• Rewards and evaluation lean on other proprietary models (gpt-5.4, Qwen3.7-Max), so semantic-fidelity scoring inherits those judges' biases.
• No licence, weights availability, or inference cost stated on this page — only a project page (wan-pe.github.io).