WanPE trains a text-to-video prompt enhancer as a shot-by-shot cinematic planner by Video-grounded reverse construction from real videos plus a semantic-consistency Group Relative Policy Optimization (GRPO) pass, rescuing ~50 preference points at the 30-second horizon where raw user prompts collapse.
Modern Text-to-Video systems like Wan3.0 and Seedance 2.5 can now render up to 30 seconds of coherent video with camera moves, lighting, and audio. That length shifts the burden onto the text prompt: it has to lay out multiple shots, timing, dialogue, and camera choreography, not just describe a scene.
Today’s prompt enhancers were built for the previous generation (e.g. Wan2.2), where a short caption expansion was enough. They take a user request and rewrite it into a richer description via an LLM. The paper’s measurement of the problem: on a 30-second test set, feeding raw user prompts into Wan3.0’s generator scores 9.38 on expert preference, versus 60.24 when WanPE plans the shots. Whatever the old-style enhancers are doing, it is not enough for the long-horizon regime.
The authors argue two things are broken. First, forward rewriting produces text in a synthetic “rewrite” style that does not match the video-grounded captions the generator was trained on. Second, as plans get longer, user requirements (a specific character’s dialogue, a camera move, event ordering) get dropped, misassigned, or contradicted across shots.
The core trick is to flip the direction of supervised data construction. Instead of starting from a short user prompt and asking an LLM to invent a detailed cinematic plan, WanPE starts from ~1.05M real filmed video clips (up to 30s each) and works backwards.
Step 1: a multimodal captioner analyzes each real video and produces a hierarchical target y: a video-level summary plus timestamped shot descriptions covering composition, subjects, actions, lighting, camera, transitions, dialogue, music, and sound. Because these captions describe things a real crew actually coordinated, the joint structure across shots is real, not invented.
Step 2: an LLM (the paper uses gpt-5.4 with few-shot examples drawn from a pool of 2K human-written requests) compresses y back into a plausible user request x. The constraint: x may only contain requirements that y actually satisfies. This gives training pairs (x, y) where the target is guaranteed to be a real, video-grounded plan.
Step 3: standard SFT on those pairs, then a reinforcement stage called Semantic-Consistency GRPO. A judge model (Qwen3.7-Max) scores each generated plan on nine dimensions (style, subjects, actions, dialogue, sound, camera, lighting, spatial relations, scene), penalizing omissions, altered requirements, wrong subject-attribute or speaker-dialogue bindings, and cross-shot contradictions. That scalar reward drives Group Relative Policy Optimization (GRPO).
# Data construction (per real video v_i)
y_i = caption_video(v_i, category_instruction) # hierarchical shot plan
if not passes_structure_and_consistency(y_i):
skip
x_i = llm_reconstruct_request(y_i, few_shot_examples) # gpt-5.4
dataset.append((x_i, y_i))
# Training
pi = sft(pi_base, dataset) # video-grounded SFT
for x in sc_grpo_prompts:
ys = [pi.sample(x) for _ in range(G)]
rewards = [judge_9dim(x, y) for y in ys] # Qwen3.7-Max, 0-100
pi = grpo_update(pi, x, ys, rewards) # standardized advantages
Four sizes were trained (4B, 9B, 35B, 397B) initialized from Qwen3.5 variants, on 512 GPUs.
Evaluation is on WanPEval, a 249-request human-curated testbed spanning 5-30s durations and three request-detail levels (intent, scene, shot), scored by ~11K blind pairwise expert comparisons of generated videos plus a Bradley-Terry loss fit. All numbers below are expert preference score S (win + 0.5 * both-good, out of 100) with Wan3.0 as the downstream generator.
•
Enhancement helps more as videos get longer. WanPE-397B beats raw user prompts by +10.66 at 5s, +16.28 at 10s, +18.84 at 15s, and +50.86 at 30s. Long-horizon prompts really do collapse without planning.
•
Reverse construction beats forward rewriting. At the SFT-only stage, WanPE-397B-SFT scores 49.86 vs 39.49 for a heavily-tuned forward-rewriting baseline (27 iterations of prompt engineering with gemini-3.1-pro-preview) and 35.17 for SFT trained on forward-generated targets. So the gain is not just from having an LLM in the loop, and not just from SFT; it is from targets drawn from the real-video caption distribution.
•
SC-GRPO lifts semantic consistency by 18.6-23.3 points across all four model sizes (judged by gemini-3.1-pro-preview). At 397B, the “perfect” rate goes from 29.7% to 85.5% and the failure rate drops from 36.9% to 2.8%. This transfers to video: expert preference on WanPE-397B rises from 42.70 (SFT only) to 49.69 after SC-GRPO.
•
Head-to-head vs commercial systems (5-15s). WanPE-397B leads overall at 50.61 S / 61.85 BT, ahead of Seedance 2.0 (43.42), MiniMax-H3 (36.40), HappyHorse 1.1 (24.89), Kling 3.0 (20.80), and LTX-2.5 (17.30). On the 30s subset it scores 60.24 vs Seedance 2.5’s 59.76, essentially a tie.
•
Cross-generator transfer. After format adaptation via GPT-5.4, WanPE-397B improves preference by +14.45 on LTX-2.5’s generator and +5.17 on MiniMax-H3’s generator over each system’s native enhancer.
Caveat on interpretation: the head-to-head numbers compare full pipelines (enhancer + generator), so the WanPE-vs-commercial gaps also reflect whatever advantage Wan3.0’s generator has over the others. The isolated enhancer effect is the raw-vs-WanPE gap under a fixed Wan3.0 generator, and the reverse-vs-forward ablation.
•
If you are building a T2V product on a modern long-horizon generator, the finding worth acting on is that prompt enhancement is now a planning problem, not a decoration problem, and that a forward LLM rewriter, however carefully tuned, plateaus below what video-grounded targets achieve. If you have access to a captioned real-video corpus, replicating the reverse-construction recipe is more promising than iterating on rewrite prompts.
•
The SC-GRPO reward design (nine dimensions, judged by a frontier LLM, penalizing binding errors and cross-shot contradictions) is portable to any setting where a plan has to preserve structured user requirements across a long output. Worth testing whenever your rewriter drops or reassigns constraints in long generations.
•
If you use a hosted T2V API where the enhancer is baked in, the paper does not give you a way to swap it out. The cross-generator experiments required separately accessible enhancer and generator components (which MiniMax-H3 and LTX-2.5 expose but most others do not).
•
Weights and training data do not appear to be released in the supplied text; only a project page is mentioned. Treat this as a recipe paper, not a drop-in artifact.
•
The semantic-consistency reward and one of the evaluation metrics both come from LLM judges (Qwen3.7-Max for training, gemini-3.1-pro-preview for evaluation). The final arbiter is 60 human film-industry experts doing pairwise comparisons, which is the strongest signal, but the intermediate LLM-judge numbers should be read as consistent with, not independent of, the training objective.
•
“Beats commercial systems” is a full-pipeline claim. WanPE always runs on top of Wan3.0’s generator in those comparisons, so the generator’s own quality is baked into the numbers. The cleaner isolated claims are (a) WanPE vs raw prompts under Wan3.0 and (b) reverse vs forward construction.
•
Training required 512 GPUs, 1.05M curated video clips, a strong captioner, and a frontier LLM for both request reconstruction and reward scoring. This is not a small-lab reproduction.
•
The 30-second win over Seedance 2.5 is narrow (60.24 vs 59.76) and category-dependent: WanPE dominates on animation and speech but loses on action and advertisements. The headline is roughly “competitive at 30s,” not “decisively better.”