ResearchStudio-Reel: Automate the Last Mile of Research from Paper to Poster, Video, and Blog¶
🕒 Published (v1): 2026-07-05 00:00 UTC · Source: HuggingFace · link
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
💬 Ask ChatGPT✦ Ask Claude
TL;DR¶
ResearchStudio-Reel is a five-skill agentic pipeline (Paper2Assets, Paper2Poster, Paper2Video, Paper2Blog, Paper2Reel) that converts a single paper PDF into three editor-ready dissemination artifacts—print-ready conference poster, narrated talk video, and bilingual blog—sharing one upstream extraction stage. It replaces soft VLM-preference quality scores with hard categorical render gates and ships all artifacts in native authoring formats (PowerPoint, Word). On the Paper2Poster benchmark, generated posters win the overall score on 84–93% of papers against prior automated systems and single-shot frontier LLMs.
Problem¶
Three structural gaps recur across prior paper-to-X automation: (G1) isolated extraction—each system independently re-extracts figures, captions, and metadata, breaking cross-artifact factual consistency; (G2) one-way renders—artifacts ship as PDF/PNG/MP4 that authors cannot reopen and edit; and (G3) soft quality gates—continuous VLM-as-judge or aesthetic scores accept layouts that score 7.8/10 even when load-bearing sections are empty. Additionally, no prior system produces all three artifacts in a unified, cross-referenced deliverable.
Method¶
Shared extraction (Paper2Assets): One idempotent pass over the source PDF produces a stable bundle—full body text, per-figure manifest, cleaned figure images, nine-section structured summary (Problem/Motivation/Contribution/Method/Dataset/KeyResult/Ablation/HeadlineNumbers/Takeaway), metadata, institution logos (fetched from Wikimedia/Wikidata), QR codes, and a reading-order narration clip list. A visual-AI + sub-agent verifier loop handles figure cropping and splitting. All downstream skills consume this bundle verbatim and never re-open the PDF.
Measured-fill loop (Paper2Poster): The poster is assembled along four orthogonal axes (column layout, visual theme, title-band header, QR block) and grown from a lean draft via a staged-fill loop. Each iteration renders the page in headless Chromium and computes a scalar fill ratio \(r = h_\text{content}/h_\text{card}\) per section, quantized into five categorical verdicts:
| Verdict | Range | Remediation |
|---|---|---|
| EMPTY | \(r < 0.70\) | append withheld paragraph / promote optional section |
| SPARSE | \(0.70 \le r < 0.90\) | polish up: expand prose or widget |
| FULL | \(0.90 \le r \le 1.00\) | no-op |
| SPILLAGE | \(1.00 < r \le 1.10\) | polish down: tighten prose |
| OVERFLOW | \(r > 1.10\) | drop supplementary paragraph or optional section |
The loop exits at a categorical fixed point (every section FULL; every figure \(\ge 70\%\) on one axis) rather than a learned-score plateau, bounded by a circuit breaker (~12 rounds). A final render-time expand stretches under-filled cards to 98% via row whitespace growth without resizing figures. The PowerPoint export reconstructs native shapes directly from the live DOM via getBoundingClientRect/getComputedStyle → EMU coordinate mapping, preserving inline bold/italic runs, hanging-indent bullets, and print-resolution images.
Paper2Video ships an editable deck, narration-aligned MP4, burned-in and sidecar subtitles, and a shared timeline JSON that Paper2Reel navigates by section. Paper2Blog produces two .docx files (WeChat and English-research-blog registers) gated by a layout-aware DOCX check flagging near-blank pages, orphan tails, and under-filled images. Paper2Reel binds all three into a self-contained HTML viewer where section-level clicks synchronize the video, slide thumbnails, captions, and blog passage.
Key Contributions¶
- Five-skill composition sharing a single Paper2Assets extraction, eliminating per-artifact re-derivation (G1) and guaranteeing cross-artifact figure/section consistency via stable identifiers.
- All artifacts ship in native authoring tools (PowerPoint for poster and video deck, Word for blog), making every output author-editable rather than a one-way render (G2).
- Categorical measured-fill loop with hard render gates replaces soft VLM-preference plateau quality signals (G3).
- Paper2Reel unified interactive HTML surface linking poster, video, and blog at section granularity—a capability absent from all prior single-artifact systems.
- State-of-the-art poster quality: leads every aesthetic and information sub-criterion on the Paper2Poster benchmark, exceeding the authors' own posters by \(\approx 0.6/5\) on aesthetics under two held-out VLM judges.
Results¶
- Wins overall score on 84–93% of papers vs. prior automated systems (PaperAgent, PosterForest, P2P, PosterGen, etc.) and single-shot frontier LLMs on the Paper2Poster benchmark.
- Exceeds authors' own posters by \(\approx 0.6/5\) on aesthetics under two held-out VLM judges.
- Leads every aesthetic and information sub-criterion on Paper2Poster benchmark.
- Only pipeline in the capability audit to ship all three editable artifacts (poster PPTX, video deck + MP4, bilingual DOCX blog).
Limitations¶
- The fill loop requires a circuit breaker; when convergence fails, the best-measured intermediate state ships rather than a guaranteed FULL poster.
- The 90% categorical gate is deliberately loose and requires a post-hoc render-time expand to reach the visual 98% target—a two-stage workaround acknowledging that tightening the gate markedly increases round count.
- Institution and venue logo fetching is best-effort; unresolved logos silently hide the titlebar slot rather than degrade gracefully.
- Figure cleanup (the visual-AI + sub-agent verifier loop) is acknowledged as the most time-consuming stage; no throughput or latency numbers are provided in the visible text.
- Evaluation relies on the Paper2Poster benchmark for poster quality; video and blog quality are assessed only by a capability audit (which artifacts are produced), not by a comparable quantitative benchmark.
Relevance to Vision-Language Models¶
VLMs appear in this work in two roles: as held-out quality judges in the Paper2Poster evaluation (establishing that the generated posters surpass author originals under VLM aesthetics scores), and implicitly as the single-shot frontier LLM baselines the system outperforms—demonstrating that structured agentic decomposition with deterministic render gates strictly dominates monolithic VLM calls for layout-constrained artifact generation. The paper's critique of "soft VLM-preference gates" is a direct methodological claim relevant to researchers using VLMs-as-judges: continuous aesthetic scores plateau and fail to detect empty sections, arguing for categorical, render-grounded quality signals instead. The figure-cleanup stage deploys a visual-AI crop verifier, showing a concrete pattern for integrating VLM perception into a deterministic pipeline with idempotent re-run semantics.