Academic CV in Crisis: The Debate After GPT-6

Episode 1 · Michael Black Questions the Role of Academic CV Research in the Era of Large Models (2026-09-09, 7 posts)

Michael J. Black, a veteran computer vision researcher at the Max Planck Institute who has attended ECCV since 1992, posted a long essay before ECCV 2026 questioning the role of academic CV research in the era of large models, drawing responses from peers including MIT professor Phillip Isola.

Confirmed

• Black's paper this year, VIGA, uses an agentic approach to convert images into 3D Blender scenes (a Vision-as-Inverse… idea). It was rejected on first submission and, after delayed publication, was already surpassed by stronger models.
• Black's central questions: with models growing ever more capable, what is the point of academic CV research? Is GPT-6 "Astra" a step change? How can researchers have impact in academia today?
• Phillip Isola responded that one direction is to do less "task solving" and more "understanding solutions": he argued a paper analyzing how Astra plus open-source models work could have greater value.
• In a follow-up, Black elaborated that academia's inability to evaluate top-tier models is a key problem requiring coordination among model companies, governments, and academic leaders.

Why it matters

• The discussion reflects the shock that rapidly iterating foundation models are delivering to traditional academic research: papers are obsolete on arrival, and academia's comparative advantage is shifting from building models to analysis and understanding—even evaluation itself is becoming a task academia cannot shoulder alone.
• Black and Isola represent two voices—the reflector and the path-giver—offering young researchers a framework for choosing research directions in the era of large models.

• CV veteran Michael Black: ECCV papers are two years stale — academia must rethink how it works — Michael_J_Black · 2026-09-09
• ECCV veteran asks what academic CV research means in the age of frontier models — CharlotteHase · 2026-09-09
• MPI's Michael Black: Learning About VIGA at ECCV 2026 Means You're Late to the Party — Michael_J_Black · 2026-09-09
• Phillip Isola: Study How Big Models Solve Tasks, Don't Race Them on New Benchmarks — phillip_isola · 2026-09-09
• Michael Black on academic CV research's role in the age of powerful large models — CSProfKGD · 2026-09-10
• What is academic computer vision research for in the age of large models? ECCV veterans weigh in — prof_kamilov · 2026-09-10
• Michael Black on academia's role in the era of powerful large models — Michael_J_Black · 2026-09-10

Episode 2 · Post-GPT-6 Gloom at ECCV: CV Papers Already Two Years Behind (2026-09-17, 6 posts)

Since GPT-6's release, a palpable sense of frustration has spread through the computer vision academic community. Letian Wang posted on X that several senior scholars attending CVPR — including researchers who have worked in the field since the 1990s, such as Michael Black — expressed considerable disappointment, and he only understood why after attending ECCV, where attendees openly discussed how much of traditional CV research would be absorbed by large-scale foundation models. The frustration at ECCV was even more visible than at CVPR. The take resonated widely, drawing extensive reposts and discussion.

Confirmed

• Letian Wang relayed that multiple senior CVPR/ECCV attendees felt disappointed, with the mood heavier at ECCV, against the backdrop of GPT-6's release.
• Michael Black (Director at Max Planck) said at ECCV that the papers being presented today are likely already two years behind — and two years is enough to render research meaningless.
• Black cited his own VIGA paper as an example: VIGA (Vision-as-Inverse-Graphics Agent), released by UC Berkeley and other institutions, is a multimodal agent that takes a single image as input and, through an analysis-by-synthesis loop with interleaved multimodal reasoning and an evolving contextual memory, generates 3D scenes in Blender via vibe coding — described as the first system to solve this problem.
• The VIGA authors reflected that academia lacks the compute to benchmark top-tier models; their ECCV paper is already two years out of date, with their method overtaken by a Claude Code-based solution.

Why it matters

• This is not an isolated sentiment but the CV community's second upheaval in fifteen years (deep learning previously upended traditional methods once before), touching fundamental questions about the positioning, funding, and compute allocation of academic research.
• The VIGA case concretely demonstrates how quickly academic work can be overtaken by general-purpose large models/agents, providing empirical backing for the claim that "papers become obsolete in two years."

• ECCV seethes as GenCeption shows video models can swallow traditional CV research — CSProfKGD · 2026-09-17
• CV researchers frustrated after CVPR: foundation models threaten to absorb the field — IgorCarron · 2026-09-17
• ECCV attendees voice frustration: will foundation models absorb traditional CV research? — broodsugar · 2026-09-18
• Michael Black at ECCV: CV papers are already two years behind frontier models — FinanceYF5 · 2026-09-18
• ECCV papers are two years stale; VIGA overtaken by Claude Code solutions — FinanceYF5 · 2026-09-18
• VIGA agent vibe-codes editable 3D scenes in Blender as author warns academic CV research lags 2 years — FinanceYF5 · 2026-09-18

Episode 3 · VIGA thread: exhaust LLMs first; paper cycles lag behind models (2026-09-18, 2 posts)

The VIGA thread argues researchers should exhaust existing LLMs and analyze their failures before proposing new methods, with insights that survive the next model generation. Black also notes the 2-year publication cycle means papers may rest on knowledge already obsolete, as LLMs overtake traditional problems mid-publication.

• Academic papers ship two-year-old ideas, already superseded by LLMs — FinanceYF5 · 2026-09-18
• Start research by exhausting frontier models first, only then claim a durable insight — FinanceYF5 · 2026-09-18

Episode 4 · Scholars argue research should first stress-test existing LLMs (2026-09-18, 2 posts)

Michael Black argues that research projects should first push existing LLMs to their limits and analyze failures before proposing new methods. He also criticizes review culture's obsession with technical novelty and suggests papers add a section evaluating current LLMs on the task.

• Scholar urges academics to benchmark frontier models first before proposing new methods — FinanceYF5 · 2026-09-18
• Papers should add a section benchmarking frontier models, argues VIGA thread on broken peer review — FinanceYF5 · 2026-09-18