Pavilion Tokyo 2021 / Generative AI / Data Scraping

TOKYO 2020-2021

For this project, Daito Manabe was responsible for planning, concept development, installation design, and the selection of the data and AI libraries used in the work. The installation connected anonymous public comments, the Tokyo 2020 Games Vision, GPT-2 + rinna, and VQGAN+CLIP to reconstruct Tokyo under pandemic conditions through generative AI.

2021.07.01-09.05 Watari-um open space Tokyo Tokyo FESTIVAL Special 13

TOKYO 2020-2021 installation at Pavilion Tokyo 2021 — Photo: Muryo Homma (Rhizomatiks)

Context

Within Pavilion Tokyo 2021, which proposed an urban landscape through architecture and objects, this work approached the same moment from the layers of information, data, and generative technology, addressing the social tension created by the overlap of the Tokyo Olympics and COVID-19.

The source material is built from two contrasting types of language: official Olympic vision statements and anonymous news comments. By placing Olympic slogans and massive public comment threads about the event, infection conditions, and media narratives into the same generative pipeline, the work visualized an information environment where celebration, politics, anxiety, and distrust were inseparable.

Generative AI is used not only to create finished images, but also as a medium that preserves the immaturity, repetition, failure, and risk of the models available at the time. The physical mosaic lens is part of the control system: passersby could not easily read the text, while viewers who intentionally approached the installation could decode it.

Technical Details

Based on the production notes, this section summarizes the collection scale, filter conditions, models, prompt design, and output-selection policy.

Articles

577

Selected mainly from Yahoo! News articles with large comment counts.

Comments / Replies

1.25M

Approximately 1.25 million comments and replies were collected.

Fine-tuning Corpus

9,725

Comments and replies passing a 390-character filter.

Output Selection

300+

GPT-2 outputs of 300 characters or more were primarily used.

Data Scraping

Node.js and Puppeteer were used to collect articles, comments, and replies, primarily from Yahoo! News items with roughly 1,000 or more comments. Later notes also considered selecting major topics from comment rankings so that the corpus would not be limited only to Olympic or COVID-related material.

Target: Yahoo! News comments and replies
Articles: 577
Collected scale: approximately 1.25 million comments / replies
Training data: 9,725 entries of 390 characters or more

Language Model

The Japanese GPT-2 model `japanese-gpt2-medium` from rinna was fine-tuned using Hugging Face Transformers and SentencePiece. The model absorbed the repetition, abrupt logic shifts, and collective pressure characteristic of anonymous comment threads as a generative writing style.

Model: rinna / japanese-gpt2-medium
Framework: Transformers
Tokenizer: SentencePiece
Selection: outputs of 300 characters or more

Image Generation

Image generation used the VQGAN+CLIP workflow. In this method, VQGAN generates images while OpenAI CLIP evaluates the similarity between the image and the text prompt, iteratively steering the image toward the prompt. For this work, phrases selected from the official Tokyo 2020 Games Vision were translated with Google Translate and DeepL, then converted into prompts for image generation. The prompts were prepared in English because the CLIP-based workflow available at the time generally responded more reliably to English-language prompts. The outputs were not selectively curated; failures, noise, and unstable visual logic were retained as part of the record of the model's state at that moment.

Generation model: VQGAN / CompVis taming-transformers
Text-image guidance: OpenAI CLIP
Prompt source: Tokyo 2020 Games Vision
Prompt preparation: English translation with Google Translate / DeepL
Method: iterative VQGAN latent optimization guided by CLIP score
Output policy: VQGAN+CLIP outputs were used without cherry-picking

Exhibition Control

Generative models at the time had few of the safety mechanisms that are now common, so discriminatory or aggressive sentences could appear. The installation treated the mosaic lens as a physical readability filter, allowing only viewers who deliberately approached the work to read the generated text.

Public readability: difficult to read from the street
Viewer action: readable only at close range
Role: the exhibition space functions as a filter

Generated Outputs

Examples of VQGAN+CLIP images and GPT-2 + rinna text outputs preserved in the production memo. The official Games Vision language and the style of anonymous public comments were transformed into unstable visual and textual outputs.

Image prompts Phrases selected from the Tokyo 2020 Games Vision were translated with Google Translate / DeepL and sent into VQGAN+CLIP. The generated images shown here are production-memo outputs rather than curated final stills.

VQGAN+CLIP output for Frenzied Tokyo — 熱狂する東京**Frenzied Tokyo**

VQGAN+CLIP output for The city of Tokyo is in a frenzy — 熱狂する東京**The city of Tokyo is in a frenzy.**

VQGAN+CLIP output for Sport has the power to change the world and the future — スポーツには世界と未来を変える力がある。**Sport has the power to change the world and the future.**

VQGAN+CLIP output for Sports have the power to change the world and the future — スポーツには世界と未来を変える力がある。**Sports have the power to change the world and the future.**

VQGAN+CLIP output for The most innovative convention in history — 史上最もイノベーティブで、世界にポジティブな改革をもたらす大会**The most innovative convention in history, bringing positive change to the world**

VQGAN+CLIP output for The most innovative and positive reform competition in history — 史上最もイノベーティブで、世界にポジティブな改革をもたらす大会**The most innovative and positive reform competition in history**

VQGAN+CLIP output for Personal bests for all — 全員が自己ベスト**Personal bests for all!**

VQGAN+CLIP output for Everyone is personal best — 全員が自己ベスト**Everyone is personal best**

VQGAN+CLIP output for O MO TE NA SHI — おもてなし**O MO TE NA SHI**

VQGAN+CLIP output for Proof that humans have defeated the new coronavirus — 人類が新型コロナウィルスに打ち勝った証**Proof that humans have defeated the new coronavirus.**

Generated text excerpts The excerpts below come from GPT-2 + rinna outputs fine-tuned on Yahoo! News comments. They are shown as generated artifacts, not as factual statements; their repetition, abrupt logic, and unstable tone are part of what the installation preserved. The full original Japanese outputs are available on a separate page.

Original Japanese output excerpt 01