theodoruszq/imagegen-tracker：追踪 GPT Image、Nano Banana、Midjourney 等 AI 图像生成模型动态

🌐 网站 / Site: https://theodoruszq.github.io/imagegen-tracker/ · News · Demos

This repo tracks (A) releases and major updates of the top AI image-generation models: GPT Image / ChatGPT Images, Nano Banana / Gemini / Imagen, Midjourney, FLUX, Seedream, Grok Imagine, Ideogram, Recraft, Qwen-Image, Hunyuan/Hy Image, Kling/Kolors, and any new SOTA contender. It also collects (B) standout showcase posts from X, preferably ones that share their prompts. GitHub Pages serves it from main:/docs.

Site

A single page (docs/index.html) with two centered tabs:

News (#news): model releases and updates, newest first. Each item has a bullet summary in the original post's language, with key facts in bold, and labelled source links, 官方 first and then 社区. Older verified milestones from data/timeline.json are merged in.
Demos (#demos, #demos/<model> for a filter): a compact masonry wall of ~220px columns (2 on mobile, up to 6 on desktop) with model filter chips. It renders in batches of 30 as you scroll. Clicking a card opens a lightbox (Esc closes it, ← → navigate, swipe works on mobile) showing the full-size image(s), model/maker, author, date, the full prompt with a copy button, and the original post link. #<uid>, e.g. #x-2107671007029076318, is the permalink that opens it.
Demo data comes from one compact file, docs/data.json. The wall loads only ~400px thumbnails with width/height attributes and loading="lazy"; the full 1600px webp loads only in the lightbox.
English help is copied from x-digest: one remembered 中文 switch shows inline Chinese glosses for harder English terms. Visual feel is also borrowed from theodoruszq/x-digest: its color palette with light/dark theme and its Lato + PingFang font stack (Lato woff2 is vendored in the repo). The layout is the site's own. No CDNs and no external requests.

Layout

data/items.jsonl      one JSON object per line: source of truth + dedup index (schema below)
data/timeline.json    verified historical milestones [{date, model, maker, change, lang, glossary?, source_url, verification}]; `change` is written in the source's language
data/state.json       {last_run_utc, ..., skipped_uids}; last_run_utc = since-time for the next run
docs/index.html       GENERATED (News + Demos tabs)
docs/data.json        GENERATED compact demo data for the wall/lightbox
docs/demos.html       GENERATED redirect to ./#demos
docs/assets/          GENERATED copy of scripts/static/ (css, js, fonts, icon)
docs/images_blobs/<h[:2]>/<h>.webp          full images (max 1600px wide), stored exactly once
docs/images_blobs/thumbs/<h[:2]>/<h>.webp   thumbnails (max 400px wide), keyed by the same h
images_blobs/README.md                      pointer only (no bytes at repo root)
scripts/make_blobs.py  originals -> full webp + thumb (needs Pillow)
scripts/build.py       data/ -> docs/ (stdlib only, no network, idempotent; never touches images_blobs or .nojekyll)
scripts/static/        css / js / fonts / icon sources

UID rules (no duplicates)

Demos: x-<post id>, using the original post's id (never a repost).
News: news-<maker>-<model>-<version>, lowercase with hyphens (e.g. news-openai-gpt-image-2, news-google-nano-banana-2-1). Several posts about one release collapse into one item; add them to sources instead of creating a new item.
Before adding, check data/items.jsonl and data/state.json.skipped_uids. Skip reposts and near-duplicate posts of the same images. build.py stops with an error on a duplicate or malformed uid.

items.jsonl schema

Common: uid, type (news|demo), date (YYYY-MM-DD, UTC), model, maker, author (X handle, no @), source_url (primary/official), source_urls [all], title, image_hashes [sha256], archived_run.

News only (no prompt field):

title: a headline in the original post's language.
lang: the source language (en, zh, ...).
bullets [str]: a complete, faithful summary of the original post(s) in the original post's language (English source → English, Chinese source → Chinese; never auto-translate), with wording kept close to the original. Cover what's new, capabilities, specs and parameters, pricing, availability and channels, license and open weights, and links. Bold the key facts with **...** (price, speed, quality/benchmark gains, resolution, step counts, availability); ==...== also works for a highlight. The build HTML-escapes everything else. Anything not confirmed by the official source is marked 未经官方确认 in every language, for consistency with the Chinese UI.
sources [{url, label: 官方|社区, by}]: official sources first.
notes_internal: not shown on the site.

Demos only:

prompt (str|null): the full prompt; null is shown as 不详.
prompt_source: e.g. the author's self-reply URL; kept out of prompt so 复制 copies a clean prompt.
prompt_status (found / partial ... / not found).
text: the original post text.
lang / prompt_lang: the language of the post / prompt (en, zh, ...).
metrics, created_at, summary: optional.

English help: inline Chinese glosses (copied from x-digest)

The page stays English-first, with no full translations; do not write *_zh translation fields.

glossary [{term, gloss}] on any item with English content: 3–10 harder English words or phrases from the title, bullets, post text or prompt.
- term must be the exact surface form as it appears in the text.
- gloss uses x-digest's format: part of speech plus a short Chinese meaning, e.g. {"term": "artifact buildup", "gloss": "n. 伪影累积"}. Allowed prefixes are n., v., adj., adv., phr., idiom.
- Skip trivial words and product names.
The build marks the first occurrence of each term per item with x-digest's markup: termn. 中文.
Glosses are hidden by default. The one header switch 中文 (x-digest's .zh-switch, remembered in localStorage as igt.zh = '1') adds html.show-zh, which underlines each term and shows its gloss inline.
There are no hover popovers (same as x-digest), and no runtime translation API.

Image policy

Download the original media (pbs.twimg.com/...?name=orig) and run python3 scripts/make_blobs.py <files>:

Full image: WebP, max 1600px wide, quality 85. h = sha256(final full webp bytes), saved to docs/images_blobs/<h[:2]>/<h>.webp.
Thumbnail: WebP, max 400px wide, quality 80, saved to docs/images_blobs/thumbs/<h[:2]>/<h>.webp.
Identical images collapse automatically. Videos are not stored (link only); original JPG/MP4 files are never committed.
Keep it tasteful: leave out demos whose images are suggestive beyond normal fashion photography, and add their uid to skipped_uids.

Build

python3 scripts/build.py                  # docs/index.html, docs/data.json, docs/assets/
python3 -m http.server -d docs 8000       # preview at http://localhost:8000/#demos

Cadence

A routine runs twice daily at 09:03 and 21:03 Asia/Shanghai. Each run searches X since last_run_utc with roughly 40–60 read calls (cap 1000/day), adds the new news items and the 3–8 best demos, rebuilds, and pushes one commit.