Image Fallback Sourcing
The 6-Tier Chain (always run top-to-bottom, never skip)
Tier 1 — Polinations.ai (free, no key)
GET https://image.pollinations.ai/prompt/{urlencoded_prompt}?width=1024&height=1024&seed=42&nologo=true
- ⚠️ As of 2026-06: returns HTTP 402 (USDC paywall). Still try — sometimes the legacy endpoint bypasses.
- If 402, immediately drop to Tier 2.
Tier 2 — Stable Horde (distributed free SD, no key)
POST https://stablehorde.net/api/v2/generate/async
Header: apikey: 0000000000 (any string works)
Body: {"prompt":"...","params":{"width":512,"height":512,"steps":25,"sampler_name":"k_euler_a","n":1},"models":["Edge Of Realism"],"r2":true}
Then poll GET /generate/status/{id} every 20s.
⛔ CRITICAL GOTCHA: ~30% of workers return 92-byte FAKE images to scam kudos. Verify before accepting:
- File size MUST be > 30 KB
- First 4 bytes MUST be PNG header
89 50 4E 47(not86 db 69 b3= random noise)
Best models (low queue as of 2026-06):
Edge Of Realism(7 workers, queue=0) — photorealistic, top pickmajicMIX realistic(7 workers, queue=0)AbsoluteReality(queue=22M — too slow, skip)
Check live: GET /api/v2/status/models?type=image → pick queued < 10000.
See references/stable-horde-fake-workers.md for the full verification pattern.
Tier 3 — HuggingFace Inference API (free, no key for some models)
POST https://api-inference.huggingface.co/models/{model_id}
Body: {"inputs":"...","options":{"wait_for_model":true,"use_cache":false}}
- ⚠️ First request to a model takes 1-3 min cold start. Use
--max-time 180in curl. - ⚠️ Python urllib on this machine throws SSL EOF errors. Use curl, not Python.
- Realistic models to try:
dreamlike-art/dreamlike-photoreal-2.0,stabilityai/stable-diffusion-2-1-base,prompthero/openjourney
Tier 4 — Pexels CC0 (real photos, real but pay attention)
- Photo CDN 直链(curl 通):
GET https://images.pexels.com/photos/{photo_id}/pexels-photo-{photo_id}.jpeg?w=1200- ✅ 不限 IP、不限 UA、不限频、不需要 Key
- 必须用已知 photo ID(见
references/pexels-photo-id-bank.md)
- ⚠️ Pexels 搜索页 HTML 已挂(2026-06-09 实测:Cloudflare 验证页拦截,沙盒浏览器都过不去)
- 不要再尝试爬 https://www.pexels.com/search/
- 想"找"新 ID → 走 Tier 7 Bing 抓 murl
- Pexels 搜索/英文关键词对"设备/养殖/中文食材"基本 0 结果
Tier 5 — Unsplash Source API
GET https://source.unsplash.com/featured/?{keyword}- ⛔ As of 2026-06, returns 503 on most queries. Skip unless desperate.
Tier 7 (added 2026-06-09): Bing 图片搜索 murl + PIL 底部裁水印
当 Pexels 401 / Unsplash 503 / Wikimedia SSL 失败时的兜底。中文搜索 ~70% 结果来自
nipic/nximg.cn(带水印),用黑名单过滤 + PIL 切底部 14% 解决。完整脚本 + 关键坑见
references/bing-pil-watermark-removal.md。
关键坑:
- Bing JSON 里
"被转义为",正则必须用murl":"..." - 黑名单必须含
nximg.cn(nipic 的 CDN 域名,常见陷阱) - 永远
vision_analyze抽查 2-3 张再批量裁,不要相信文件大小
Tier 6 — vision_analyze verification (mandatory)
- After ANY download from Tier 1-5, always verify with
vision_analyzeto confirm the image matches the request. - Pexels / Unsplash / search engines often return wrong images. Trust nothing without verification.
- If wrong, try next ID from the bank. If all wrong, stop and report (don't keep burning time).
Tier 7 — Bing 图片搜索 HTML scrape(2026-06-09 新增,实际救场)
- URL:
https://www.bing.com/images/search?q={query}&form=HDRSC2 - ⛔ 正则坑:Bing HTML 把
"转义成",必须用这个正则:
而不是常见的urls = re.findall(r'murl":"(https?://[^&]+)"', html)r'"murl":"(https?://[^"]+)"'(那个匹配 0 个) - ✅ 中文关键词命中率 100%(35/35 murl),英文关键词基本 0
- ✅ 国内 CDN 实测通:699pic/126.net/sinaimg/alicdn/baike/baidu 都返真图
- ⚠️ 水印雷区:09_蛋白分离器 抓出"禅"字书法 + 渔帮手 logo;06 抓出沙漠岩石;10 抓出汽修工人
- ⚠️ 每张必须 vision 验图(不验=50% 翻车)
- 跑法:抓 4-6 个候选,>80KB 保留,验图不对就跳下一个
Workflow pseudocode
# Phase 1: AI gen (1-3 tiers in parallel, 60s timeout each)
for tier in [pollinations, stable_horde, hf]:
img = try_generate(tier, prompt, timeout=60)
if img and verify_is_real_image(img): # size>30KB, PNG/JPEG magic bytes
return img
# Phase 2: Real photo fallback (Tier 4)
for photo_id in curated_ids_for_category(prompt):
img = curl_download(pexels_url(photo_id))
if vision_confirms_match(img, prompt): # vision_analyze
return img
# Phase 3: Honest report
return "I tried 6 tiers. All blocked or returned wrong images. Please drop a real photo into [target path]."
Output to user
On success: ✅ Got image via [tier] from [source]. Path: [full path]. (1 line, no extras)
On partial: Got N candidates, [M] verified by vision as matching. Using [ID] at [path].
On failure: All 6 tiers failed. Last option: please drop a real photo at [path] or paste any image URL. (5-second fix, I take it from there.)
NEVER ask "which tier do you want me to try" or "should I use AI gen or stock photo" — that's the user's job to read the report, not decide infrastructure.
User autonomy pattern (CRITICAL — 2026-06-08 signal)
When user says any of:
- 「自己想办法」
- 「看着办」
- 「我只要结果」
- 「不可以再问我」
Run the full chain silently and report only the final outcome. Do not:
- List A/B/C options
- Ask "should I use AI or stock photo"
- Ask "what category of image"
- Show the tier-by-tier progress
The pattern fires repeatedly. Future sessions will see it again. Capture the "fully autonomous output" report as the only acceptable response shape:
✅ Got [N] images. Saved to: [path1], [path2], ...
or
❌ All tiers failed. Please drop a real photo at [path]. (5-second fix.)
Both formats are < 2 lines. No explanation, no options, no "what do you want to do next."
Key pitfalls
- Python urllib SSL is broken on this machine — most HTTPS calls fail with EOF. Use curl.
- Stable Horde workers are NOT trustworthy — verify file size AND PNG header bytes.
- Pexels "shrimp" search returns city models — use curated IDs, never search.
- Pexels search HTML is Cloudflare-blocked (2026-06-09) — don't try to scrape, use Bing instead.
- Bing HTML escapes
"as"— regex must bemurl":"...", not"murl":"...". - 设备/养殖/中文食材必须中文关键词(英文 0 命中,中文 35/35)。
- 国内 CDN 常带水印(商家 logo/书法站/装饰画)—— 抓完每张必 vision 验图。
- HF first request is slow (1-3 min) — set
--max-time 180and don't panic. - Vision verify is not optional — even "looks right" Pexels photos can be wrong (mosques, sushi, fried chicken all returned for "boiled shrimp" search).
- Don't ask the user for images — they will say "自己想办法". Use the fallback chain.
Verified photo ID banks (Pexels CC0)
references/pexels-photo-id-bank.md— base curated list (verified via vision, not search results), focused on shrimp/seafood.references/pexels-cc0-id-bank-food-equipment.md— water/food/equipment-aquaculture 8-image kit (5 美食: 白灼虾/堆虾/海鲜大餐/寿司/树桩料理; 3 设备/场景: 塔吉锅/鱼缸/工业罐). Reuse strategy covers 10 笔记 × 4 张 = 40 张配图.references/pexels-known-bad-ids.md— 23 IDs vision-verified as wrong (city models, fried chicken, red rocks, bathroom brush, goldens, ...). ~80% miss rate. Always vision-verify before reusing.scripts/fetch_pexels.sh— bash one-liner to batch-download any IDs with progress + size info. Usage:bash fetch_pexels.sh /path/to/out 725992 566344 566345.scripts/fetch_image.py— full 6-tier chain implementation in one script (Tier 1 pollinations / Tier 2 stable_horde with size+header check / Tier 3 HF / Tier 4 Pexels CDN / Tier 5 Unsplash / Tier 6 vision verify). Usage:python fetch_image.py "boiled red shrimp" 03.jpg.
Stable Horde verification patterns
references/stable-horde-fake-workers.md— exact size + PNG header check code (the canonical recipe).references/bing-pil-watermark-removal.md— Tier 7: Bing murl 提取正则 + 黑名单图库站 + PIL 裁水印 + Windowsos.startfile兜底开文件夹。references/stable-horde-hack-workers.md— the Zikeri-class hack node phenomenon (30-50s fast-done + <200B base64 + no PNG header), the legit-worker names (smellycat1,Roaring 3050series,Gravitate7706 Dreamer), and defensive code.