ãŠä¹…ã—ã¶ã‚Šã§ã™ï¼ã—ã°ã‚‰ãアップãƒãƒ¼ãƒ‰ãŒå°‘ãªã‹ã£ãŸï¼ˆå®Ÿéš›ã¯1週間も経ã£ã¦ã„ã¾ã›ã‚“ãŒï¼‰ã®ã«ã¯ç†ç”±ãŒã‚りã¾ã™ã€‚
ç§ã¯1ã¤ã®ã‚·ãƒªãƒ¼ã‚ºã§æ•°ç™¾ã€œæ•°åƒæžšã®ç”»åƒã‚’生æˆã™ã‚‹ã“ã¨ãŒã‚りã¾ã™ã€‚
ã—ã‹ã—ã€ãれらをã™ã¹ã¦ç›®è¦–ã§æ•´ç†ã™ã‚‹ã¨ã‚¨ãƒ©ãƒ¼ãŒå¤šã発生ã—ã¾ã™ã€‚
「画åƒãŒå…¨ããªã„状態ã€ã‚ˆã‚Šã¯ã€ã€Œå¤šå°‘ã®ã‚¨ãƒ©ãƒ¼ãŒã‚ã£ã¦ã‚‚多ãã®ç”»åƒã‚’æä¾›ã™ã‚‹ã€æ–¹ãŒæ£ã—ã„ã¨è€ƒãˆã¦ãã¾ã—ãŸã€‚
ã¨ã¯ã„ãˆã€ã“ã®æ•´ç†ä½œæ¥ã«ã¯æƒ³åƒä»¥ä¸Šã«æ™‚é–“ãŒã‹ã‹ã‚Šã€ã‚¨ãƒ©ãƒ¼ã‚‚é¿ã‘られãªã„ãŸã‚ã€LLMを活用ã™ã‚‹æ–¹ãŒè‰¯ã„ã¨è€ƒãˆã¾ã—ãŸã€‚
ã—ã‹ã—ã€Gemini ã‚„ ChatGPT ã®ã‚ˆã†ãªå•†ç”¨ãƒ¢ãƒ‡ãƒ«ã¯ NSFW ãªç”»åƒã‚„テã‚ストã¸ã®æ¤œé–²ãŒå¼·ã„ãŸã‚ã€
ç¾åœ¨ã¯ãƒãƒ¼ã‚«ãƒ«ç’°å¢ƒã§ Qwen2.5VL ã‚„ Kimi-VL-Thinking モデルを使ã„ã€ç”»åƒã‚’æ•´ç†ã™ã‚‹æ–¹æ³•ã‚’ç ”ç©¶ã—ã¦ã„ã¾ã™ã€‚
ã¾ã æ–°ã—ã„分野(VLMï¼‰ã«æŒ‘戦ä¸ã§ä¸å…·åˆã‚‚ã‚りã¾ã™ãŒã€æ–°ã—ã„å¯èƒ½æ€§ã‚’発見ã§ãã¦å¬‰ã—ãæ€ã„ã¾ã™ã€‚
ã“れã‹ã‚‰ã‚‚より良ãã€ãã—ã¦ä¸€è²«æ€§ã®ã‚ã‚‹ç”»åƒã‚’ãŠå±Šã‘ã§ãるよã†åŠªåŠ›ã—ã¾ã™ã€‚ã„ã¤ã‚‚ã‚りãŒã¨ã†ã”ã–ã„ã¾ã™ï¼
It’s been a while! Even though it hasn’t even been a full week since my last upload, there’s a reason for the recent slowdown.
I sometimes generate hundreds to thousands of images in a single series, but manually reviewing and organizing them often results in many errors.
I’ve always believed it’s better to provide more images, even with minor mistakes, rather than have nothing at all.
However, sorting them takes more time than expected and still leads to errors, so I thought it would be better to leverage LLMs for the task.
Unfortunately, commercial models like Gemini and ChatGPT tend to have strict moderation against NSFW images and text.
For that reason, I’ve been experimenting locally with models such as Qwen2.5VL and Kimi-VL-Thinking to manage and organize my image sets.
While this new area (VLM) still comes with bugs, I’m excited to have discovered new possibilities.
I’ll continue working to deliver better and more consistent images. Thank you, as always, for your support!
好久ä¸è§ï¼è™½ç„¶è·ç¦»ä¸Šä¸€æ¬¡ä¸Šä¼ 还ä¸åˆ°ä¸€å‘¨ï¼Œä½†æœ€è¿‘æ›´æ–°é¢‘çŽ‡å˜æ…¢æ˜¯æœ‰åŽŸå› çš„ã€‚
我有时会在一个系列ä¸ç”Ÿæˆæ•°ç™¾åˆ°æ•°åƒå¼ 图片,但é€ä¸€æŸ¥çœ‹å¹¶æ•´ç†å®ƒä»¬æ—¶ï¼Œç»å¸¸ä¼šå‡ºçŽ°å¾ˆå¤šé”™è¯¯ã€‚
我一直认为,å³ä½¿æœ‰äº›è®¸é”™è¯¯ï¼Œæä¾›æ›´å¤šçš„图片也比什么都没有è¦å¥½ã€‚
ä¸è¿‡ï¼Œæ•´ç†è¿™äº›å›¾ç‰‡æ‰€èŠ±è´¹çš„æ—¶é—´æ¯”é¢„æƒ³çš„æ›´å¤šï¼Œè€Œä¸”ä»ç„¶éš¾å…出错,所以我觉得应该利用 LLM æ¥å®Œæˆè¿™é¡¹ä»»åŠ¡ã€‚
ä½†åƒ Geminiã€ChatGPT è¿™æ ·çš„å•†ç”¨æ¨¡åž‹å¯¹ NSFW 图片和文å—çš„å®¡æŸ¥æ¯”è¾ƒä¸¥æ ¼ï¼Œå› æ¤ï¼Œ
我目å‰åœ¨æœ¬åœ°ä½¿ç”¨ Qwen2.5VLã€Kimi-VL-Thinking ç‰æ¨¡åž‹ï¼ŒæŽ¢ç´¢æ•´ç†å›¾ç‰‡çš„æ–¹æ³•。
虽然这是一个新的领域(VLM),还å˜åœ¨ä¸å°‘问题,但我很高兴å‘现了新的å¯èƒ½æ€§ã€‚
我会继ç»åŠªåŠ›ï¼Œå¸¦æ¥æ›´å¥½ã€æ›´ä¸€è‡´çš„图片。感谢大家一直以æ¥çš„æ”¯æŒï¼
FullVersion 161, Cookie 256, Draft 112
https://mega.nz/file/qxYwETpD#R0gvGvt2Hq_DS1oLCif-VEaNjWbskj0fRoppxkvdTno