Adobe 的生成式 AI 平台是問題的完美精煉

去年，Adobe 多次更新了 Firefly 生成式 AI 平台，最近一次更新是在 9 月。在那段時間裡，依賴該技術的 Lightroom 和 Photoshop 工具變得越來越糟糕，系統選擇在海鷗照片上添加比特幣徽標是問題的完美昇華。

Adobe updated its Firefly generative AI platform multiple times last year, the most recent coming in September. Over that time, the Lightroom and Photoshop tools that rely on the technology have gotten steadily worse and the system’s choice to add a Bitcoin logo to a photo of a seagull is a perfect distillation of the problem.

Adobe 去年多次更新了 Firefly 生成式 AI 平台，最近一次更新是在 9 月。在那段時間裡，依賴該技術的 Lightroom 和 Photoshop 工具變得越來越糟糕，系統選擇在海鷗照片上添加比特幣徽標是問題的完美昇華。

Last week, I struggled to get any of Adobe’s generative or content-aware tools to extend a background and cover an area for a thumbnail I was working on for our YouTube channel. Previous to the updates last year, the tasks I asked Photoshop to handle were done quickly and without issue. Since, however, it’s been a rocky road.

上週，我努力使用 Adob​​e 的生成工具或內容感知工具來擴展背景並覆蓋我正在為 YouTube 頻道製作的縮圖的區域。在去年的更新之前，我要求 Photoshop 處理的任務很快就完成了，沒有任何問題。然而，從那時起，這條路就崎嶇不平。

All I was trying to do was make a little bit more room on that side of the frame so I could reposition the camera and lens Chris was using.

我想做的就是在框架的那一側留出更多的空間，這樣我就可以重新定位克里斯正在使用的相機和鏡頭。

Eventually had to resort to the old-fashioned way of doing this and manually cloned out the area to produce the thumbnail we eventually published:

最終不得不訴諸老式的方法來執行此操作並手動克隆該區域以生成我們最終發布的縮圖：

The issue with my request is apparently known by Adobe. When I reached out to the company for comment, a company representative pointed me to an article on Lightroom Queen where it says asking Generative Remove or Generative Fill to work in a space requires that the entire subject and anything related to it be selected or else it will try and replace it with something.

Adobe 顯然已經知道我的請求有問題。當我聯繫該公司徵求意見時，一位公司代表向我指出了Lightroom Queen 上的一篇文章，其中提到要求生成刪除或生成填充在空間中工作需要選擇整個主題以及與之相關的任何內容，否則它會被刪除。

“Select the entire object/person, including its shadow, reflection, and any disconnected parts (such as a hand on someone else’s shoulder). Otherwise, the AI tries to rebuild the object based on what’s left behind. For example, if you select a person and miss their feet, Lightroom tries to rebuild a new person to fit the feet,” the article reads.

「選擇整個物體/人，包括它的陰影、反射和任何斷開的部分（例如放在別人肩膀上的手）。否則，人工智慧會嘗試根據留下的內容重建物件。例如，如果您選擇一個人但錯過了他們的腳，Lightroom 會嘗試重建一個新的人以適合腳，」文章中寫道。

While this kind of makes sense if you don’t think about it too hard, it also is completely counterintuitive to the concept of the name of the tool and the result an editor is expecting.

雖然如果你不仔細思考的話這是有道理的，但它也與工具名稱的概念和編輯器期望的結果完全違反直覺。

If I am selecting a body part and asking a tool to fill or remove that space, zero percent of the time would I want it to replace my selection with its eldritch nightmare version of that exact same thing. What I, and any editor doing this, want is for what is selected to be removed as seamlessly as possible.

如果我選擇一個身體部位並要求一個工具來填充或刪除該空間，那麼我希望它用完全相同的東西的可怕噩夢版本替換我的選擇的可能性為零。我和任何這樣做的編輯都希望盡可能無縫地刪除所選內容。

Also, this method does not always work, as I demonstrated last year:

此外，正如我去年所演示的那樣，這種方法並不總是有效：

This is a repeat of the problem I showcased last fall when I pitted Apple’s Clean Up tool against Adobe Generative tools. Multiple times, Adobe’s tool wanted to add things into a shot and did so even if an entire subject was selected — which runs counter to the instructions Adobe pointed me to in the Lightroom Queen article.

這是我去年秋天將 Apple 的清理工具與 Adob​​e 生成工具進行比較時所展示的問題的重複。 Adobe 的工具多次想要將內容添加到鏡頭中，即使選擇了整個主題，也會這樣做 - 這與 Adob​​e 在 Lightroom Queen 文章中向我指出的說明背道而馳。

Adobe also pointed me to an Adobe Community post that has some tips for getting better results out of its generative tools, and while I can confirm these do help, we’re still seeing weird results even if we follow the instructions to the letter.

Adobe 也向我推薦了一篇AdAdobe 社群帖子，其中提供了一些利用其生成工具獲得更好結果的提示，雖然我可以確認這些確實有幫助，但即使我們嚴格按照說明進行操作，我們仍然會看到奇怪的結果。

This loops us back to the Bitcoin situation. Yesterday, photographer Matthew Raifman shared a bizarre result Adobe’s Generative AI produced in Lightroom. The Generative Remove tool saw a selection of a reflection and decided to replace it with a Bitcoin logo.

這讓我們回到了比特幣的情況。昨天，攝影師 Matthew Raifman 分享了 Adob​​e 的 Generative AI 在 Lightroom 中產生的奇怪結果。生成刪除工具看到了一個反射選擇，並決定將其替換為比特幣徽標。

“Adobe has officially jumped the shark. Their AI remove feature in lightroom just added a bitcoin to my gull bird in flight photo,” he shared on Bluesky. “A bitcoin!?!”

「Adobe 已正式躍躍欲試。他們在 Lightroom 中的人工智慧刪除功能剛剛為我的海鷗飛行照片添加了比特幣，」他在 Bluesky 上分享道。 “比特幣！？！”

Raifman shared a screen recording with PetaPixel that verifies this wasn’t added on purpose and was the first suggestion from Adobe’s AI.

Raifman 與 PetaPixel 分享了一段螢幕錄像，證實這不是故意添加的，並且是 Adob​​e AI 的第一個建議。

To its credit, two of the three options Generative Remove suggested did provide usable alternatives. Unfortunately, the Bitcoin option was the first one, which (whether Adobe intends this or not) tells an editor that it is what the platform feels is the best result.

值得讚揚的是，產生刪除建議的三個選項中的兩個確實提供了可用的替代方案。不幸的是，比特幣選項是第一個，它（無論 Adob​​e 是否有意）告訴編輯，這是平台認為最好的結果。

It’s not so much that Adobe’s tools don’t work well, it’s more the manner of how they’re not working well — if we weren’t trying to get work done, some of these results would be really funny. In the case of the Bitcoin thing, it just seems like it’s trying to replace the painted pixels with something similar in shape to the detected “object” the user is trying to remove. But that doesn’t make any sense in how editors would expect the tool to perform.

與其說 Adob​​e 的工具不能很好地工作，不如說是它們不能很好地工作的方式——如果我們不努力完成工作，其中一些結果將會非常有趣。就比特幣而言，它似乎試圖用與用戶試圖刪除的檢測到的「物件」形狀相似的東西來替換繪製的像素。但這對於編輯者期望該工具的執行方式沒有任何意義。

Editors don’t want something replaced with an object akin to what they select to remove, they want it replaced with what is around it. But, somehow, Adobe’s AI just isn’t coded to understand this and it repeatedly generates the weirdest stuff because of it.

編輯者不希望某些東西被類似於他們選擇刪除的物件替換，他們希望它被它周圍的東西替換。但是，不知何故，Adobe 的人工智慧的編碼無法理解這一點，因此它不斷地產生最奇怪的東西。

Generative Remove and Generative Fill have become so unreliable that some members of the PetaPixel staff have stopped using it entirely. As I pointed out, I had to go back to the manual clone stamp method to get the task I wanted completed.

產生刪除和產生填充變得非常不可靠，以至於 PetaPixel 員工的一些成員已經完全停止使用它。正如我所指出的，我必須回到手動克隆印章方法才能完成我想要的任務。

“Overall, Adobe is aware and actively working to resolve,” Adobe tells PetaPixel.

Adobe 告訴 PetaPixel：“總的來說，Adobe 已經意識到並積極努力解決這個問題。”

When Adobe is pushing AI as the biggest value proposition in its updates

當 Adob​​e 將人工智慧作為其更新中最大的價值主張時