Copy link to clipboard
Copied
With the recent addition of flagging all images that use Generative Fill, a wide range of questions has come up. I have searched the forum for answers but have yet to find solid information. I understand this is a new feature, so definitive answers may be challenging to provide. However, any clarification on the questions below would be helpful!
Metadata Display of AI Use:
Distinguishing Between Uses:
Testing and Deletion:
Content Credentials Display:
Copyright Considerations:
7.
List of Tools that flag as AI use
Copy link to clipboard
Copied
This post is a discussion in the user-to-user forum, so I doubt you will receive any reply from Adobe here.
I have examined the RAW data in the metadata UI but cannot find any indication of AI use. If the metadata is affected:
- What is it called?
- Where is it shown?
- By @J.C.C.1
File > File Info > Raw Data isn't an exhaustive list of metadata, it's only what Adobe has decided to make available... Using ExifTool, there appears to be a new CBOR group:
[CBOR] ActionsAction : c2pa.edited
[CBOR] ActionsSoftwareAgent : Adobe Firefly
[CBOR] ActionsDigitalSourceType : http://cv.iptc.org/newscodes/digitalsourcetype/compositeWithTrainedAlgorithmicMedia
[CBOR] ExclusionsStart : 3164
[CBOR] ExclusionsLength : 11448
[CBOR] Name : jumbf manifest
[CBOR] Alg : sha256
[CBOR] Hash : (Binary data 32 bytes, use -b option to extract)
[CBOR] Pad : (Binary data 9 bytes, use -b option to extract)
[CBOR] Title : Generated Image
[CBOR] Format : image/jpeg
[CBOR] InstanceID : xmp:iid:<UUID code here>
[CBOR] Claim_generator : Adobe Photoshop/25.7.0 adobe_c2pa/0.7.6-2 c2pa-rs/0.25.2-patch
[CBOR] Claim_Generator_InfoName : Adobe Photoshop
[CBOR] Claim_Generator_InfoVersion : 25.7.0####
[CBOR] Claim_Generator_InfoComAdobeBuild: 20240415.r.504 130b225 mac
[CBOR] Signature : self#jumbf=c2pa.signature
[CBOR] AssertionsUrl : self#jumbf=c2pa.assertions/c2pa.actions, self#jumbf=c2pa.assertions/c2pa.hash.data
[CBOR] AssertionsHash : (Binary data 32 bytes, use -b option to extract), (Binary data 32 bytes, use -b option to extract)
[CBOR] Alg : sha256
[CBOR] Item0 : (Binary data 3266 bytes, use -b option to extract)
[CBOR] Item1SigTstTstTokensVal : (Binary data 3639 bytes, use -b option to extract)
[CBOR] Item1Pad : (Binary data 3043 bytes, use -b option to extract)
[CBOR] Item2 : null
[CBOR] Item3 : (Binary data 256 bytes, use -b option to extract)
I'm not sure if/how this relates to another set of metadata (JUMBF) which appears to be related to JPEG2000:
[JUMBF] JUMDType : (c2pa)-<UUID code here>
[JUMBF] JUMDLabel : c2pa
[JUMBF] JUMDType : (c2ma)-<UUID code here>
[JUMBF] JUMDLabel : adobe:urn:uuid:<different UUID code here>
[JUMBF] JUMDType : (c2as)-<UUID code here>
[JUMBF] JUMDLabel : c2pa.assertions
[JUMBF] JUMDType : (cbor)-<UUID code here>
[JUMBF] JUMDLabel : c2pa.actions
[JUMBF] JUMDType : (cbor)-<UUID code here>
[JUMBF] JUMDLabel : c2pa.hash.data
[JUMBF] JUMDType : (c2cl)-<UUID code here>
[JUMBF] JUMDLabel : c2pa.claim
[JUMBF] JUMDType : (c2cs)-<UUID code here>
[JUMBF] JUMDLabel : c2pa.signature
You can head over to the ExifTool Forum and search and or ask for more about CBOR.
Copy link to clipboard
Copied
Awesome, Thanks for the insight!
Copy link to clipboard
Copied
Through some more in depth testing, I do have a CBOR group, and I do have mentions of Adobe Firefly.
I am using exiftool-12.87 on a Windows Platform, drag and drop on exiftool(-k).exe, using it on tif's and jpg's
Photoshop Version is: 25.9.1
Only Jpgs's show Firefly data for me at the moment.
Looking at https://contentcredentials.org/verify as well, the metadata does not affect Content Credentials through adobe, just when Content Credentials (Beta) is enabled
Copy link to clipboard
Copied
If someone uses generative fill and doesn't want it to be known, that's probably an indication that they shouldn't be using it, purely from an ethical point of view.
By nature, generative fill is for making content. For minor dust-ups, there is very little point in using it. The traditional tools will be more than sufficient - and probably faster, more effective, and with a better final result.
Personally, I'm not touching it. The borderline case is the remove tool. I do use that, but sparingly and only when absolutely necessary for reasons I couldn't control when shooting.
Copy link to clipboard
Copied
Totally agree, I think if you use it, it should be shown. Just a lot of ambiguity and potential for updates to copyright issues going forward so I am trying to ensure I am ahead of the ball when changes do come. There are definitely some situations where the gen fill works wonders for standard retouching, but for the most part other tools are faster and much more efficient .
Copy link to clipboard
Copied
I use it extensively for removing small items from the backgrounds of photoshoots where other Photoshop tools are inadequate. That is 99% of my use of it. I only use it where I have to because it is so much slower than the other tools, and the bigger problem is that I edit photos for Play boy and Firefly mostly won't edit photos that have scantily-clad females in them due to Adobe's moral stance.
Copy link to clipboard
Copied
I believe this topic is significantly aggrivated because of the "fuzzy" definition of AI; there are too many very diverse different techniques and approaches that are all given the name "AI".
Again to me: the differentiator is the data used for the immage. If new data is added I call it "generative AI". However, if it is using only original data, but manipulating it in different ways [i.e., doing what you could do in traditional editing], I call it "Large-Language-Model (LLM) AI". There is a world of difference in these two approaches and I feel LLM should not be held to the same, or simular, judgement criteria as "generative AI".
This issue will not go away until a common, discriminating naming concept is uniformly applied to these two approaches!
In the mean time, I am applying only LLM AI approaches to my images.
Vince Higbee