Hi gang;
I've been using Adobe Firefly for about two weeks now and I'm quite excited about the prospect of it.
Prior to this I was fairly involved with MidJourney, especially from its initial release. You can see my AI work (predominantly created with MidJourney) here:
[Weblink removed as per forum guidelines]
I will be adding my Firefly work to it day by day.
I thought I would offer some of my insight to the Adobe team regarding my initial experiences so far. I want to preface this by saying I am not partial to any particular software. As someone who has been working in the computer graphics and animation industry for close to 30 years with all types of apps, I love see these wonderful technological advances and how they can be creatively combined, regardless of who makes what. The only reason I mention MidJourney is because there isn't much else to compare to, aside from Dall-E.
- I'll start by saying that I find Firefly's beta commercial work restriction quite foolish. If you want to properly beta test software, there is no better way to do it than in a production environment. That is where you truly find issues that you won't find in any other environment, including time-sensitive matters, compatibility constraints, file management concerns, quality nuances, and more. Simply "playing" with it isn't really going to reveal many important flaws like it would in a production environment. I've seen this many, many times occur in my other software beta tests. However, that is Adobe's decision and I respect it.
- I'll start with the most important feature: output. Although it's a great start, Firefly still has a long way to go. This also seems to be the general concensus among various other beta testers in my LinkedIn social circle. The output simply doesn't yet compare to MidJourney and although I don't like making comparisons, it needs to in order to make it viable for production purposes. Almost all of my prompts were set on the photo filter tag, and even included a "Photorealistic" text term, but they all end up looking like artwork (for the most part). Many pieces lack fidelity while others include a lot of distortion in areas it shouldn't (ie. inorganic shapes like tables, etc). Others still are heavily pixelated. (I have all these examples, btw). I realize MidJourney has several years of development over Firefly so I understand it's early in the game but I encourage you to pay close attention to this. Perhaps this is also an issue of the training data. I understand Firefly is trained off the Adobe Stock database which I assume is much smaller than the training data MidJourney had. Perhaps this is also due to Adobe's presumably more responsible approach to trainining, which also limits the training data. Whatever the reason, it's simply not quite there yet.
- One thing I found to be really well done with Firefly was the staging. All my subjects were correctly staged within the frame. I had a lot of trouble with this using the initial MidJourney where the subjects would often be cut off the sides of the frame. Perhaps this has already been correctly in MJ v5 but I don't know.
- In my opinion, there are pros and cons to the way in which the interface was designed. For example, Firefly provides a dropdown featuring various presets. This is great for beginners and hobbyists but a professional will need much more control than that. Why not provide the ability to specify any aspect ratio numerically, for example? Yes, you've specified many of the most common but many of us work in cinemascope, or other less conventional aspect ratios. Same with color tone, lighting and composition. Those dropdowns are more apt to being text prompts rather than presets.
- The "Show Similar" button is really helpful. Similar to the MidJourney "V" button. Although in MidJourney I believe it regenerates all 4 images, I prefer what Adobe has done which is to keep the one you like, and regenerate the other three.
- It would be nice if Firefly paid attention to prompt color specifications. For example, typing in this prompt: gorilla vector using hexadecimal colors 0f7814 and 6a0f78 doesn't result in any of those colors being used.
- One feature I have been eagerly waiting for is alpha channel / transparency support. This is really, really important. I see it's been implemented in the Text Effects section (which is great!) but I also encourage you to add it to the "Text To Image" section. For example, one might use Firefly to generate specific elements to bring into After Effects or Photoshop for use in a larger composition or video. These will require alpha channels / transparency in order to be used effectively. I'm sure sooner or later Firefly will have the ability to export as a layered PSD which will solve that issue bit still... forward thinking.
- Speaking of "Text To Image". I didn't use it much because although I think it's got enormous potential, much of the generated content was simply cut-out by the letters, no different than if I used a mask to reveal an image. Yes, some graphical elements creep outside of the letters occasionally, which is what supposedly makes this tool unique but for the most part, there is little of it. The letter "i", in particular, is a great example of how it almost always just behaves like a cut-out mask, instead of having the graphical elements creep outside of the edges and reshape the "i" into something more interesting. In many of my tests there was also quite a bit of artefacting in what should have been fully transparent areas.
- I found that Firefly was definitely better at some things over others. For example, photorealistic vehicles, environments, etc, simply didn't work well. However, stylized cartoon characters were great great (see below). This suggests to me the training was more biased to those kinds of datasets but what do I know... At the end of the day, you want a tool that excels at all types of imagery.
- I really like the fact that any regenerated thumbnails tend to use the same palette, style, etc, allowing one to use various of them in a consistent set. I assume this is done by using a similar seed for each user session. If one wants a radical change, how would one do that? Perhaps actually providing a seed parameter might be interesting.
- The interface could use some love. It feels a bit clunky & amateur. It could use a designer's touch, as well as a darker interface color (or the ability to specify one).
- Firefly seems to generate high resolution imagery much faster than MidJourney. This is a plus for sure. Not sure why or how, but it's a welcome benefit.
- Oh, and nice work on hands. Firefly seems to be capable of doing a reasonable job, given the known difficulties in generting those!
Anyway, I hope some of this feedback is helpful. I'll try to add more thoughts and insight as I continue to work with it, and I'll post more content as well.
Here's some select images from my Firefly experiments (from many, many more I've generated in my tests). These are some of the better results.
Regards,
-Richard
2
Replies
AdChoices