P: Maintaining a scene or character across different prompts

I appreciate your appreciation! 🙂 

Let's face it, generative AI is still relatively uncharted territory. It's been an "on the job" experience to say the least. 

In terms of prompts, I'm not exactly sure if having more or less information for the tool to bite onto is positive or negative. In my own expereience, and in also using other generative AI image tools outside of Firefly, the more detailed information you can feed it the better. This is why I've developed my own prompting convention to help keep this stuff approachable and manageable, otherwise it gets thick into the weeds really quickly. 

I have some possibly interesting takeaways to share after having just completed my last frame of the script, which was a hurclean feat and a half.  As follows: 

• Generative AI text-to-image (or video) is NOT ready for prime time in the context of multiframe narrative where you need the same character throughout the storyline. No jobs are getting replaced here. The toolset will get there within the next year or two, especially if there's a ton of demand for that out there, but as it stands today you'd better have a full body of hair on your head because half of it is getting pulled out and the other half is going grey. True story. 
  
  
• Firefly's "Style Reference" feature is a great bonus in terms of getting similar results, BUT it most definitely needs refinement, such as locking or resusing the same seed generation, which would go a long way to help ensuring you get the desired result. Again, if used, ALL characters in your scene get the same likeness. Everyone looks like a sibling. It's currently working in a pinch to varied results, but it feels rudimentry, or too "user-friendly" as it stands at the moment. Adobe, give us a little autonomy here. 
• Maybe this is a placebo due to starring at the screen for rash of uninterprupted consecutive hours, but as I went along in my generation journey of my character generation results seemed to get progressively worse. My results started to get progressively "muddier". I'm not sure how Firefly is working under the hood, and I doubt Adobe will let us take a peek, but resetting the UI seemed to bring it back to results I would expect. 

• Hands. Lord help us all with AI hands. Pick your poisonous generation tool, they all suffer from this and it will be the bane of my AI existence until it's reliably addressed. 

• Production time: Thought you were going to save time=money, did you? Ha! Sure, I didn't have to illustrate each frame which obviously DOES save time=money in the grand scheme of things, but there's plenty of time=money in having to sit there and massage the deep learning models to behave itself like an onry toddler. There's hope here; As the models data mine more information the equation will eventually be, "understanding=awesome sauce". 

Cheers, and happy holidays, you crazy creative cloud kids! 

Matt,

Thank you so much. I (and I'm sure others) are so appreciative of you sharing this with us. We're all trying to get a grip on what prompts deliver what we want and this is one of the longer prompts I've seen. I wasn't even aware it would generate an image with this much info so I think we all just learned something.

I've mentioned a number of times that I'm trying to create a children's book my daughter wrote and until I can generate consistency of characters between scenes, I can't begin the process. To date this is the closest I've seen.

Again thank you for your kindness and willingness to share this important prompt.

All Firefly, using the Style Reference feature, which is crucial to teaching the model that you want "x" character's likeness in each generation. 

The main qualifying prompt is: 
"a 40-year-old woman with tan skin, medium-length black wavy hair parted to the left, with hazel eyes and a long slender nose, with medium-thin lips"... (This gets inserted in the front of every subesquent prompt)

Followed by an attire prompt: 
..."wearing collarless white pleated blouse with a saturated magenta blazer and dark grey business casual pants with purple flat shoes"....

Followed by an action prompt: 
..."sitting inside a blue car holding the steering wheel and driving on a busy city street, with her head pointed forward at the road"...

Followed by a camera prompt: 
..."Wide side view"...

Followed by an enviroment/lighting prompt: 
..."with back lighting."

The complete prompt looks like this:
a 40-year-old woman with tan skin, medium-length black wavy hair parted to the left, with hazel eyes and a long slender nose, with medium-thin lips, wearing collarless white pleated blouse with a saturated magenta blazer and dark grey business casual pants with purple flat shoes, sitting inside a blue car holding the steering wheel and driving on a busy city street, with her head pointed forward at the road. Wide side view with back lighting.

You can switch the order, and insert your own unique prompts. This has been working for me thus far, some scenes better than others. When I get to scenes that involve multiple characters (plus my lead character) Firefly reallys starts to shirt the bed. 

I'm still trying to figure out getting reliable/effective camera angles and NOT having the character look at the camera like a fool. Seems random at best. If anyone has any solid prompts for that I would greatly appreciate the share. Before anyone jumps in I've added "looking at camera" in the exclude list under the advanced field on the sidebar. Does not matter. 

Did you say you did or did not use FireFly for this sequence of images. If you did, I think this is very close. If not, it gives me hope that at some point we'll have something like this in FF. The without clothes is clearly an issue but as you showed here, neck and shoulders isn't really naked but you probably need to use that in the prompt which would raise a red flag. Moving in the right direction.

The real answer is that in ANY AI tool (at the current moment) it's not failproof, given the very nature of generative content, which means "random", but the short answer is yes, ROUGHLY speaking. It's never going to be spot on perfect until the engines start refining their learning models. 

Observe: (Currently working on this project)

Generated source used a a "Style Reference"
In my explorations I've discovered it's better to have a human character without clothes (Adobe has an obvious limitation here) to that the learning model doesn't try to "force" the clothing in any subsequent generations/scenes. 

Now, we generate specific scenes being VERY descriptive. If you're lazy with your prompts your results will be more "guesswork". It's important to note I included the prompt I used in the source image in every new scene/generation, followed by my specific scene descriptors. 

      
As you can see, the results are close, but not spot on. Again, even in Midjourney using the “/prefer option set” you're NOT going to get 100% likeness in each generation. This is the nature of the AI beast. That being said, Firefly's interface is money.

Overall, I find that as it stands at the current moment, multi-scene application using the same character is not refined enough by the current toolset. It's been a somewhat contentious and frustrating journey trying to force a square peg in the proverbial round hole. This method starts to break down whenever you start adding more varired characters into one scene. Firefly can do it, but the other characters look VERY related to your main character.

I have a deadline, so I'm going to have to soldier through it, but I can't wait to tell my Art Director in post mortem that, "this ain't it".

Right, time to brew more coffee. 

In Adobe photoshop it will be very useful to generate same thing on different photos. For examle in one photo I replace some background stuff and I have almost the same photo but I can generate the same background as on previous picture. PLEASE it is very important button generate as on privious image =)))

Bonjour,

ce serait pertinent d'avoir la possiblité de pouvoir travailler à partir de la même image afin de la faire évoluer dans différentes pauses ou actions. 

Ex : un homme qui sourit, face caméra.

le même homme qui fronce les sourcils et s'interroge, les yeux levés en l'air

le même homme qui salue quelqu'un (hors champs)

etc

je vous remercie par avance de votre écoute. 

Bonne journée,

[Removed by Mod]

Vous pouvez travailler sur la suggestion en permettant à l'utilisateur d'importer achaque etape de sa modification l'element similaire qu'ils voudrais editer. Cela donnera encore plus de facilité de une correspondance encore plus avancé de ce que desirs le createur