firefly-custom-models

[vestal kestrel]

Welcome to the Firefly Custom Models Beta! 🎨🧠

You made it to the early access crew. This space is for you to experiment, break things (in a helpful way), and help us shape what Custom Models becomes for creative pros like you.

What are Custom Models?

Custom Models lets you train Firefly on your style or characters so you can:

• Generate new work in a consistent illustration or photo style
• Keep a **recurring character ** and vary them across scenes
• Explore variations and concepts that feel unmistakably yours without constant prompt retweaking

As a reminder, you'll use the generative credits from your existing Creative Cloud plan when training & generating. Credit details are shown in-product.

Quick Start 🚀

1. Open Firefly in Chrome on desktop:
   🔗 https://firefly.adobe.com/custom-models
2. Make sure you’re signed in with your Adobe ID and select Custom Models in the left nav if you’re redirected to the main screen after logging in.
3. Pick a use case:

• **a. Illustration style **
• **b. Photographic style **
• **c. Character ****

4. Upload 10–30 high-quality JPG/PNG images (1000px+, same-ish aspect ratio).

• Important: Only upload content you have rights and consent to use.

5. Check your Training set score and aim for 85+.

For details and more tips on how to best train your model visit our Help Guide

Feedback/Bug Report Instructions 💬

Use this Discord channel to:

• Provide feedback with as much detail as possible — **screenshots are very helpful! **
• Ask questions, share tips, and suggest UX improvements.
• Call out any steps that felt confusing, slow, or broken (upload, captioning, training, generation, sharing, etc.)

Suggested format for feedback:

• Model type & use case (e.g., “character: mascot illustrations”)
• What you did (upload details, training set score, prompt)
• **What happened **(screenshots of good/bad outputs)

The more specific you are, the more we can actually fix!

Helpful Links 📎

• Product access: https://firefly.adobe.com/custom-models
• Overview Help Guide: https://helpx.adobe.com/firefly/web/firefly-custom-models-in-creative-cloud/train-firefly-custom-models.html
• Best Practices: https://helpx.adobe.com/firefly/web/firefly-custom-models-in-creative-cloud/best-practices-for-creating-custom-models.html

Thank you for being part of this early wave. Your experiments here will directly shape how Custom Models ships to the broader creative community. Let’s build the future of Firefly together!

[vestal kestrel]

@earnest ore @forest osprey @daring comet Welcome to the custom models private discord channel! We're waiting for a few more people to join the channel, but soon we'll be posting creative challenges, polls, and questions to understand more about how custom models are working for you. In the meantime, I wanted to welcome you here. If you have any questions or thoughts you want to share, feel free to post them here and I'll make sure the team sees them!

[buoyant hull]

https://tenor.com/view/grogu-star-wars-mandalorian-baby-yoda-gif-9168999412781595068

[strong wind]

Hi <@&1436446216513785947> friends! Nice to see a few people filtering in after getting access to the beta experience.

We are hosting a small discussion group this Thursday (Dec 18) at 1pm PT to hear about your initial impressions and experience. Please DM me or 👍 this message if you would like an invite! We have a couple spots left.

[runic coral]

UI Feature Request:

1. Allows upload of captions in text format (same filename as the image files)
2. It would be nice to upload all images and captions in a zip archive
3. The instruction is not clear about whether higher resolution images (>1000px) will improve quality of the custom model
4. When an image cannot be processed, an explanation is needed. I can understand that there needs to be certain level of safety filter, but currently, the filter is not very consistent

[strong wind]

Thank you! Sharing this feedback with the product team.

[mental snow]

Thanks for having me ♡
Curious --
Will there be a way to export / back up / transfer our custom models?
Is there currently, or will there eventually be a way to use a reference image with the models we've trained (like we can with firefly models)?

[strong wind]

Hey @mental snow - thanks for joining the beta and discord!

• what type of export / backup would be helpful for you? are you thinking about the ability to move it to another platform, share with others, or just have a backup?
• yes, the team is working on additional capabilities and settings when you generate using your custom model (like reference images). this is still very much an early version.

[strong wind]

Hey @mental snow and @hazy fiber (or anyone else who is interested): here is the sigup link for tomorrow's 1pm PT discussion group. only requirement is that you have trained a model prior to joining the session and are open to walking us through your experience (what worked, what didn't, anything surprising!). We're offering a $50 gift card as a small thank you!

[hazy fiber]

@strong wind there's a small chance I'll have a call tomorrow at 4pm or 4:45p est (1p/1:45p pt), but if not then I will join.

The signup link says it's fully booked as well, but I'll keep a lookout if it frees up and time allows for sure.

Trained a custom model on Cinestill 800T photos and it's been great at emulating it for sure.

[strong wind]

🤦‍♂️ sorry, the link was busted. now working! would love to have you if you can join. otherwise would love to hear your feedback here (whats worked, what hasn't, anything surprising).

[hazy fiber]

Hey you all. Hope you're each navigating your day well.

Ran into an issue when trying to train two custom models at once for today (2 being the daily limit and I was at 0 of 2 trained for today 12/18):

Dataset 1 ("DSLR - Moody") - This group was training and was at 1% within the training process. This used 500 credits.

Dataset 2 (GoPro Hero - Action...") - Upon clicking "train" for this one after Dataset 1 was already started and at 1%, both groups presented the dialogue box below and Dataset 1's training was halted. It says my daily limit has been reached and prevents me from restarting training on either of these datasets for the time being. The 500 credits from Dataset 1 have not returned either.

[hazy fiber]

Update on this similar to the call: That Dataset 1 ("DSLR - Moody") completed, so it must've still gone through with the process regardless of the halt & dialogue box. The training process wasn't trackable by that percentage bar it usually shows when training a model.

[hazy fiber]

@strong wind thanks for the cool call, the extra credits, and your time too.

A few questions came up as it relates to training more custom models for testing:

1. Is the stuff we're putting into these custom models being used for ML training or commercial products from Adobe down the line? or is the data not held, closed off from other products, and or not used for training?

2. Similar to #1, is likeness of people being stored for anything commercial or beyond the private testing or nah?

---

I cannot remember the answers to these two questions, but remember the prompt before training the first model alluded to us having ownership on those things. That prompt no longer appears for me when creating the other models.

Just wanna make sure not to limit myself on what can be placed in or not, but also not jumping into it with other stuff unless clear on training & commercial stuff.

[strong wind]

Thanks for joining and sharing your experience! Super helpful. TLDR: we do not, have not, will not train on any of the data you upload. the longer answers are attached. and our public commitment to commercial safety here

[strong wind]

noted. that was a weird state that we will look into.

[hazy fiber]

Thank you for the quick and detailed response, and copy that on data. Lifts that hesitation for me. Onward into da unknown!

[hazy fiber]

@strong wind good afternoon.

Are any of Adobe Firefly custom model outputs able to be shared or shown in review just yet? or not just yet.

[strong wind]

Hey Christopher - are you asking if you can use these outputs in your actual work and/or publicly?

If so, since this is a "public beta" (anyone can join the waitlist, although we are letting people in slowly in waves) it is not under NDA and you are free to both use the outputs for your production work, share and talk about it publicly.

[hazy fiber]

@strong wind more so just to showcase it to others vs being used in a commercial workflow as of now. Like if someone in the beta wanted to do a "review on Adobe Firefly custom models" or make a post about it.

Got you on talking about it publicly. Think it's a strong product so far that is in line with copyright laws as well which is huge for adoption in my neck of the woods.

[strong wind]

got it. yes, absolutely, you are free to post about it!

[hazy fiber]

Good afternoon you all.

Tried training two models and ran into some hiccups that could be good for feedback:

• "Nova" a character model based on a female comic protagonist of mine.
  -- Little to no captions altered, so the model was able to detect who "Nova" was from text.

• "Sci-Fi Comic Aesthetic" an illustration model based on a comic of mine.

Not sure why some picked up on one model and not on the other, and white background vs filled frame, etc.

---

Here are some newer generations from a photography model ("GoPro Hero - Action Photos") trained off of old GoPro Hero 3-4 camera photos of mine.

None of the training data had bears, gingerbread men, clowns, young children, cubes, etc.

Not many straight on photos of people with their faces, so it may have struggled with that as you can see from the clown and running photos.

[kind imp]

Hi, I am just getting to explorations but do not see the use cases to begin making a custom model—just an empty screen. What do you advise?

[strong wind]

Hey @kind imp - thanks for joining! Sorry for the slow reply. Adobe is on our "winter shutdown" this week 😉 Re: the empty screen, are you on a team plan (vs individual/solo)? I believe in that case your admin has to manually approve you for the beta update (otherwise you will see a blank screen). If not, let me know and we will investigate. Happy New Year!

[strong wind]

Thanks @hazy fiber - can you share your prompts to generate? The model should match style but are you saying it placed in the gingerbread man / bear / clown without prompting?? 🧐 Thanks!

[hazy fiber]

Yes will do @strong wind . And it does match the style well. No I was saying it did a good job incorporating a Gingerbread man in the desired style without any former training data for the custom model 👍

Happy almost new year to you!

[kind imp]

@strong wind - excellent, many thanks + Happy New Year! I'm on a team plan within my corporation. It's very possible that this is an admin / approval issue. I've started working with my admins to find out what's going on—I'll come back here as soon as I know more. My teams are also on winter break so I likely won't see any traction until next week but will keep you in the loop; we're super excited about this feature

[hazy fiber]

Prompts for you on these photos @strong wind :

1. A floating cube in the middle of a volcano

2. Gingerbread cookie Running from a large bear

3. Swimming through the water, facing the camera, swimming goggles on, little hispanic boy with duck swim trunks, and floatation devices on his arms

4. Group of toddlers running from a clown shooting a water gun

5. Wide angle shot of a person with shoes running from another person, facing the camera, low angle, low to high angle, person with shoe's shoe is close to the camera

[strong wind]

Hello <@&1436446216513785947> crew! Hope your new year is off to a great start!

We are seeing a relatively small % of folks who have been admitted into the beta actually train a model. I am curious to hear from the group here if anyone has not gotten around to it yet (totally fine!) and what the barriers might be. The more people that kick the tires the more we can improve the experience.

[runic coral]

I am heading out so I will find my biggest complaint first. The trained model is like its own island. I can't use it with other options of the Firefly generator. Content type, composition, styles, effects, etc.

[blazing wind]

The 500 premium credits is the issue for me.

[hazy fiber]

@strong wind good afternoon. Hope you’re having a nice weekend.

Sent you a DM about potentially sharing some stuff incase it gets like 1 more person on board w/ testing.

[runic coral]

Continuing my previous message. I think it's a missed opportunity not allowing a custom model to be used inside Photoshop. Imagine this:

I am working on an illustration. At the last minute, I am told to change the outfit of a character in my illustration. If my custom model works in Generative Fill, I can quickly generate the new outfit using my own style. I am not even looking for someone perfect, just something that lets me meet my deadline.

[eternal stream]

I work extensively in product based content and advertising and have been experimenting with using custom models to train on specific products. Ive been trying different approaches with the training assets and Im getting some intersting results. Still a ways to go but im encouraged by the process. Overall I find that I can get the broad strokes like can shape, branding logo etc very accurate but no so much when it comes to detailed type and design elements. I going to take the last set of images i trained on and remove all type except for the front label design and see what i get.

[vestal kestrel]

Did you end up trying removing the type aside from the front label design. Curious how that worked out. Glad to see that the broad strokes have been accurate!

<@&1436446216513785947> Thanks to everyone in this channel for trying things out and reporting back to us! Starting today, we'll be posting a weekly creative challenge to give you some ideas to try out with custom models. For each challenge, we'll select one participant at random to receive some fun Adobe swag! This week's challenge is:

🫵 **Create a Model That Represents You **: Train a small model using a few images, designs, or prompts that reflect your personal style or voice. Test it out, then share your favorite output here for a chance to win swag! Let us know what you were hoping the model would capture and where you felt it was or was not able to achieve that!

[eternal stream]

I did try training with images that only included the front label. Im still getting really nice results on the overall shape, design, layout, texture of the can and accuracy in the Logo but still not achieving any real accuracy with the small type details. How much does the captioning influence the training? Im thinking that for the next model i will be much more detailed in captioning the images
with label info.

[strong wind]

Hey @eternal stream - thanks for kicking the tires! The use case you are attempting (take a product image and reproduce it / place it consistently across a variety of contexts) actually isn't one that is well suited to custom models. Custom models work well at inventing new images, scenery and characters in your style but not reproducing a subject (like a product) with perfect fidelity (text, dimensions, detail). But this is great feedback because its an area of confusion a number of folks have mentioned so i think we need to do a better job of educating on this in the product experience.

The great news is, there is another approach for what you are trying to accomplish and it already exists. That is using reference images. You can go to image generation on firefly.adobe.com (or, my personal favorite firefly boards) and choose an image model that accepts reference images and upload an image of the can. then you prompt to say where you want to place the can / what layout / what background.

I ran a few tests in Firefly Boards trying to recreate what you had and here are the results. Happy to chat more about this!

[hazy fiber]

@vestal kestrel this was a favorite output today with a custom model.

Dataset: trained on 14 Ilford style b&w 35mm film photos.

Prompt: Landscape shot, a small humanoid being glowing white in a dark enchanted forest, fireflies in the air, there are also several other forests in the distance, there is fungi glowing, a bright river going through the composition and the river is lit from below by glowing fungi and fish. Overall I thought it captured the idea well.

Pros: It definitely captured the black and white, grain, and compositions very well. It did a solid job with the glowing water, fungi, and adding fish even though I didn't have any training data with that specifically. It also did a good job with contrasting light from shadow night or evening.

Cons: It didn't quite capture the realism of the fungi or fish towards the back. It definitely didn't show other forests in the distance and was a bit more confined than I imagined. It also didn't have any glowing fish.

[eternal stream]

Fantastic! -Thanks Colin!

[vestal kestrel]

<@&1436446216513785947> Hey Custom Models Beta testers! 🌟 Special shout out to everyone who responded to our January feedback survey. Your input is already helping us prioritize what to improve next. We will be adding a short instructional video to the experience soon. 🎉

As always, we love to see what you are creating as you experiment, so keep sharing what you’re making as you test. If you’re comfortable, include:

• What you trained (photo/illustration/character)
• How many of your assets you uploaded to train
• What you were trying to generate with the model
• What surprised you (good or bad)

We’ll be adding in the final wave of private testers in the next couple weeks, so stay tuned for more inspiration from the community!

What will you create a custom model for this week?

[toxic yarrow]

hi this is all i see when i open the custom models tab! there is no where for me to start creating one

[ember plover]

I am running into the same issue.

[dense ice]

Same issue.

[kindred raven]

Bumping, having the same isue

[round vessel]

@toxic yarrow @ember plover @dense ice @kindred raven thanks for sharing, can you each confirm your AdobeID (privately if you wish) so we can troubleshoot.

[kindred raven]

Just Dm'ed it

[dense ice]

Responded.

[vestal kestrel]

<@&1436446216513785947> Welcome new members to the Custom Models Beta – and a big shout out as well to those who’ve been on this journey with us through the winter holidays! You all have early access to build, train, and test your own models so you can personalize and speed up your creative workflows. We can’t wait to see what you do with them!

For those just joining, here’s a Welcome Challenge For You:

Create a Model That Represents A Favorite Project
Train a model using a few images or designs that reflect the style or subject of a project you’d like to expand on. Test out your model to see what it will generate and how it could save you time. Share your favorite output and what you were hoping the model would capture.

👉 To get started, visit https://firefly.adobe.com/custom-models and sign in using the same Adobe ID you used during sign-up.

Using the same Adobe ID is important so the feature shows up correctly.

Enjoy Experimenting!

[vestal kestrel]

<@&1436446216513785947> An open question for you. Feel free to answer whether you just joined or have been with us for a long time.

1️⃣ 👁️ What were your first impressions of the Custom Models Beta? Anything surprising, interesting, or confusing so far?

[toxic yarrow]

sent you a dm!

[toxic yarrow]

i finally figured out what the issue was, for those that had the same issue if you are on a company account you need to switch to a personal profile! one thing that i'm a bit disapointed in is that you cant train it on products , custom models for illustration are great but what i was really looking for is a way to train a model to use my products in the right way and proportions. I work in marketing on CPG. I haven't had much luck with firefly boards or the image gen with a reference as it seems to change everytime i run it. Any suggestions?

[ember plover]

Hello Theresa, I just DM’d you. Thanks!

[hazy fiber]

1. The custom model idea and forward thinking with ethics and generative Ai was refreshing to see off rip. Especially since Adobe is a large Silicon Valley & likely ai forward company.

Found it interesting how a system / product like this can be created + the labeling side of stuff & how whatever is helping a user build the model is interpreting that user input.

Confusing how to properly train a model to 100% score, what certain types of media registers and doesn’t register within training, and what to do if running into errors (some of which have stopped me from further testing.)

[round vessel]

Thanks for the additional feedback and insights, we are always looking for ways to make this experience better. Since it sounds like you have access to a corporate account, you could try Object Composite. Here's a link to more info and a demo for that feature: https://helpx.adobe.com/firefly/web/work-with-enterprise-features/create-object-composites/object-composites-overview.html

[vestal kestrel]

<@&1436446216513785947> A quick poll for you above! If you feel like it, tell us why you chose the answer you chose? Was there anything that got in your way?

[vestal kestrel]

<@&1436446216513785947> We have a fun challenge for you this week!

**Build a Mood or Aesthetic Model **
Train a model focused on a specific mood or look—like bright, calm, cinematic, retro, or bold. Share a few of your favorite outputs.

[strong wind]

hey @toxic yarrow - id be curious to hear more about what you are trying to accomplish and what isnt working with the reference image workflow (where you say stuff "seems to change very time"). feel free to share thoughts here or DM me and we can setup a time to chat if helpful.

[vestal kestrel]

<@&1436446216513785947> Hi all! Huge thanks to everyone who shared feedback in our March survey – we really appreciate it. Excited to share that Firefly Custom Models is now in public beta and available worldwide (no more waitlist 🎉). Along with that, this channel is now public, so anyone in the Firefly server can join!

🎯 New here? Check out the quick start video inside the experience — you'll be training your model in no time!

🚀 The community is growing — and so are the creations! Whether you're brand new or already testing, we want to see what you're making.

__Drop your results below and tell us: __
📸 What you trained (photo/illustration/character)
🗂️ Assets uploaded
🎯 What you tried to generate
😮 What caught you off guard

New members are joining every day — let's show them what's possible! 🔥

📣 Know someone who'd love this? Share this with friends, colleagues, or anyone curious about training their own model — and help us shape the future of this feature with your feedback!

Access the public beta here

[static citrus]

My Custom Model is up to 87%!!🎉

FYI: With my plans, to use my custom model, it's 20 credits for a generation and you get 2 images per.

[kind imp]

Hi there, tuning back in! I now have access to the beta custom model area. I do however have an issue that I am unable to solve. One of my team members is unable to access. Is there a beta access list on your end? I would like to add a user

[vague pewter]

whats the error message they're getting?

[kind imp]

@vague pewter My team member sees "You don't have access to train models. If custom models have been shared with you, go to Model Inventory to view them." I, however, am able to access, just fine.

[vague pewter]

do they happen to live in illinois or texas?

[kind imp]

hi, they are not

[vague pewter]

are they on an enterprise or team account where an account manager would need to give them access?

[kind imp]

we're on a team account and definitely have an account manager and a team that grants access. sounds like it would be an internal access thing?

[vague pewter]

I would check with them yeah

[cedar dust]

<@&1085962062636797962>

[vague pewter]

yes, we've been aware, we need to wait for an admin

[old sleet]

I see that custom model training is in beta, and the options for custom model methods are cool, but is their a setup to have a truly "custom" method?

I would love to talk with someone with further information, I have a concept of training data, but doesn't seem to fit the approaches allowed currently.

[vague pewter]

what sort of customization beyond the current system are you looking for

[old sleet]

I have an avatar system that is a single Female image, that I have clothing for (which are layers over top of the female body :: dresses, pants, shirts, hairs, etc.) I'd love to be able to train the concept of creating clothing that layers overtop of the female body layer

Currently when you use Firefly (really any model) when you want to create a layer of clothing, the body becomes a part of the new layer, which then defeats the purpose of the clothing/hair/etc being it's own layer

So I'd love to be able to provide the base body (with 0 clothing, hair, etc.) and provide images of the clothing/hair layers so that a model could fully understand that these are overlayed, rather than attached to the body itself -- the body in itself is a flat object that never moves, always in the same position, etc.

[static citrus]

@old sleet Having a nude anything is against Adobe Firefly T&Cs. I do not see this being a workflow that would work.

[old sleet]

This is the image I was meaning, not necessarily nude

[static citrus]

@old sleet Too much skin for Firefly.

[old sleet]

I guess a question then is: How would you ever train any sort of clothing data?

It's also a fake cartoon character

[static citrus]

@old sleet I would not if you want the image never to move. You could try "Try On" preset in Firefly boards.

[old sleet]

Where is that @static citrus ?

[static citrus]

@old sleet In Firefly Boards.

[old sleet]

I don't see a "Try it on" preset

Also, would this create a brand new image? Or a layered image as I was mentioning is a requirement

[static citrus]

Under Presets.

[vague pewter]

it would not create a layered image, you'd have to do that in something like photoshop with background remove, or you could try using gen shape fill in illustrator to create clothing by creating a path that generally fits what you want to make.

[vestal kestrel]

If you're new to this channel, Welcome to the Custom Models Beta. We want to give abig shout out to those who’ve been on the waitlist. You made it! You now have access to build, train, and test your own models to personalize and speed up your creative workflows. We’re just getting started here and can’t wait to see what this community creates together.

To kick things off – we’re launching a Custom Models Welcome Challenge! Choose from the options below and go with whatever inspires you:

✏️ Illustration — build a model around your unique art style
🤵 Character — create a consistent character across scenes
📸 Photography — craft a signature photography aesthetic

Visit https://firefly.adobe.com/custom-models to get started and share what you build right here. Let's see what this community can do!

Quick reminder – To protect your work and your privacy, your custom models remain private by default and Adobe only uses the images you upload to train your custom model. Your content is never used to train Firefly’s foundational models, and you’re always in control of who can access what you create.

[mossy dirge]

My immediate reaction is that the process is quite cavalier: It decided on a name without asking. The name is a long way from what I intended or would have chosen, but I could not discover any way to change it. Annoyance #2 came when I edited captions to get rid of the all caps name and change it to what I wanted, only to discover that the unwanted name was compulsory.

After that, the process is remarkably uninformative. It comments on images that could be improved, but doesn't highlight them in the set. Tiny thumbnails are not necessarily useful, and the commentary is mostly vague, giving no concrete examples of what is meant by "normalize the lighting" or why "varying lighting and texture quality" should be an issue.

I pressed on after I got the model score to 84. We'll see how it turns out. So my answer to the poll is "None of the above"!

[mossy dirge]

Having trained it, I was able to generate a pair of images that were true to the character. Then I got "Something went wrong" when trying to generate another, then the model was gone. The model inventory is empty.

[unkempt fox]

I need to play with it again. But so far I didn’t see a way to change resolution/image size. And I feel like it should be cheaper than 20 credits per image.

[mossy dirge]

Aaand, the saga continues. Now the custom model is back.

[undone fable]

Hey, I feel like i am getting conflicting recommendations when trying to train my first custom model. First it tells me my subject pool is too diverse, then it tells me to make it broader?!?

[vague pewter]

they're just recommendations, like it says "if the goal is broader style training" you have to decide what you want the model to do and make changes based on that

[undone fable]

I see. 500 credits is a lot to spend on trial and error!

I assume once the model is made, you can't just go back and edit it without paying again, right?

[vague pewter]

correct

[undone fable]

OK, I clicked "Train" on my first model and walked away to let it run. The credits were deducted, but it still says "Draft". How am I supposed to actually use it?

[vague pewter]

it's likely not done yet, when it's done you can select it from the models drop down in text to image

[brisk hill]

almost got to 90 but then reduce to 85 😭

[brisk hill]

(going to sleep now)

[vestal kestrel]

Hey there, did this ever resolve? Were you able to access your model?

[vestal kestrel]

85 is still very good! These little birds with backpacks and beanies are adorable!

[undone fable]

I contacted support, and they asked for screenshots and said they would look into it. I just checked, and now it says trained, so I guess they figured it out! Now to go test it!

So do the custom models only work for text-to-image? Or is there a way to do img-to-img?

[brisk hill]

thank you 🙂

[vestal kestrel]

Okay good! Glad it's working now.

Great question. At this time custom models are text to image only!

[vestal kestrel]

Hi Alan, Glad the model returned. Thanks for this detailed account of your experience getting started with custom models. Sounds like you ran into a few things that made the process a little bumpy and less than ideal. I'm going to bring this up to the custom models team. Feedback like yours is really helpful to us as custom models is still in beta and we're working to improve it every day.

[vestal kestrel]

Hey there! thanks for sharing this. Really great feedback for us to consider as we work through this beta period!

[mossy dirge]

The model struggles with cartoonish content. Here it complains about text in images that have no text at all.

Also worth noting: There seem to conflicting requirements for variation on one hand, and uniformity on the other. If the user wants uniformity in the output, this is quite unhelpful. The model tends to see similar images as "near identical," and assumes that what is wanted is variety.

[vestal kestrel]

Ah that's interesting. It's assuming the image has text not because it recognizes text but because this particular art style is often accompanied by text. Hmmmm.

Your note about uniformity vs. variety is also super interesting. Thanks for digging in to the patterns you're noticing and sharing them with us. I'll pass this on to the custom models team.

[vestal kestrel]

FYI I did share this with the team and it turns out you can edit the name in the RHS panel that appears on the right hand side (see screenshot). Thanks to your note, we've updated the Custom Models help page to make it more clear where you can go to edit this.

[mossy dirge]

Yes, I stumbled across that right-hand panel and discovered the editing capability. Good move to update the help, because it was definitely non-obvious!

I think it would be very helpful to have an intro page that walks through these steps in just a little more detail than the current one. Get someone who doesn't know about custom models to work their way through it, user testing Paul Krug style. You'll find out exactly where you need to add detail. Someone who is familiar with the product can't do it alone, because they don't have the objectivity. (Don't ask me how I know that! 😸)

[signal isle]

do custom models work with partner models?

[vague pewter]

not currently no, you can make a request #1087561412269854840

[vestal kestrel]

<@&1436446216513785947> New week, new challenge to give some ideas for how to use Custom Models.

🦹‍♀️ This week's challenge is — Meet Your Character!
One of the most powerful things Custom Models can do is keep a character consistent — same face, same style, same world — across completely different scenes and stories.

Your challenge: Build a custom model around a character and use it to create at least two scenes featuring that character. It could be a storybook, a cartoon series, a graphic novel panel — whatever fits your creative world.

🍎 Educators — this one's especially for you. We'd love to see how you bring characters to life for your students.

✨ Share your character and the scenes in this channel. Tell us who they are — we want to meet them.

[undone fable]

I actually just brought this up myself: #firefly-custom-models message

[vestal kestrel]

Thanks @undone fable and @mossy dirge I'll bring this up with the team!

[undone fable]

I plan to try a character training next. Should all the images be full body images? Or should I include some facial closeups as well?

[undone fable]

I definitely meet this requirement. I am just a hobbyist, pretty new to AI as a whole, so I am definitely stumbling my way through things!

[icy prawn]

I was excited to see if a custom model would generate images that looked like I drew it when I am frequently using AI gen to do the opposite and make up for my lack of proper drawing skills. I would call these results a success because this is what my hand drawing looks like (unfortunately, haha).

[mossy dirge]

This isn't specifically a custom model issue, but Firefly seems to struggle with the concept of a full-length portrait. Feet are often cut off, but this was a bit too much:

Here are some Indira images from the custom model that the model builder decided to call "enchantress."

[vestal kestrel]

@undone fable @mossy dirge Any chance you would be able to share the image sets you were using to train when you got the contradictory messaging around variation vs. specificity? You can DM me or email me at [e-mail].

[undone fable]

Sent an email

[undone fable]

Thought I would share this comparison of Firefly custom character and a Freepik custom character trained on the same data set. I am not particularly happy with either model, as both of them insist on making my character skinny, when I designed her to be chubby. But at least Firefly manages to stick with the anthropomorphic zebra character. Freepik consistently generated one zebra, and one human.

[fiery frigate]

[MASTER STYLE] Single figure on a narrow trail through a geometric forest

[static citrus]

@fiery frigate Firefly is not a Discord bot.

[fiery frigate]

where the Discord bot ?

[static citrus]

@fiery frigate There is no Discord bot. Go to https://firefly.adobe.com/

[mossy dirge]

Custom models reflect their legacy with appallingly bad rendering of hands.

[brisk hill]

true, in most cases, it does with faces

[mossy dirge]

There's a rookie photographer mistake that seems to be embedded in the Firefly paradigm for full-length portraits. You'd correct any beginner or photography student for cutting off the subject's feet or the top of their head. Firefly is so addicted to this, that it becomes almost comical to see how far it will go to avoid having feet in an image. Custom models are Image Model 3, I think, and they have all of its flaws.

[cold marten]

I have uploaded 10 images of the exact same resolution (2048 x 2048). 5 of them were accepted, and 5 received the error "can't process" with no indication why. What does "can't process" event mean? The images uploaded successfully, but did not auto-generate a caption. Can I still train the model on images that can't be processed? I'm getting a good score with the images (85%). Please help.

Gonna go work out. Please tag me in responses.

[cold marten]

OK, I'm back

[proud yacht]

create image one like tree

[vague pewter]

Adobe Firefly is not a discord bot, try prompting at https://firefly.adobe.com/

[vague pewter]

sorry im unable to access custom models, but it could have been a temporary server issue

[round vessel]

@cold marten are you able to send a screenshot of your training set and the error so we can share with the product team? here or DM is good.

[cold marten]

Actually, I made some changes to block telemetry in Windows. I broke a lot of Adobe. So, gonna restore the previous settings and see if that helps.

[cold marten]

OK, I restored the settings. Still having the same issue. Firefly can't process images. CC app doesn't show half the tabs and content that it should. Adobe home page won't even load in browser.

[round vessel]

@cold marten suggest you reach out to tech support to get your CC working again post-OS changes. In terms of your Custom Model images, that's likely too much skin for the Firefly model.

[cold marten]

Tech support spent a couple hours remoting in to my PC last night. They couldn't figure out the issue. They said that I probably have malware, which is ridiculous. There is zero indication that I have malware. I ran two different malware detectors and found nothing. I feel like they just gave up because they couldn't figure it out. Meanwhile, I still cannot update or get full use of software that I'm paying a ridicuouls price for. Thanks again Adobe.

[vague pewter]

have you completely uninstalled and reinstalled creative cloud? it seems like whatever changes you made affected the program files. It's reasonable to suspect malware if they weren't aware of the changes you made to the os as that wouldn't be something a standard user would mess with.

[undone merlin]

I can't access the custom models. I get this page:

[vague pewter]

seems like your subscription type might not have access to this feature? can you access other premium features? are you on an enterprise or education account that doesn't have this feature enabled?

[brisk hill]

tested out today during prompt jockey event

[desert steeple]

Hello Custom Model Channel, apologies for my ignorance as new to discord and new here, Should I be able to access Custom models in the UK yet? We have an enterprise agreement for all apps, I can get firefly but no access to custom models ?

[static citrus]

@desert steeple Your admin may need to activate custom models since it's newly released.

[desert steeple]

Thank you - I wasnt sure if it was just a US release to begin with

[static citrus]

@desert steeple I'm in Germany and have access.

[cold marten]

Completely reinstalled Windows, and that still didn't fix it. I have a Pro account. I have all of these features. I have used them before.

[vague pewter]

that was a reply to a different user, with a different issue. Did the reinstall fix your issues with the cc apps not displaying the UI? did you reinstall your browser? is Windows firewall or third party browser extensions blocking the site from loading? we're not tech support here, Adobe support would be your best bet.

[brisk hill]

another fun way to explore character model in response to a challenge posted on X

[stray parcel]

ALGUÉM ME AJUDA ? NÃO CONSIGO FALAR COM A ADOBE E PRECISO ENTREGAR TRABALHOS !

[modern vessel]

And it was a custom model you built or some other model?

You can ask for help in #ai-video or if its a specific partner model you're having trouble with, post the details #partner-models. Be sure to include more than this screenshot. Include the video model you're using, the reference image you're using and the prompts you've tried. Then perhaps someone can assist.

[cold marten]

No, Adobe support is not my best bet. They tried and gave up. Yes, I reinstalled everything beginning with the OS, then all apps. Browser extensions were the first thing I eliminated before even contacting support. I finally got it to work by switching VPN protocols from Wireguard/TCP to OpenVPN/UDP. There is no indication anywhere that I could find that this should be an issue. Even support didn't ask or say anything about it. Total 💩 show

[vague pewter]

Well you didn't mention you were using a VPN either when we asked if anything could be interfering with the website loading, and using a vpn would very much do that. I'm glad you got it sorted out but this does seem beyond the scope of basic tech support

[cold marten]

I don't expect support to know how to make Adobe work with my VPN. But, a simple mention that it might be causing a problem would have been a nice start. I don't disable my VPN for ANYTHING! It's a matter of principal for me.

[vague pewter]

I can't speak for support, but you're yelling at us in the discord, and we had no idea you were using a vpn.

Had we known that, we would have suggested disabling it

[cold marten]

Sorry if you think I'm yelling. I'm not yelling. I was just asking for help.

[sonic fox]

Hi, I just joined. I'm also new to Firefly. I set up a "model" to use in Firefly. Somehow I can't get it to work right. What am I missing?

[vague pewter]

are you getting an error message?