P: Model needs more training

[AlanGilbertson]

Firefly is struggling with winged insect representations. Here's a typical response to a prompt calling for a Chinese-style black ink of a dragonfly. Eight wings is the most I've seen, but I've only seen a four-winged version once out of more than a dozen tries. Most have six wings, one had seven.

[Rhonda26679422a0py]

This is not really a bug. I would say Firefly does not know what cloud computing means. It knows cloud and it knows computing, however I am getting puffy cloud like technology images. How can we as users help train the model to grab the right images?

[Dave_Williams]

OK, I've noticed that every AI will put horns on a VIking helmet. It's inaccurate and comes from movies etc rather than historic fact. I used this prompt: a silhouetted viking warrior without horns on their helmet in winter amongst the mountains of northern norway and with northern lights.

They all had horns. Not the end of the world, but an observation you may be interested in.

Cheers

Dave

[Tracy.H]

Thank you for the feedback!. This is an ongoing process. The quality of our models will improve with time.

[efsb]

The accuracy of images in relation to text prompts isn't very accurate - for example, the phrase "cajun accordion" generated an inaccurate representation and when this was changed to "melodeon", it was even worse.

This is just one example - I've tried several others and at the moment, the results are way off the mark. Another example - an image of Cologne cathedral with a group standing in front of it added random buildings that don't exist between the people and the cathedral.

Mac Ventura - online text to photo.

[Tracy.H]

Thank you for your feedback. The quality of our models will improve with time, and generated content will better respect the prompt.

[keithl72430567]

Hey, I am asking the AI to generate some Images within my text. I am going for a themed font. I run an Agency named CYA. I want the (C) to represent the strategy so I was asking it to make Graphs, and images that represent the strategy but it literally makes the same thing every time. Some whoosh balls and random lines.

[Tracy.H]

Thank you for the feedback. I am able to reproduce this issue. Since we are training on Stock images, when I search for graphs and charts, this is what Stock offers which is likely why you're getting all the colorful swirls which aren't representative of what you are aiming for. I'll open a bug for this.

[Tracy.H]

Please let us know if there are areas you think our model needs more training. (Merging all similar threads here so the team can access all requests on this issue)

[keithl72430567]

Graphs(Line, bar, Dual Axis Chart, Scatter Plot Chart, Bubble Chart, Waterfall Chart, Funnel Chart) Train it on Social media logos, Website or mobile app mockups, Analytics dashboards, Email newsletters and Digital ad banners. Things like this. Many people using this software are agencies and these are things agencies will be attracted to.

[Balamurugan29099931slm8]

When I typed "shih tzu dog snowboarding on solitude ski resorts utah". I get the beutiful images almost matching what I asked except the location. images match with ski resort. But don't show anywhere with the name or landmark of solitude ski resort. At least name board of "Solitude ski resorts' in backgound would satisfy the customer.

[Tracy.H]

Thank you for the feedback. Our model does have some difficulty with locations and cities. Please use the thumbs up/thumbs down buttons to provide feedback on the image.

[Junior Rocha]

Firefly doesnt know what is a Diatonic Harmonica or a Blues Hamonica, a simple musical instrument, firefly always place other instruments instead.

[bors___]

1) prompt: `wombat gets in the way of a tourist couple's selfie. pov is the selfie camera`

- model doesn't always draw a wombat (often kangaroo or koala)

- model sometimes replaces the human head with that of the animal

2) model doesn't handle numbers or amounts well, you have to really overemphasise the amounts to force it to draw more than a handful of items
eg `infinite number of monkeys sitting on a house, crammed, too many` vs `lots of monkeys sitting on a house` or `50 monkeys...`

[bors___]

Another example of model problems - it doesn't know specific animal, insect or plant names, especially if they're less common. Some examples:

- `c0ck-of-the-rock` or `Rupicola peruvianus` - 0% hits on either common name or latin name

-  `common firecrest` or `Regulus ignicapilla` - common name gets it right about 25%, latin name 0%

-  `mouse lemur` - only ringed-tailed lemurs are shown, and for its latin name `microcebus murinus`, all sorts of weird creatures arshown.

[JoshCatchur]

More specific might help. Stacked Bar chart is almost great. It could at least be a good starting point for post work. Columnar charts is a little better. Histogram chart is better. Bar graph. Sunburst chart might be interesting for the C, not so much for the YA. Stacked Area Chart created some pretty moments. Multi-level Pie Chart is interesting. And this was pretty good too: "multi-dimensional Columnar Graph [avoid=sphere and circle], heat map coloring"

[Bob_Hallam]

I've interacted with Firefly over the past week and have to agree.  So far the AI does not perform well on any tasks without much user interaction to rate and train.  I have not gotten a good image from Firefly yet but it appears that I am doing the training.  Not beta testing at all. 

[JR Boulay]

Firefly seems not to have learned the difference between a dolmen and a menhir!

I'm looking forward to it.

😉

 

[Nikola2955864008fj]

Hi,

How to introduce some new model or shape to system? For example, Firefly has no idea what platypus is, and what to do with it, in any context.

Merging all similar threads here so the team can access all requests on this issue

[AlanGilbertson]

In terms of improving the model, there are some areas where Firefly continues to be reliably unreliable  and could stand improvement:

• Any prompt that calls for multiple animals (say an elephant, a crocodile, and a snake -- nothing rare or unusual) will result in chimeras: a snake with the texture of an elephant's trunk and the head of something that might be a crocodile but it's hard to tell, for example, or a couple of tiny elephants with snakes instead of trunks.
• More than one person in a scene results in very distorted bodies, faces, body parts.
• Many human figures have grotesque distortions: extra ears, a nose in the middle of a forehead, extra or fused limbs, fantastical shapes where a face should be, bizarre hands or feet.
• Firely is beginning to learn that "four fingers plus a thumb" is the normal complement for a primate hand, but hands and feet are rendered correctly less than half the time and it is very rare for any human figure to have two sensible-looking hands and two sensible-looking feet.
• Introduce unusual color to a prompt ("profile of young man with blond hair and green beard") and Firefly cannot make the beard green or the hair blond. The beard is mid-brown and the hair green. The subjects are all youngish white guys.
• Firefly has real trouble with vehicles if there are people in them. A car almost never has a steering wheel, and if there is one it is usually at an impossible angle. People protrude from the car where the windshield should be (not a coupe; a sedan). 

These seem to be less frequent with generative fill in Photoshop.