Is Firefly 2 a first-class image generation model?

By admin
Adobe just released the

second
ORDINAL

version of

Firefly
ORG

, its image generation model. While models such as Stable Diffusion and

Midjourney
PRODUCT

are the market leaders in terms of result quality, the uncertainty (warranted or otherwise) about the datasets they’ve used in their training means that it’s important to have strong rivals that are trained on licensed data.

I had high hopes for

Firefly
ORG

when it was released, but unfortunately it fell short in many areas. So is Firefly 2 any better? In short, yes. But is it good enough to be a viable contender? Well, let’s take a look.

Below are

three
CARDINAL

tests I’ve been tracking. For each I’m going to show the prompt I used, followed by the results in (

1
CARDINAL

) Firefly

2
CARDINAL

, (

2
CARDINAL

) Firefly 1, and (

3
CARDINAL

)

Midjourney
PRODUCT

.

close-up portrait photograph, a young

Korean
NORP

woman with pink hair, holding a tan Shiba Inu puppy, smiling joyfully, studio lighting

With a simple, single-subject image, all

three
CARDINAL

models give a decent result. Firefly 1 handled this well already, and

Firefly 2
PRODUCT

is a little improved. This is the prompt that got the worst result from

Midjourney
ORG

; the image is fine, but it doesn’t seem to know what a Shiba Inu is.

candid photograph, a group of young boys and girls standing in a

London
GPE

street, mixed ethnicities, natural light

Firefly 1 struggled badly with shots featuring multiple subjects; some of the poses were awkward, the faces have strange details, and the scale is very weird. Firefly

2
CARDINAL

shows a huge improvement here. But

Midjourney
ORG

really nails this shot; it looks great.

a

Highland
LOC

cow on a

Harley Davidson
ORG

motorcycle, wide mountain road, golden hour, freedom, speed

This is an unusual combination of subjects;

Highland
LOC

cows riding motorcycles isn’t very common in photography. Firefly 1 didn’t give a great result; it’s forced and unnatural, the ‘golden hour’ lighting isn’t very strong, and it doesn’t give any feeling of freedom and speed. Firefly 2 improves the composition, and the lighting is much better, but it’s still not a great, coherent image. Again

Midjourney
ORG

gets this spot-on; creative combination, strong lighting, and emotive.

The improvement of

Firefly 2
PRODUCT

over Firefly 1 is clear, and it’s now competitive with

Midjourney
ORG

in some cases; but it’s still a lot less creative. However this is a far from comprehensive test, and I’m sure as I use Firefly 2 more I’ll discover its strengths.