It’s the little details that bring everything together.

Can AI really juggle 10 REFERENCES? Let's put it to the test!! 👀

👁️ Goal: Create a Halloween poster: a girl doomscrolling in bed while Munch's Scream screams right behind her, and she doesn't notice, the title is "Screen time: 7h 42m. Time left: 0.", while mixing drastically different styles in my refs (anime, baroque chiaroscuro, Munch's expressionism, pixel art). ⚒️ Model: FLUX 3 Image Klein 🌪️ Process: I generated my refs (I love complicating my life yes) and garthered the original paintings, fed it to Flux with a detailed prompt on how to use each of the ref I uploaded.

I uploaded the two best results I got, but it was a few back and forth to get every detail correct! My rating on how the model respected each of my refs (average of both outcomes):

- Palette 5/5 - The Scream 3/5 - De La Tour lighting 3/5 - Paper texture 5/5 - Face 5/5 - Pose 5/5 - Velvet & linen 3/5 - Sky 2/5 (swapped my swirls for a sunset on the first one) - Title 5/5 - Sketch 4/5 (because why was she filled on the second one!!) ✨ Overall rating: 4/5

✍️ Takeaway: Keep the refs very specific and your prompt very detailed. The model doesn't understand "vibes" (the paint swirls became a sunset) only but is very good at following precise instructions (pixel title was perfectly copied), unlike MidJourney for example which is very good at interpretingvibes.

Do you prefer outcome 1 or 2? 😱