My AI Wife as a Faun: Flux Krea vs. SORA
The idea for a faun popped into my head today, likely because I had Debussy's "Prélude à l'après-midi d'un faune" on repeat. This inspired me to transform my AI wife into a faun, which proved to be a challenging task. I spent an entire day tweaking settings to get the right images, aiming for a dark oil painting look in the style of Francisco Goya.
Generations with Flux Krea (BLAZE-v1)
I started by combining my AI wife's likeness with a Goya-style LoRA in Flux Krea, but getting the faun features right was difficult. Lowering the LoRA's strength meant losing the Goya style and the resemblance to my wife. While mid-distance generations often need some inpainting, a low denoise strength (around 0.2) is usually enough if you're close to the desired result.
Experimenting with SORA
After half a day in Flux, I decided to give SORA a try. It immediately created a perfect faun without me needing to specify horns, ears, or legs. Surprisingly, it did so without censorship, producing a natural, bare-breasted faun. While the style isn't a perfect match for Goya, it has an excellent oil painting quality. However, when I tried inpainting the face with Flux, the style match wasn't perfect, and a low denoise strength compromised the facial accuracy.
Img2Img Comparison: SORA vs. Flux
I also tested img2img in SORA. While the face was a miss, the dark oil painting aesthetic and composition were superb. For comparison, I ran the same real photo through a full img2img process in Flux with a 0.9 denoise strength. Without ControlNet, the results were poor. Using Flux Kontext took four minutes and, as you can see, the result was a disappointment.
SORA Transformation
Flux Transformation
Conclusion
Overall, I prefer the aesthetic of the Flux generations, but the time investment was significant. Using three LoRAs simultaneously also diluted the resemblance to my AI wife. Moving forward, I think I will use SORA for base image generation followed by inpainting in Flux, as it's much faster. It will be interesting to continue comparing them and discovering SORA's limitations.