This is an AI image generation comparison for
reference-to-image 
prompt:
Medium shot of this woman from @image1 in white t-shirt with "AI creators tools" print and baggy purple silk pants standing left of center in this street from @image2. To the right, the storefront reads "AICREATORS.TOOLS', red bench underneath. Also in the backdrop a high-tech modern silver-metallic... tone statue is seen holding red viny plant, its petals scattered underneath....
Log in to see full prompt.
Tested: September 16, 2025
Kolors 2.0, now Image 2.0 with two references. Still cartoonish look.
Tested: September 16, 2025
Pretty good result. With more tries could likely gert it perfect, except the t-shirt print text
Tested: September 16, 2025
Nano Banana does everything right in this one.
Tested: September 16, 2025
Flux Kontext in Freepik at least tends to do these noisy faces for complicated prompts. Mind you, multi reference is still in Beta for this model. Overall great prompt following, details preservation and text handling
Tested: September 16, 2025
Even if asked for closeup, Flux Kontext Max, in Freepik at least, tends to do these noisy faces for complicated prompts. Note, multi reference is still in Beta for this model. Overall great prompt following, details preservation and text handling
Tested: September 16, 2025
That's wide shot not even medium, and I've asked for loose closeup. Our woman isn't looking so hot. But there is likeness. And backdrop is preserved well.
Tested: September 16, 2025
Now we're talking! Following up in the same chat on my base prompt, I've asked the assistant: 'can you try again and this time make sure woman is closer to the camera, framed from head to waist?'
Tested: October 10, 2025
Prompt with 2 images to combine. Not impressive tbh. Subject is completely out of focus.
Tested: October 11, 2025
It's not bad for an open-source model. I've made the subject the 1st image though and I should've placed her in the 2nd slot so I'll retry.
Tested: October 11, 2025
When I switch places for img 1 and img 2 the model loses face reference completely and also makes funny eyes
Tested: October 14, 2025
Used the default prompt from demos and a full body shot of the woman for this test, and it does better. But the superimposed woman is only very similar, lots of details are modified. Background image is untouched so 100% same.
This is a test with 2 images used as subject + setting reference. The AI models should preserve character's likeness as much as possible AND the environment characteristics. The subject's photo is a high-resolution closeup portrait.
Check out the results from Vidu AI (Vidu Q1 Image) vs Kling AI (KOLORS 2.1) vs Runway (Runway's Gen-4 Image) vs Freepik (Gemini 2.5 Flash) vs Freepik (Flux Kontext [Pro]) vs Freepik (FLUX Kontext [Max]) vs Reve (Reve Image 1.0) vs Reve (Reve Image 1.0) vs Fal AI (Qwen-Image-Edit 2509) vs Hugging Face (DreamOmni2 Edit) vs Fal AI (DreamOmni2 Edit) vs Hugging Face (DreamOmni2 Edit) for similar or identical prompts side-by-side.
Girl with pearl earring realistic photo with text
Tourist photo with photobomber
