A Very Different Comparison

Introduction I've done a lot of work with women, and most of the time I compare models, I use images of women to do it. The two previous blogs were that sort of thing. This time, I thought to do it with something non-human, so I chose a tree. This is the tree of forgotten lore, which is carved into it's surface. Prompt A forgotten lore carved into the very fabric of a massive, ancient tree. Intricate runes and symbols glow with a faint, internal light, pulsating with dormant magic. The tree stands in a serene, moonlit clearing, surrounded by ancient standing stones. The style is dark fantasy concept art, with photorealistic rendering of the bark and glowing inscriptions, and a mysterious, epic atmosphere. The Results of this comparison may surprise you. Contents - GPT 1.5 High / Medium / Low --- GPT 1 High / Medium / Low - DALL•E 3 - Flux 1 - Pro Ultra / Pro / Dev / Krea / Schnell - Flux 2 - Max / Flex / Pro / Dev / Klein - Flux Kontext Max / Flux Kontext Pro - HiDream I1 - Full / Dev / Fast - Hunyuan Image 3 --- Imagen 4 Ultra - Ideogram v3 - Quality / Balanced / Turbo - Lucid Origin - Ultra / Standard --- Minimax Image 01 - Nano Banana Pro / Nano Banana - Qwen --- Z-Image Turbo --- P-Image - Seedream - 4.5 / 4 / 3 - Wan - 2.6 / 2.5 / 2.2 Comparison GPT 1.5 High / Medium / Low --- GPT 1 High / Medium / Low I've been very impressed overall with the GPT 1.5 series of models. This comparison continues to show off its versatility. The same can not be said for its predecessor, GPT 1. GPT 1 has always leaned towards a darker, hazier image. It isn't intuitive how to make that go away,and even Chad had troubles with it. It's supposed to have been a more painterly model, applying a default style to everything. I have in the past made some successful images with it,but not very often. One other thing I noticed, comparing the two is that all levels of GPT 1.5 produce beautiful, complete images. With GPT 1, Low was almost always unusable. It was intended to let you test prompts at a lower cost, and Medium was the low end image generator. DALL•E 3 This model has always been the one to use to get fantasy styled imagery. It's something it excels at,and even today,it's hard to beat in some areas. GPT 1.5 has surpassed it in many ways, though, but it still manages to pull some great images,like this one. I would say that in this comparison, DALL•E 3 beats GPT 1.5. Flux 1 - Pro Ultra / Pro / Dev / Krea / Schnell In thisround of images, it's interesting to see Dev making a nice image. Pro Ultra's interpretation is a little weird, so I'm not crazy about it's image. As far as I'm concerned, Pro has the best image. Schnell's image isn't bad, either. I like it. Flux 2 - Max / Flex / Pro / Dev / Klein Max is, in my opinion,the best image in this group. Flex, which I usually prefer, made a less than great showing. Pro is in my opinion, a weak model, and it shows here. Dev, on the other hand, did pretty nicely. Klein, the new kid, made a lovely image. It might not be quite as nice as Max, but it is very nice, and when you factor in the incredibly cheap price, it's the real winner here. Flux Kontext Max / Flux Kontext Pro I don't know what tosay about the Kontext models. They were supposed to be the next great models, but they're not. These images look good, especially Max, but they don't compare favourably to Flux 2. HiDream I1 - Full / Dev / Fast When you look at allthe other models, including the low end models, HiDream is just not comparable. They do some things nicely, but they are boring. They aren't as bad here as they were in the last comparison, but they aren't great. Hunyuan Image 3 --- Imagen 4 Ultra Hunyuan is clearly one of the best models in this comparison, it proves to be every day. It always makes high quality images. Imagen,on the other hand, often makes very nice images, but missed the mark here, in my opinion. Mostly just interpretation, I think. Ideogram v3 - Quality / Balanced / Turbo I don't know what to say about these three images. I have seen Ideogram make some great stuff in the past, but lately they've been unimpressive. Lucid Origin - Ultra / Standard --- Minimax Image 01 In my personal opinion, none of the images here are very good. What I can say, though, is that @Cheinia's blog about Lucid is spot on. THe Standard model doesn't make a great picture, but it definitely let's you see what the real image might look like when you make it in ultra. Use it to iterate until the prompt is how you want it, then remake it in Ultra for a final image. Nano Banana Pro / Nano Banana I really love the Pro image. I think it turned out quite good. This standard image isn't my cup of tea, but I can still see that it could be pretty good with a different seed. Qwen --- Z-Image Turbo --- P-Image I think they all turned out pretty good, but expecially Z-Image Turbo and P-Image. Qwen seems a little low-effort. Seedream - 4.5 / 4 / 3 4.5 Is one of my favourite images in this comparison. I think 4 and 3 did decently, but not as good as 4.5 Wan - 2.6 / 2.5 / 2.2 I wanted to get 2.2 in here as well. I have the image. Unfortiunately, The blog currently has a max image cap that I just crossed. The 2.2 image turned out quite nice, though, so hopefully, I will get to show it somewhere. All in all, Wan made some nice images. I like all of them. Final Thoughts Comparing images of people is hugely different from comparing non-human things. It is clear to me that where people are concerned, the models vary more than they did here. I hope you all like this comparison. These are my favourites, in roughly that order. 1. DALL•E 3 2. GPT 1.5 High 3. GPT 1.5 Medium 4. Hunyuan 5. Seedream 4.5 6. GPT 1.5 Low 7. Nano Banana Pro 8. Z-Image Turbo 9. Flux 2 Max 10. Flux 2 Klein 11. Flux 1 Pro