Something positive? Anatomy is VERY good. Other things too.
I'm not talking about NSFW, that has to come from the community as usual. I'm talking arms, hands, legs, people in close, complicated, multi people interactions and so on. This is possibly the best at that i've seen in several months.
This model, coming completely out of the blue, it is VERY good at anatomy. and geometry in general. Even in complex multi people scenes with multiple interactions, where Flux.2 Klein 4b and 9b and even dev may make already a mess of limbs, i've only see a messed up arm once or twice.
I'm using mostly the editing model. NOT for editing images (flux2/QE2511...) are much better at that. I use it basically as general image model, just with reference images for people and so on. For completely newly prompted (complex) scenes just with the reference person/clothing or whatever involved. Which works well enough.
But turbo t2i works equally well.
Hands even in fairly complex situations are mostly alright. And in simpler situations they are always alright. Arms and legs are basically always correct. Weird clothing glitches and alike are also rare. This is well beyond Flux Klein4b and even 9b in Anatomy, (but MILES off them in image quality!).
Not to mention Ernie image, which couldn't make a correct person just standing there. Also Boogu is NOT extremely biased towards making people look chinese. It feels much more versatile and flexible.
All of that with the caveat that it needs for example a zit image upscaler (template in comfyui) to fix the glaring texture and (Lizard) skin weakness. But Z-image upscale/improve works a charm as so often.
Why not just use Z-Image Turbo? Boogu seems to be better with Text, Better with variety (use the base model+lora). Skin may not look as photo realistic as inKlein9b or Zit, but it feels more capable in handling and following complicated prompts (i.e. complex scenes prompted). Even with many moving parts+text and in normal language prompts, it gets things usually right. While Z-image turbo especially tends to throw it's hands up when it gets too much.
So despite the image quality issues, this is a good base! With that qualifier of needing some postprocessing, this Boogu (0.1!) is definitely the best recent model (Ernie, HidreamO1, Longcat image, ideogramv4). The editing version is not brilliant but ok at keeping people or styles or clothes identical, often it is more of a ok approximation. But it can handle very complex multi person prompts with text and stuff, very well. And it doesn't need short story as prompt! OR a ridiculously complicated json prompt.
I hope they keep at it, fixing the low texture and skin quality, improve the persistence from the editing model. But this is actually promising, now That Qwen image 2 for example has moved behind the paywall.
Training:
Sadly It trains considerably slower than HidreamO1. Which is quite good for some things but very blurry in skin and stuff. So for the quick person lora, that remains my new goto. It also needs an postprocessing upscaler (against the ridiculous low res blur) but the training speed is insane. Not just the it/sec but also how qickly it "gets it".
While it is also better with reusing people from reference images in new scenes. But not as flexible/creative or ...good at creating new scenes