Something positive? Much, especially Anatomy is VERY good. I use it a lot now.
I'm not talking about NSFW, that has to come from the community as usual. At least, If that means simply nakedness. But because of the below it is quite good for example with the anatomy and geometry of kinkyness* without messing stuff (Arms, legs hands) up as much as others. It is flatout better at this stuff than Flux2 Klein. Even Flux.2 dev. Maybe even Qwen Image.
So i'm talking arms, hands, legs, people in close, complicated, multi people interactions and so on. 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.
This is ( in THAT aspect) possibly the best at that i've seen in several months.
edit/update to "Why not just use Z-Image Turbo then?": TL;DR Because Boogu is actually better (Boogu Fp8 base+lora >> boogu turbo model) . With only minor qualifiers.
How did it happen? Since the image quality has improved (base+hotfix turbo lora), i noticed the upscale result wasn't much better anymore. So i used my ZiT upscaler in the workflow for some tests. I set denoise to 1, steps to 9 and give it the same prompt. So basically a complete z-image run with the same prompt. Easy to compare every image ran. (Z-Image Turbo is so fast, it is no problem to run both).
My result: Boogu (base+lora) is hands down better in EVERY aspect except (slightly) in image quality, than Z-Image. Almost every result. The only thing Z-Image is better is it's trademark hardcore nitty gritty realism look. But it also tends to overdo that.
after the patches even the image quality is actually ok. So Boogu has sneakily slowly become my most used image gen model. for all but precise editing. It's easy to use and good. No weird prompting nonsense needed, even simple, un enhanced prompts work just fine (base+lora!). It just works.
Boogu is much better at understanding the prompt, it is much better at placing text. At placing people. While Z-image is quite ok with anatomy, even it messes things up more than Boogu. Also Z-Image of course has no editing function at all. Not so convinced of the Boogu Turbo model. Maybe a settings thing but it lack in image quality. Since i can ran base+lora very fast, i do.
I'm using mostly the editing model. NOT for simply editing images (flux2/QE2511.. are 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. Same works of course For Flux2, QE and others). Using them just to edit an image is a criminal underuse.
Likeliness in Boogu is not always perfect but works well enough. And because it is so good at anatomy it is good at putting reference subjects in different situations.
So for that reason it's very similar to the non edit model. When using non edit, i use mostly base models+turbo lora. lora 1.0, 10to maybe 14 steps, pretty good image quality.
little update on the edit model for actual editing:
The results are a LOT better (i.e. closer to the prompt) without the speedup lora. But then it takes of course much longer. But again, in challenging multi people interactrions it (then) messes up anatomy less than Qwen (things like additionally hands placed somewhere nonsensical...) Boogu seems to be better at geometry and spatial understanding.
Generally: 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 still off them in image quality!). In precise editing tasks that change only what is wanted, flux2 (klein) is also better.
Of all the recent models i know, Boogu feels the most underrated.
- ideogram needs a fricking script as prompt. WAY too much hassle. and doesn't seem to come with much benefit. Except maybe for text heavy infographics and stuff.
- Ernie image is just ... shit. It couldn't make a correct person (2arms, 2 legs, 10 fingers...) just standing there. And it turns everybody chinese, no matter what you prompt. It also feels vastly less flexible and capapble than boogu in general.
- hidreamO1: looked interesting, but extremely low detail and blur in large ares (skin, clothing texture) , only useful with heavy postprocessing/upscaling. too much hassle while no advantages (ok lora trainign is FAST). .
- longcat image not really tried. Seems very fast but nothing special.
- there is a joyai one coming to comfy main (hopefully) that looks interesting because it is actually large, while everything else seems to go the XB turbo mode. Hoping for a new qwen image edit, but likely won't get it...
With Boogu being officially 0.1, i hope it gets better over time without messing up what they already have. this is a good base!
Training:
Sadly Boogu trains considerably slower than HidreamO1 (which is the fastest training (modern) model i know). Arguably Hidream images are so low detail in textures and stuff, that's probably why it learns so fast. But HidreamO1 is quite good for some things, like using people as reference images, and because of the fast training ( it/s AND iust 200-400 steps needed) ), for the quick person lora, that remains my new goto (to then be heavily post processed.
HidreamO1 it is not remotely as flexible/creative or ...good at creating new scenes as Boogu though, so it can only do (relatively) simple.
*Talking of kinkiness: unlike any model but Chroma1Hd, it knows what handcuffs look like :P