I haven't had any luck making a good-enough FP8 safetensors version of this fun little merge yet, but I'm still working on it, and will share if successful. This is an Int8 ConvRot merge of the latest version of Boogu Image Edit Turbo, plus a little something extra to help with unlocking its fuller potential. For purely artistic reasons, of course. <a target="_blank" rel="ugc" href="https://huggingface.co/Boogu">Boogu is actually a pretty impressive little family of models</a>, and I like using this merge, so I figured maybe share, and help raise awareness.I've tested each flavor in various ways. The Image model can edit, the Edit model can generate, and both come in Turbo versions, plus you can even use the Base version with a Turbo LoRA to get better diversity with detail and composition. It's a groovy thing. I tried the turbo editing model, liked it enough to look for a way to get it to do just a little more natural beauty (boobies) too, and so here we are. I daresay Boogu can outperform Klein 9b, and I didn't expect anything within my grasp to do that so soon. It's a pretty smart model for its size class, especially if you already know how to talk to a Qwen3 text encoder.Boogu is a 10-billion-parameter model, so it's a little bit heavier than Klein 9b, and it uses a Qwen3<a target="_blank" rel="ugc" href="https://www.google.com/search?q=qwen3vl+what+does+the+vl+stand+for">VL</a> 8b text encoder. I prefer the Heretic/Abliterated version, just because otherwise it can balk at some prompts (it's extremely sensitive and prudish for a text encoder, but then I guess it's seen a lot in its time). And I know I probably just lost everyone in the "potato" crowd as soon as they saw "10b" and "8b", but hear me out. I'm practically on a potato, too. I'm using a 3060 ti with only 8GB of VRAM, and I'm able to use this and it's as fast as some of my other smaller models.Despite the diffusion model and text encoder each being almost 10 gigs in size, I get generation times of around 2-3 seconds per step, and editing times of about 5 seconds per step, and I get usable outputs at only 4 steps. Thank you, ComfyUI, for native Int8 support. So, even with a smaller GPU, this might work for you, too. And apparently, you can go as low as 3 steps. You can use as many as 8 or probably 12, but I usually just use 4 steps.I favor this one particular version of the model, since I use Boogu more for editing than generating, but it does a fine job of both (all samples were generated using it, and then some were edited). If it had more LoRA support already, this model would stand a fair chance of putting my beloved Klein on a shelf. All those hundreds of LoRAs, all put away in the old archive SDD, and I haven't even tried them all yet D:Much like Flux2 Klein, editing tends to produce a little bit of color shift (depending on what's being added/taken away/changed), but you can mediate most of this (and prevent staticky artifacts) with just the denoise settings. I find a denoise of .87 works best for me, but you should always experiment on your own, as well. It probably helps to use more steps than I do, also, or do a detail-fix upscale pass, or something. It's still new, and I'm still learning it, too.I'm including my own personal-use copy of a Heretic/Abliterated Qwen3VL 8b text encoder (and I can't recall where I got it or I'd provide a link and credit), along with the Flux1 VAE, which you might need to rename to "Boogu_VAE" if you are using my sample workflow or any of the sample images' workflow innards (or just change it in the workflow). The VAE link in my listing goes to the official Civitai page for Boogu-Image Base, but just download the VAE and rename it (if necessary) and make sure your models are in the right folders, and you should be good to go.I'm also including a sample workflow, along with a pack of blank image latents to use with it. I like using blank images, because then I can simply use the same workflow for genning and editing, and I never have to remember which multiples of 8 or 16 to type into that stupid blank latent node this way. Drop one into the Load Image node and you're all set.I've set the workflow's denoise to 0.87, because this seems to work best when editing, and helps prevent some of the artifacting and color drift that can otherwise occur. You can set it to 1.0 for generating images, but always set it lower for editing. You can experiment with this and see if you can get it to work better than I have. Almost all of my sample images were generated with a denoise less than 1.And lastly, though this isn't really a NSFW model (it's general-purpose), it's capable of doing some NSFW stuff, so please prompt responsibly. If you generate or edit wrongness into the AI-osphere it will all be on you anyway, so do think first. I'm just giving you a tool, it's up to you to use it responsibly.You can thank <a target="_blank" rel="ugc" href="https://civarchive.com/user/herpderpmerp">this guy</a> for the added bits, though he claims no credit in his listing. Give him some credit (or Buzz) anyway, because thanks to him, this model knows how to do bewbies. This is a fun as well as useful model, and I'm looking forward to seeing it get more attention in the future. It would be nice to see it get more LoRA support from the community.Boogu is licensed under the <a target="_blank" rel="ugc" href="https://www.apache.org/licenses/LICENSE-2.0">Apache2.0 license. </a>So yay~! Maybe that will help.And eventually I'll get around to writing a simple guide for all of this crap. I know it can be confusing for the less-initiated.

Boogu Image Edit Turbo Int8 ConvRot

Files with SHA256: 6e505b11c321ad455b3f503bb8422f2e3ab50c45741196244e7f80f63e35fc2b