<h3 id="version-2">Version 2</h3>Welcome to V2. I've updated this because I heard you could get better results by combining these elements instead of the default vanilla ones. I've combined the fp8 Krea2 turbo diffusion model, Huihui's abliterated Qwen3VL 4B-Instruct, and the Wan2.1 VAE (larger than the Qwen-Image VAE but still compatible), along with the Krea2 filter bypass, just for good measure. I used v1 (2-layer) of the filter bypass because during testing v2 (3-layer) was adding small details I didn't ask for, and v1 didn't do that. If you like the 3-layer bypass better, I'm pretty sure you can still safely append it and get the results you want.I didn't do any of this to make the model spicier in content, but rather because it frees the model up to handle prompting a little better, and I like the visual results I'm getting with it so far. Images seem a little less "staged", without having to use the raw model with a turbo LoRA to get more spontaneity.If you do spicy stuff, it should work fine, but you'll probably still need your smexy LoRAs.I'm looking forward to what these guys put out next, and I can't wait to see if they make a model that can do editing. I finally get good at generating with Klein, and then this thing comes along and knocks the weeny out of the bun, lol. If they make an edit version, my Klein might end up on a shelf.Drag images to ComfyUI to use the simple workflow I made them with. You'll probably need to update the checkpoint loader (it's inside the purple subgraph), as I renamed the file for upload. 8 steps is the recommended number for the turbo version, but, much like some other turbo models I know of, I even get surprisingly good results with as few as 4 steps, and it takes far fewer tries to get images I'm okay with.If anyone wonders why I chose fp8 rather than some other quant (to get a smaller filesize), I find that fp8 just works better and faster on my 3060 ti (8GB VRAM) than GGUFs do (at least since some ComfyUI update in May, I think it was), and I have yet to find a smaller quant of the text encoder that isn't a GGUF. I also hear that the BF16 text encoder is far superior, but I've settled for fp8, like with the diffusion model.The latest ComfyUI gives me times of about ~3 seconds per iteration/step with this checkpoint. LoRAs make it a bit slower, especially large ones, but I have yet to see anything over 4.-something seconds/iteration. Your mileage may vary, as no two computers are made the same anymore, and for all I know I got lucky.<h3 id="version-1">Version 1</h3>Just a merge of the fp8 Krea2 Turbo diffusion model, the Qwen3VL 4B fp8 text encoder, and the Qwen Image VAE, thrown together for my own convenience more than anything else. I like checkpoints. This model is fantastic and I love it, and this checkpoint runs in the latest ComfyUI with only 8GB of VRAM (which is what I have), spooting out images at ~2-3 seconds per step/iteration. Works fine at 8 steps, and can handle more.I've included the basic workflow I've been using for the past few days, but you can always just drag the sample images into ComfyUI as well (though the animated one is of course a Wan video). If you don't have or want rgthree nodes, just swap out the LoRA loader or something.A 12-billion-parameter model that runs as fast, or faster, than Z-Image Turbo used to run for me, I can't help but like it a lot, and I like how much support it's already gotten from the community at large, and I love the quality of the outputs this model is capable of.I also like how, with the right settings, the body horror is rare enough to actually be startling. I get runs of upwards of 20+ images without having to cull for freakshow anatomy. Everything old is new again, lol. Nothing's perfect, but this thing is great. Just know that some LoRAs may produce strange visuals with some samplers. XDI hope that you like it as much as I do, if you take the time to download it. This particular merge has no native/built-in NSFW, because I prefer to merge basic, "pure" models without all the extras, but whatever other people want to do with it is up to them. Naturally, <a target="_blank" rel="ugc" href="https://www.krea.ai/krea-2-licensing">the Krea2 license</a> applies, and I make no claims to anything or changes to the license. I'm just a messenger, sharing something I've found that's useful and fun. And awesome.

My Krea2 Turbo Checkpoint

People said that incorporating a bypass LoRA and using an abliterated Wan3VL 4B-Instruct and throwing in the Wan2.1 VAE would improve image quality, composition, and prompt adherence, so I tried it, and it actually seems to work. I like the results I've gotten so far anyway, so I'm sharing v2.

Files with SHA256: 733403eda39ea1d77cc1cf1923e3ce4fd8ce33e4c1eb064b2792fde3412c6fed