I've been testing Boogu-Image Turbo since release and built a workflow around it. Sharing here with the settings that actually worked for me, in case it saves someone the trial-and-error.Quick take on the modelThe realism is solid. Details, light and shadow are handled well, and in a few generations I noticed some really natural reflections and highlights I haven't seen from other models lately. Worth a serious look if.Settings that actually workedThe official recommendation is 4 steps, CFG 0. In my tests, CFG 0 produced random outputs that ignored the prompt completely — could be a sampler/scheduler interaction, but it was consistent enough that I'd flag it. What worked for me:Sampler: LCMScheduler: bong_tangentCFG: 1.0 (above this it got worse, below — see above)Steps: 5 (4 works, 5–6 gave me a bit more stability)I tested the fp8 variant.What's in the workflowA clean Boogu Turbo generation branch, plus an SDXL refiner branch that only runs on regions detected by UltralyticsDetector — not on the full image. Every branch has a one-click toggle, so you can pick exactly what you need per generation.Detector nodes included:nipples — for soft NSFW. This is the one I actually keep on. Boogu handles the rest of the body fine on its own, but this detector + SDXL pass cleans up the result nicely.face — left in for completeness. In my tests Boogu's faces are already good, so the refiner pass rarely adds anything.eyes — same as face. Marginal in my tests.pussy — honestly, leave it off. Boogu doesn't seem to understand the anatomy here, so the detector ends up trying to refine something the base model already broke. No detector is going to save that.Most of my generations are done with the refiner branch off entirely — Boogu Turbo alone is fast and clean. I only flip the nipples branch on when I'm doing soft NSFW.Heads up — Empty Latent sizingOne thing I haven't figured out yet: to get an output at a given resolution, I had to set the Empty Latent width and height to double the target — e.g. 2560×3840 in the node to get a 1280×1920 output. Not sure if this is a Turbo quirk, a Flux2 latent thing, or something in my own setup that I'm missing.If you've run into the same thing or know what's going on, drop a comment — I'll update the workflow once I figure it out.RequirementsBoogu-Image Turbo checkpoint (I used fp8 — bf16 and nf4 should work the same)Qwen3-VL 8B text encoderFlux VAEAn SDXL realistic checkpoint for the refiner branch (pick your favorite)UltralyticsDetector + SAM models for the detectionrgthree-comfy nodesCreditsFull credit to the Boogu project , original work by their team, mirrored on Civitai. Project page and weights: github.com/boogu-project/Boogu-Image.Workflow put together by ShinobiSat. Comments and corrections welcome — still learning what this model can do.

Boogu-Image Turbo + Refiner — V1.2

Settings that actually workedThe official recommendation is 4 steps, CFG 0. In my tests, CFG 0 produced random outputs that ignored the prompt completely — could be a sampler/scheduler interaction, but it was consistent enough that I'd flag it. What worked for me:<ul><li>Sampler: LCM</li><li>Scheduler: bong_tangent</li><li>CFG: 1.0 (above this it got worse, below — see above)</li><li>Steps: 5 (4 works, 5–6 gave me a bit more stability)</li></ul>

Files with SHA256: 1a690b90e8d1c7c6d9f15c611a0ce190763fe616ffe0fb42f000889afa03463d