Socamerge Turbo XL - 1.2 (Hybrid)
CRUCIAL INFORMATION: 1-3 CFG, 5-10 Steps, DPM++ SDE Karras
This model produces photorealistic images of pretty women, but it'll give nice results for all ages and genders. This is a merge of SDXL Turbo with my Socamerge Realism model, a fistful of sexy women loras, Tertium, DreamShaper, Ishtar's Gate; and with 1.2, Hephaistos and RealVisXL.
Sampler: DPM++ SDE Karras or DPM++ SDE. DPM++ 2M SDE Karras, DPM++ 2S a, DPM++ 2s a Karras will work with more steps. Euler a will work (and be faster) but has much less detail. Other samplers might work with higher steps (~20). This is a legacy of SDXL Turbo.
Steps: 6-10 As a hybrid SDXL turbo merge, it moves away from the 2-5 steps of SDXL Turbo. This is an intentional choice, as my testing showed the increase in image quality from merges that used less Turbo and more Socamerge Realism to be worth the extra steps.
CFG: 1-3. SDXL Turbo requires MUCH lower CFG than normal SDXL models.
This model will not work at low steps with many samplers, and as an evolution of SDXL Turbo requires a different approach to steps and CFG. It is critical you follow the above guidelines or your images will be burned, distorted, or poor quality.
-Resolution Recommendations: 1024x1024, 896x1152, 1152x896
-Make sure to turn OFF Face Restoration, SDXL Turbo models actually lose image quality/definition from it
-I strongly recommend ADetailer to make the faces better: search for ADetailer in the WebUI Extensions panel or go to: https://github.com/Bing-su/adetailer
NSFW Prompting:
- (Nude naked NSFW) in positive prompt = very naked people. This checkpoint merge is extremely good at anatomically correct NSFW for cis women and trans men with prompting, and with trans women and cis men with an appropriate penis Lora (which work surprisingly well with this checkpoint)
- No (Nude naked NSFW) anywhere = a bit on the skimpier side for clothing.
- (Nude naked NSFW) in negative prompt = more conservative clothing, can reinforce descriptions of clothes that cover-up more.
This checkpoint has a VAE baked in, so be aware of that if you do further merges (just make sure to bake a vae in as you merge). Automatic VAE selection will work just fine when you generate images with this checkpoint.
Special thanks to FrenzyX, Lykon, Afroman4Peace, and SG_161222, who have done most of the hard work stabilizing SDXL Turbo and getting it to produce acceptable anatomy and proportions at 1024x1024 and related resolutions.
Civitai Generation is turned off, due to the usage requirements of SDXL Turbo. Should this change, I will enable it.
Description
Additional Merges
FAQ
Comments (8)
Update 1.2 is here! Now with better prompt following, colors, and details thanks to the new merges included.
Turbo models still have issues with long torsos so I suggest (long torso:1.5) (stretched torso:1.5) and/or (tall torso:1.5) in the negative prompt.
As a turbo hybrid model, negative prompts still have a SLIGHT effect, but must be emphasized.
It seems like this model is perplexing. When I set the steps to 10 and CFG to 1.5 as recommended, and generate an image, the result is excessively noisy and produces distorted images. I tried copying the data of the female image you provided in the sample, but the issue persists. I also tested another SDXL Turbo model to check if there's a problem with my system, and it worked fine. Are there any specific settings required to use this model?
I see now VAE is baked in thanks it looks great gonna try it today.
I'm getting progressively more and more washed out colors as I run an image through img2img with v1.2. Possibly a VAE issue? Any recommendations?
This model is superb for realistic photographic portraits, only the eyes are not on the same level as skin texture. I hope it will be updated furthermore :) thanks for the model and your work!
This has become my default XL checkpoint for photograph pinups. Very solid.
Great Model
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.