New -> Flux2Klein9b ULT PRO
SAMPLER: seeds_3 / exp_heun_2_x0_sde / Euler / Whatever works for YOU
STEPS: 4-20 (YES 4-20 - you control how much detail/veneer you want)
CFG: 1-3 (YES 1-3 - you can use NEG prompt now and improve overall consistency)
RESOLUTION - UP TO 1280x1920
LESS DEFAULT PUBIC HAIR (use dedicated LORA and/or prompting when needed)
can generate those male thingies now (on women too)
works GREAT on img2img - no need for NSFW/"make real" LORAs anymore
Previews in gallery look way worse than the real images!
How to see >REAL< quality (examples: >Flux2Klein< >ZIT< ):
Right-click image → Open image in new tab
In new tab: again Right-click image → Open image in new tab
Click image - it will magnify to it's true resolution
Idea behind Vintage Beauties models:
Real-looking hairy bodies that feel like actual film photos - grainy, soft, a bit imperfect and alive. Visible film grain, gentle glow in bright areas, creamy shadows, vintage lens quirks (like halation and swirl), plus subtle color shifts. Nothing clean, sterile or digital. Just raw, organic, breathing analog vibes. Experimental merges, so not perfect.
No cherry-picking or upscaling/refining - simple workflow for straightforward evaluation.
Each image in gallery was generated using the same seed.
Workflows and prompts are included in the images (dl and drag to ComfyUI)
Flux Klein 9b v2 is based on FLUX.2-klein-base-9B
Flux Klein 9b v1 is based on FLUX.2-klein-9B (distilled)
Both models work with txt2img and img2img
Both models work with LoRAs (tested with up to 3 + sliders)
Settings for FLUX.2-klein-9B V2 (model for rich natural detail that looks best at ~2 MP ~ full HD)
SAMPLER: exp_heun_2_x0_sde (top choice for this model)
STEPS: 20-30
CFG: 1 (yes, ONE)
Settings for FLUX.2-klein-9B V1 (model for quick generations, but quality is still good)
SAMPLERS: er_sde, res_multistep, dpmpp_2m_sde, seeds_3
STEPS: 4 - 9
CFG: 1 - 1.4
img2img workflow: click here
Previous Models:
VintageBeauties ZIT (based on Z_Image_Turbo_bf16)
Sampler/scheduler combos to try:
dpm2_ancestral / sigmoid_offset (great contrast and shadows, depth, artistic)
er_sde / ddim_uniform (gives analog look to photo without LORAs)
lcm_custom_noise / linear_quadratic (gives that special render'ish SD 1.5 artificial look)
Krea:
Play around with the Guidance values between 1 and 13 — though going as high as 20 also works. Weird, but true.
Recommended settings:
Sampler:
dpmpp_2mScheduler:
betaSteps: 30
CFG (starting point): 5
Flux:
(v1 and v2) use at least 4.5 STR.
P.S. Use STR 3.5 for comfy backend in swarmui - thanks @snotty for the tip!
Description
Same Vintage with Flux Krea acting as a sidekick. Expect the unexpected!
FAQ
Comments (10)
there is a diff in weights between v2.0 & v3.0.
I get an error with v3.0 for float 8 that is not supported under mps Mac OS.
v3 is based on Krea, so if FP8 Krea works for you, this one should too.
Thank you for your answer - I am new in all this - I wonder if it is easy to make a fp16 version?
ioannismastoris1965 No problem. Fp16 wasn't possible due to gear limitations. But, you can try flux.krea https://civitai.com/models/618692?modelVersionId=2068000 with my LORA https://civitai.com/models/940109/vintage-beauties to get results that will be pretty close to this model.
Partisano / thank you
can i use it on forge ui?
Can't tell for sure. But, if flux.krea fp8 works for you then this one will too.
forge neo i guess
@dark_aromaa It sucks
@qek haven't tried it. i will later and see if it is any good



















