*Showcase images include embedded workflow and are post-processed with sharpening and noise grain.

In other words, your results may vary depending on your workflow.

Recommended settings are below.

Version features

Rather than improvements, each version has their own strengths.

• v1.1 - Most realistic

• v1.2 - Semi realistic - sometimes include outline

• v1.33 - Slightly less realistic than 1.1 in exchange for better lighting/composition/prompt adherence.

Version notes

v1.33

• Focus on keeping realism while improving composition, background and lighting compared to v1.1.

• Less creative/aesthetic than 1.2 in terms of background.

v1.2

• More focus towards semi realism *includes outline at certain times

• Better lighting, composition, background, aesthetic

• Structure feels more stable compared to v1.1

Model focus

• 3DCG features with more detailed skin texture. (Heavily dependent on heun sampler, not recommended if using web generator or if you don't know what you're doing)

• Face proportions closer to real life. Less anime like.

Strength/Quirks

• Facial features can respond to prompts

Weakness

• Will do beauty mark/mole under eye

• Skin pores can look too smooth/uniform/unnatural texture depending on sampler/scheduler.

• Backgrounds (res_multistep + normal with high steps 50-100 - has chance to resolve scale issues but might introduce artifacts) (better in version 1.2 and above)

Recommended settings

(can see workflow via uploaded images)

initial sampler

• res_multistep + normal/simple

• steps 30-100, cfg 3.5

• fast, reliable but slightly boring/simplistic results

• dpm_adaptive + exponential

• steps doesn't matter since adaptive, cfg 4.5

• slower but very reliable without much errors

upscale/hiresfix

• realesrgan_x2 (only because it's fast - don't think it makes a big difference which ever used)

• heun + normal/ddim_uniform(second choice)

• steps 15-20, cfg 3, denoise 0.35-0.55

• using heun+normal as the upscale/hiresfix is what brings out the noise/texture

facedetailer

• heun + normal

• steps 10, cfg 2.5, denoise 0.25-0.4

Remarks

• res_multistep is fast and converges at high steps, fixes the structural problems 90% of the time.

• usual euler/2m/3m + sgm uniform/normal/karass should work though

• i prefer heun as it produces natural noise/texture at lower step values and loses it at higher (recommended schedulers: normal/ddim_uniform/ays+)

• for prompt adherence: higher resolution can have lower cfg, lower resolution need higher cfg

• personal frequently used resolution before upscale: 896x1120 or 1152×1440 (more pixels = more space for details like hands)

• post-processing nodes with sharpening, bloom, grain, tone curves for more texture in final image. (see workflow in images)