To update the model, I have spent quite some buzz to generate LORAs.

Hence the early access, the buzz will be used to update newer versions.

If you enjoy using the model, I appreciate if you can do any of the following:

• Upload an image and tag the checkpoint. (I get buzz from each unique user that posts an image to model so even one upload helps alot. I also love to see uploads made with the model)

• Send a tip (even 10 buzz is appreciated)

• React to the latest images I have uploaded.

Thank you for the support.

Version notes

v2

• Increased colour saturation, brightness.

• Increased head size/reduced broad shoulders.

• Less dreamy, lesser artifact.

• Prompting should be more responsive.

v1

• First version that is on par/better than the illustrious mix

Version features

Rather than improvements, each version has their own strengths

• v1 - Dark, desaturated, dreamy style

• v2 - Bright, natural style

Model focus

• 3DCG/Semi realistic face; sharper/mature facial features with subtle skin texture.

Weakness

• Noise pattern on image. Recommend using WAN 2.1 fp32 VAE and film grain postprocessing node.

• Hands can be funky sometimes.

• Anime style hairstyle can be hard to prompt

Recommended settings

(can see workflow via uploaded images)

euler + simple

• steps 8-10, cfg 1

• standard, simple, stable, works most of the time

• downside - lacks texture, uniform texture, softer looking

euler + beta/(beta57-more tex)

• steps 8-12, cfg 1

• more texture than simple but can cause more artifacts

euler_a + beta57

• steps 8-15, cfg 1

• even more texture and details but also can get the image wrong

• can cause the image to look more illustrative/less realistic

deis_2m / deis_3m + beta57

• steps 3-4, cfg 1

• similar to euler+beta57 but can have more stable structure while having as much texture

• might have more realistic proportions compared to euler

• downside - can be soft?

• deis_3m is longer but might have better structure

Remarks

• Personally use 1024x1536 resolution most of the time. Lower resolutions makes the noise pattern more noticeable when upscaled.

• Upscale model recommendation: 2xLiveActionV1_SPAN. Keeps the image similar to the original without extra sharpening or lost of details while 2x resolution. Also fast.

• WAN 2.1 fp32 VAE and film grain postprocessing node for noise pattern reduction.