CinPony

CHECKPOINT
Reprint


Updated:

TL;DR; CinEro Reworked. Wide CFG range. Read params recommendations notes.

This model takes the best from CinEro (XL | PONY), XenEro, XenoBooru.

Goal: combine the Pony / Booru prompts adherence with cinematic qualities of CinEro series and creativity of XenEro series.

VERSIONS

V2_Razor

  • It renders awesome glass, metals and sharp-edged stones. It renders the picturesque landscapes. It produces a quite good crispy skin textures. Slightly biased toward feminine chars, but can easily render males if you ask it specifically.

  • Wide CFG range: 3 (realism) ... 11 (digital art) - use CFG 5 for better realism (slightly lower adherence); use CFG 9-11 for digital art style (higher adherence but lower creativity).

  • Better PONY prompt adherence

  • Higher sharpness - sharpness of small details increased

  • Better anatomy - feminine poses, hands, fingers, eyes; use HiRes Fix or ADetailer for best results

  • Less cinematic because of much greater sharpness.

V1_CintaClaws

Initial experiment of combining the UNet blocks from CinEro/XenEro and CLIP from XenoBooru with multi stage adjusting, tuning and flaw fixing via Elemental Merges.

In total I spent few weeks and about 100 iterations of merging and testing to get a satisfying result. Hope all these efforts worth this small EA donation.

Highlights

  • Good BooRu tags adherence

  • More stable anatomy (in comparison with CinEro)

  • More diverse faces

  • Stable at high CFG (tested on CFG 9)

  • Better eyes, fingers (higher chance to get good anatomy details without ADetailer)

  • Can draw furry (tested with anthropomorphic cat-woman prompts)

  • Moderate creativity

  • Cinematic lighting

  • More dynamic poses, better feminine anatomy

  • High resolution images without body duplication (tested at 1024 x 1440, 1200 x 1200, 960 x 1216)

Params Recommendations

CFG 12 == semi-Realistic image with higher contrast but slightly unstable (look for examples in main image post).

CFG 5 == low contrast, dimmed colors, smaller details, more realism.

Steps for DPM++ 2S a: 30 for balanced speed and quality; 45-60 for good quality and realism with low speed.

Steps for Euler, DPM++ SDE **: 30 - smooth, lower quality but fast; 45 - good balance of quality and speed; 70 for better quality and more detailed textures.

Low CFGs give more realistic and less contrast picture.

Euler A, DPM++ 2M SDE Heun with SGM Uniform == smoother results.

DPM++ 2S a, Heun with Karras == more realistic and detailed results BUT slow.

DPM++ 2S a with Karras at CFG 5 == good realism with smooth lighting and shadows, moderate colors, consistent background

Use DDIM-based samplers and DDIM scheduler if you need better chance of good anatomy.

DDIM CFG++ with DDIM at CFG 12 == good contrast, semi-realistic, slightly unstable but gives interesting combination of smoother textures and more contrast medium-to-small objects like fluffy clouds and foliage.

DDIM CFG++ with DDIM at CFG 9 == more stable and more realistic but still have artistic look.

Lowering CFG with "DDIM CFG++" produces smaller details but leads to instabilities.

Version Detail

Pony
V2_Razor It renders awesome glass, metals and sharp-edged stones. It renders the picturesque landscapes. It produces a quite good crispy skin textures. Slightly biased toward feminine chars, but can easily render males if you ask it specifically. Wide CFG range: 3 (realism) ... 11 (digital art) - use CFG 5 for better realism (slightly lower adherence); use CFG 9-11 for digital art style (higher adherence but lower creativity). Better PONY prompt adherence Higher sharpness - sharpness of small details increased Better anatomy - feminine poses, hands, fingers, eyes; use HiRes Fix or ADetailer for best results Less cinematic because of much greater sharpness.

Project Permissions

Model reprinted from : https://civitai.com/models/1096167?modelVersionId=1289716

Reprinted models are for communication and learning purposes only, not for commercial use. Original authors can contact us to transfer the models through our Discord channel --- #claim-models.

Related Posts

Describe the image you want to generate, then press Enter to send.