Some fixes was made. The recipe received a modification in the form of SDPO, DMPO, and was also cleared of debris that I missed when preparing the first model. This model much stable, and have superior quality.

Model have medium cfg sensitivity. CFG 4 or 5 is recommended. Recommended vae xlVAEC_f22 or f231. MaHiRo is also recommended if cfg scale on 4-4.5.

Recommend sampler - DPM++ 2m (or euler) for generating process. Scheduler - polyexponential, or beta. dpm++ 2m for inpaint (sgm uniform scheduler) and upscale. 28-35 steps.

Model can handle large tensor resolution (1536x1536), and i can recommend this on general use if you making composition like portrait, full body standing or something wide shot. Thanks to the new predictor, the model handle with this many times better than the EPS one.

Generic positive: very as2, masterpiece, absurdres, best quality, very awa, very aesthetic, year 2024, newest, ultra-detailed, amazing quality.

Generic negative: (patreon username, patreon logo, artist logo, twitter username, signature, username, logo, watermark, copyright name, copyright symbol, web address, twitter logo, gumroad username, text, artist name, gumroad logo, artist logo, logo, blurry, blurry background, blurry foreground, blurry edges, blurry vision, bokeh, depth of field:1.2), green theme, red theme, yellow theme, blue theme, worst quality, worst aesthetic, bad quality, low quality, (bad hands:1.2), censored, bar censor, mosaic censoring, artistic error, bad proportions, bad perspective, wrong hand, bad reflections, too many fingers, too many hands, too many legs, wrong feet, bad feet, extra digits, missing fingers,bad anatomy,anime screenshot, anime coloring, abstract, ai-generated, lossy-lossless, jpeg artifacts, scan artifacts,