For best outputs and usage:
Sampler:
Realistic: DPM++ 2M SDE Karras, DPM++ 2M Karras.
Rest: Euler a / DPM++ 2M SDE Exponential.
Steps: 30, cfg 8, CLIP skip 1.
VAE:
Realistic: vae-ft-mse-840000-ema-pruned_fp16.safetensors
Rest: kl-f8-anime2_fp16.safetensors
Pos prompt: this is a versatile type model, which means it can go from perfect photo-realism to nice anime. In order to achieve this, for example:
Realism: wrap your text between "(Photorealistic)" and "realistic, real photo, realistic skin texture, skin imperfections, skin pores, sharp details, bright light, 8k, HDR, f.1/2, 70mm lens, Kodak, Analog style, <lora:epiCRealLife:1>", being the LoRa "epiCRealLife" included in my photo-realistic runs (optional, but highly recommended).
Anime: "anime, (2d flat colors), plain colors, japanese animation style" at the beginning of the prompt.
Rest: just ask the model the art style you want. It knows of many different styles.
Overall note about the pos, the model cannot stand too much keyword weighting, doing it will actually decrease the quality.
Neg prompt:
Realism: "(anime), (cartoon), (3d), (cg), (render), (unrealistic), heavy shadows, fused limbs, malformed hands, out of focus, long body, monochrome, symbol, text, logo, door frame, window frame, mirror frame, epiCNegative, CyberRealistic_Negative-neg, realisticvision-negative-embedding", being the embeddings "epiCNegative", "CyberRealistic_Negative-neg" and "realisticvision-negative-embedding" included in my photorealistic example runs.
Anime: "girl, shadows, (Realistic), 3d, render, cg, (volumetric), photorealistic, fused limbs, malformed hands, out of focus, long body, monochrome, symbol, text, logo, door frame, window frame, mirror frame, EasyNegative", being the embedding "EasyNegative" included in my anime example runs.
Rest: "(realistic), ugly, worst quality, bad quality, normal resolution".
For best outputs and usage:
Sampler:
Realistic: DPM++ 2M SDE Karras, DPM++ 2M Karras.
Rest: Euler a / DPM++ 2M SDE Exponential.
Steps: 30, cfg 8, CLIP skip 1.
VAE:
Realistic: vae-ft-mse-840000-ema-pruned_fp16.safetensors
Rest: kl-f8-anime2_fp16.safetensors
Pos prompt: this is a versatile type model, which means it can go from perfect photo-realism to nice anime. In order to achieve this, for example:
Realism: wrap your text between "(Photorealistic)" and "realistic, real photo, realistic skin texture, skin imperfections, skin pores, sharp details, bright light, 8k, HDR, f.1/2, 70mm lens, Kodak, Analog style, <lora:epiCRealLife:1>", being the LoRa "epiCRealLife" included in my photo-realistic runs (optional, but highly recommended).
Anime: "anime, (2d flat colors), plain colors, japanese animation style" at the beginning of the prompt.
Rest: just ask the model the art style you want. It knows of many different styles.
Overall note about the pos, the model cannot stand too much keyword weighting, doing it will actually decrease the quality.
Neg prompt:
Realism: "(anime), (cartoon), (3d), (cg), (render), (unrealistic), heavy shadows, fused limbs, malformed hands, out of focus, long body, monochrome, symbol, text, logo, door frame, window frame, mirror frame, epiCNegative, CyberRealistic_Negative-neg, realisticvision-negative-embedding", being the embeddings "epiCNegative", "CyberRealistic_Negative-neg" and "realisticvision-negative-embedding" included in my photorealistic example runs.
Anime: "girl, shadows, (Realistic), 3d, render, cg, (volumetric), photorealistic, fused limbs, malformed hands, out of focus, long body, monochrome, symbol, text, logo, door frame, window frame, mirror frame, EasyNegative", being the embedding "EasyNegative" included in my anime example runs.
Rest: "(realistic), ugly, worst quality, bad quality, normal resolution".
