Anima Preview-2

A detailed workflow for generating high-fidelity anime art, dubbed 'Anima Preview-2,' has gone viral within the Stable Diffusion community. Shared by user Haoming02 on Reddit, the guide specifies using the Forge Neo frontend interface with a precise configuration: the T2I Er_Sde sampler, SGM Uniform scheduler, 30 steps, and a CFG scale of 4. The core of the technique involves a two-stage process, starting with a base image generation before sending it to img2img for a 2x upscale using the Multidiffusion method with Mixture of Diffusers. Key technical parameters include a tile overlap of 128, matching the tile dimensions to the original resolution, and a low denoising strength of 0.12 to refine details without altering the composition. The final touch applies the 4xAnimeSharp upscaler to achieve a crisp, illustrated finish.

The workflow is notably defined by its extensive negative prompt, a critical tool for guiding the AI away from undesirable artifacts. This prompt meticulously bans categories like low quality, film grain, scan/jpeg artifacts, and the tell-tale signs of AI generation. It also explicitly filters out bad anatomy (missing fingers, extra limbs), text elements (watermarks, logos, speech bubbles), and specific stylistic traits deemed undesirable, such as 'muscular female' or 'crazy eyes.' The creator noted that Reddit's image compression significantly degraded the posted examples, suggesting the local outputs are of much higher quality than what viewers see online. This post represents the community's move towards highly specialized, reproducible pipelines for achieving specific artistic results, moving beyond basic prompting into advanced parameter tuning and post-processing.