CIVITAI / Workflows

Anima Preview3 | Image-to-Image Anime Refinement Workflow

This workflow is designed for Anima Preview3 image-to-image generation, focusing on controlled anime-style transformation from an existing reference image. Its main purpose is to let creators upload a source image, guide it with a text prompt, and generate a cleaner Anima Preview3 result while still preserving the basic structure, pose, composition, and subject direction of the original picture. The workflow uses anima-preview3-base.safetensors as the main generation model, qwen_3_06b_base.safetensors as the text encoder, and qwen_image_vae.safetensors as the VAE. This creates a compact Anima Preview3 img2img pipeline where the input image is first resized, encoded into latent space, edited through the sampler, decoded back into an image, and then previewed or saved. Compared with a pure text-to-image workflow, this setup gives creators a stronger visual anchor because the model is not starting from an empty latent. It starts from the uploaded image and modifies it according to the prompt. The image preparation section is simple and practical. The source image is loaded through LoadImage, then passed into image_scale_pixel_v2, with the total pixel target set around 1 megapixel and alignment set to 64. This helps normalize the image into a model-friendly size before VAE encoding. The workflow is therefore useful for taking an existing AI draft, sketch, screenshot, character image, animal image, or rough composition and pushing it into a more polished Anima Preview3 style. The workflow then uses VAEEncode to convert the scaled image into latent space. This is the key difference from text-to-image. In text-to-image, the model creates everything from noise. In image-to-image, the original image becomes the base latent, so the final result can keep more of the original layout. This makes it especially useful for redraws, style refinement, concept cleanup, character reinterpretation, and controlled anime transformation. The sampling stage uses ClownsharKSampler_Beta with a 30-step setup, beta57 scheduler, linear/euler sampler route, CFG around 3, and denoise around 0.73. That denoise value is important: it is strong enough to visibly transform the image, but still keeps the source image as a meaningful reference. Lower denoise would preserve the original more strictly; higher denoise would push the result closer to a full redraw. The example prompt is simple: