Editing and Upscaling AI Anime Images

Updated 2026-10-20

The first generation is a draft. The difference between an average AI image and a good one is usually what happens afterwards: fixing the obvious errors, refining details, upscaling properly and exporting carefully. This guide walks through that post-generation workflow.

Step 1: Choose the right image to work on

Do not try to rescue a bad composition. Pick the image with the best overall pose, framing and lighting, even if the hands are wrong or one eye is off. Small errors are easy to fix; a bad composition is not.

Step 2: Inpainting

Inpainting regenerates only a masked area of the image while keeping the rest. It is the main tool for fixing AI mistakes.

How to inpaint well

1. Mask generously. Include a little of the surrounding area so the new part blends in.
2. Write a focused prompt for the masked area: “detailed hand holding a sword hilt, five fingers” rather than your whole scene prompt. Some tools work better with the full prompt plus the detail; try both.
3. Set denoising strength around 0.4–0.6 for fixes that should stay close to the original, and 0.7 or higher to replace something entirely.
4. Use “inpaint at full resolution” / “only masked” if available. The tool then works on the masked area at a higher resolution, which gives much better detail on small things like eyes and fingers.
5. Generate several variations and pick the best.

Common fixes

• Hands: the classic. Mask the hand, describe the gesture, generate a batch. If it keeps failing, change the pose to something simpler (hand on hip, hand partly hidden in a sleeve).
• Eyes: mismatched pupils, odd highlights, different colours. Inpaint both eyes together so they match.
• Faces in wide shots: small faces are often blurry. Inpaint the face at full resolution.
• Extra limbs or objects: mask and inpaint with a background description to remove them.
• Garbled text: mask it and replace with plain background, then add real lettering in an editor if needed.
• Age drift: if a face came out looking younger than intended, inpaint it with an explicit adult description, or discard the image. See safety and age rules.

Step 3: img2img for bigger changes

img2img uses your whole image as a starting point and regenerates it under a new or revised prompt.

• Low denoising (0.2–0.4): subtle changes, such as a lighting shift, colour adjustment or a style refresh.
• Medium (0.4–0.6): noticeable changes while keeping composition.
• High (0.6+): mostly new image, loosely following the original layout.

It is useful for turning a sketch or rough collage into a finished piece, changing time of day (see colour and lighting) or harmonising an image after several inpainting fixes.

Only use your own generations or your own drawings as input. Never use photos of real people.

Step 4: Upscaling

Generators work best at their native resolution, typically around one megapixel. For a large, sharp final image you need to upscale.

Types of upscaling

• Simple resize (bicubic, Lanczos): fast, but gets blurry. Fine for small enlargements.
• AI upscalers: models such as ESRGAN-family upscalers, including variants trained specifically on anime art. They sharpen lines and add plausible detail. Anime-specific upscalers keep flat colours flat and line art crisp.
• Upscale plus refine (“hires fix”, tiled upscaling): upscale, then run a light img2img pass on the larger image (denoising around 0.2–0.35) so the model adds real detail. Tiled methods process the image in pieces, so you can go large without running out of memory.

Tips

• Upscale 2x at a time, not 4x in one go.
• Keep refine denoising low; too high introduces new faces or patterns in the background.
• Check the result at 100% zoom for artefacts, especially in hair and fabric.

Step 5: Manual editing

A regular image editor (GIMP, Krita, Photoshop, Affinity Photo) is still useful:

• Crop for stronger composition. The rule of thirds and leaving space in the direction the character is looking both help.
• Levels and curves for contrast.
• Colour balance to unify the palette.
• Clone or heal small blemishes.
• Lettering for any text, logos or signatures.

Step 6: Export

Formats

• PNG: lossless, best for anime art with flat colour and sharp lines. Larger files.
• WebP: much smaller, good quality, widely supported online. Use quality 85–95 for illustrations.
• JPEG: smallest and universal, but compression artefacts are very visible around anime line art. Use quality 90+ if you must.
• AVIF: excellent compression, but not supported everywhere yet.

Step 7: Strip metadata and private information

This step is easy to forget and matters for privacy.

• Generation metadata. Many local tools embed your full prompt, negative prompt, seed, model names and settings in the PNG file. ComfyUI can embed the entire workflow. Anyone who downloads the file can read it.
• File names. Default names may include prompt text or dates.
• EXIF data. Images that passed through phones or editors can carry device info or location.
• Visible details. Screenshots can show usernames, browser tabs or notifications.

How to remove it: re-export through an editor with “save metadata” turned off, use a metadata-stripping tool, or convert to WebP or JPEG with metadata excluded. Check by opening the file in a metadata viewer before sharing.

If you want to keep your prompts for your own records, save them separately in a private text file. For more on privacy, see privacy and NSFW AI.

A compact workflow

1. Generate a batch, pick the best composition.
2. Inpaint hands, eyes and faces.
3. Optional img2img for lighting or style.
4. Upscale 2x with an anime upscaler, refine lightly.
5. Crop and colour-correct.
6. Export PNG or WebP without metadata.