Anime diffusion refers to diffusion models and workflows specialized for anime/manga aesthetics: crisp line art, cel shading, flat or softly graded palettes, and stylized anatomy. Popular bases include SD 1.5 anime checkpoints and SDXL anime models. Compared to photoreal models, anime checkpoints respond better to tag-like prompts, benefit from different CFG and sampler choices, and often use features like Clip Skip (for SD 1.5 variants) to better align with anime tagging conventions.

• Models (pick one): SDXL anime (e.g., Animagine XL), SD 1.5 anime (e.g., Anything v4.5, MeinaMix, Counterfeit, AOM-like mixes).
• Sampler: DPM++ 2M Karras or DPM++ SDE Karras.
• Steps: 20–28 (SD 1.5), 28–40 (SDXL).
• CFG scale: 5–7 (SD 1.5); 4.5–6 (SDXL).
• Resolution: 768×1024 or 1024×1024 (SDXL); 512×768 or 640×896 (SD 1.5). Match aspect to your subject.
• Clip Skip: 2 for many SD 1.5 anime models; not applicable for SDXL.
• Hires pass/upscale: 1.5–2× with Latent or ESRGAN-type upscaler; denoise 0.2–0.35.
• VAE: Use the model’s recommended VAE to avoid color shifts and muddy lines.
• Seeds: Fix a seed for iteration and reproducibility.

Structure prompts by priority: quality tags → subject → attributes → scene → style/medium → camera.

Example (positive): "masterpiece, best quality, anime, clean lineart, cel shading, 1girl, solo, full body, dynamic pose, detailed eyes, vibrant colors, scenic background, soft lighting, depth of field"

Negative prompt (general): "lowres, blurry, bad anatomy, extra fingers, deformed hands, warped face, artifacts, watermark, text"

Tips:

• Place quality tags first, then subject and attributes.
• For character- or style-specific results, add a LoRA with a weight like <lora:style-name:0.8–1.2>.
• Use concise, relevant tags; overlong prompts can cause conflict.
• If colors feel too intense, lower CFG or include terms like "flat colors, minimal shading".
• To emphasize line clarity, include "clean lineart, sharp lines" and keep denoise low on upscales.

• Base checkpoints: SD 1.5 anime mixes (e.g., Anything, MeinaMix, Counterfeit, AOM-style) and SDXL anime (e.g., Animagine XL line).
• LoRA: Add character, outfit, or style specificity. Start at 0.6–1.0 weight; raise carefully to avoid overbake.
• Textual inversion/embeddings: Compact style or tag bundles; useful but model-specific.
• ControlNet/IP-Adapter: Guide pose (OpenPose), line art/edge adherence (Lineart, Canny, SoftEdge), and layout (Tile). Great for consistency and composition.
• Regional prompting: Assign different prompts to character vs background for clean separation.
• VAEs: Use the recommended VAE per model to preserve palette and contrast.

• Text-to-image (baseline):

  1. Choose anime model and VAE. 2) DPM++ 2M Karras, 24–32 steps, CFG 5–6.5. 3) 768×1152 for full-body. 4) Hires pass 1.5×, denoise 0.25.

• Image-to-image style transfer:

  1. Provide a rough sketch or 3D blockout. 2) Denoise 0.35–0.55 for strong restyle; 0.15–0.3 for gentle clean-up. 3) Add "clean lineart, cel shading" to anchor the look.

• Pose control with ControlNet (OpenPose):

  1. Generate or capture a pose reference. 2) Strength 0.6–0.9 to lock pose while allowing details. 3) Keep prompt focused on outfit and scene.

• Lineart adherence (ControlNet Lineart/Canny/SoftEdge):

  1. Feed inked sketch or edge map. 2) Lower denoise on hires pass to retain outlines. 3) Add "sharp lines" in prompt.

• Manga panel colorization:

  1. Use img2img on inked panels. 2) Prompt palette and lighting cues. 3) Keep CFG 4.5–6 to avoid over-saturation; finalize with ESRGAN x2.

• Character + background compositing:

  1. Generate character on plain bg. 2) Generate background separately. 3) Merge and run a light inpaint to harmonize lighting.

• Mushy or fuzzy lines: Use the model’s recommended VAE; try DPM++ SDE Karras; increase steps slightly; lower denoise on upscales; add "clean lineart".
• Hands and anatomy issues: Lower CFG; try a hand/anatomy LoRA at 0.6–0.9; inpaint problem areas; guide pose with OpenPose ControlNet.
• Over-saturated or neon colors: Reduce CFG by 0.5–1.5; include "flat colors"; switch VAE; consider a different anime checkpoint.
• Detail overbake/glitter: Reduce LoRA weights; shorten prompt; lower steps 2–4.
• Orange tint or washed palette: Swap VAE; confirm color management; avoid double VAEs.
• Character drift across shots: Fix seed; use ControlNet pose; keep LoRA weight stable; avoid adding conflicting style tags mid-series.

• Upscalers: ESRGAN/RealESRGAN anime variants, 4x-UltraSharp, and Latent work well for clean edges.
• Sharpening: Apply mild sharpening after upscale; avoid halos.
• Compression: PNG or WebP (lossless for masters, lossy for web). Avoid heavy JPEG compression which harms lines.
• Color: Keep sRGB; ensure viewer/app color consistency.
• Metadata: Save seeds and settings for reproducibility.

Review model licenses and permitted uses before commercial work. Credit and comply with license terms when required. Be mindful of rights around branded or trademarked characters and follow applicable laws and platform rules.

• Anime vs photoreal diffusion differences? Anime models respond to tag-like prompts, benefit from Clip Skip (SD 1.5), and emphasize line art and cel shading.
• Do I need Clip Skip? Often for SD 1.5 anime models (Clip Skip 2). SDXL models typically don’t use it.
• SDXL or SD 1.5 for anime? SDXL anime models can offer higher native resolution and stability; SD 1.5 anime mixes are abundant and fast. Try both.
• Best resolution? Start 768×1152 (full-body) or 1024×1024 (portrait), then upscale 1.5–2×.
• Is ControlNet required? No, but it’s the most reliable way to lock pose, edges, or layout when consistency matters.