AIGC anime is the practice of using generative models to create anime-style visuals and motion assets: character sheets, keyframes, backgrounds, manga panels, animatics, and short loops. It relies on text/image prompts, reference conditioning, and control signals to produce images that follow an art bible. Human art direction remains central for style definition, selection, corrections, and narrative cohesion.

Use interchangeable modules to go from concept to delivery:

1. Pre-production

• Style targets: collect references and define palette, line weight, rendering density, and lens rules.
• Character bible: nameable tokens for hero/side characters; poses, expressions, outfits.

2. Asset preparation

• Reference embeddings/LoRA for character identity and outfit variants.
• Background plates via text-to-image or kitbash, with depth/line consistency.

3. Shot creation

• Base generation: text-to-image for wide exploration; seed-lock promising candidates.
• Control passes: pose (skeleton), depth/normal, lineart/edge to keep composition and anatomy.
• Img2img refinement: low denoise to preserve identity; mask-inpaint for faces/hands.

4. Post

• Upscale with detail-preserving models; fix line chatter and moiré.
• Batch color management to enforce palette and contrast targets.
• For motion: generate adjacent keyframes, interpolate, and run de-flicker.

5. Review

• Art director pass for on-model checks, continuity, and narrative pacing.

Keep a tight review cycle so the model follows your vision:

• Define: style guardrails (palette, lens, line density, texture budget) and a no-go list.
• Guide: provide 3–5 best references per element (character, background, effects) per shot.
• Select: shortlist outputs by readability, character identity, and action clarity.
• Correct: annotate issues (hands, eyes, tangents, perspective), then re-run masked fixes.
• Log: save seeds, prompts, control strengths, and model versions for reproducibility.

Use structured prompts to reduce variance. Template:

Core template [subject], [character token or descriptor], [pose/action], [camera], [composition], [lighting], [palette], [style cues: line weight, screen tones], background: [setting], quality: [keywords]; negatives: [failures]

Example 1 (character close-up) Heroine A (short brown bob, sailor uniform), determined expression, 3/4 view, medium shot, rule-of-thirds, backlight rim, pastel city dusk, clean thin lines, soft cel shading; negative: extra fingers, blurry eyes, messy hair, watermark

Example 2 (dynamic action) Swordsman B LoRA, leaping slash, Dutch angle, motion lines, speed effects, dramatic key light, limited palette (crimson, charcoal, off-white), rooftop night; negative: broken anatomy, double weapon, text overlay

Tips

• Keep camera and composition explicit (e.g., 50mm, low angle, centered portrait).
• Put character identity early; place scene style and quality cues later.
• Use concise negatives for common failures (hands, eyes, extra limbs).

• Identity: train or use character LoRA/embeddings; maintain a stable token naming (e.g., heroine_A_v2).
• Seeds: fix seeds per shot; vary only when exploring.
• Control: combine pose + lineart/depth; start with moderate strengths and tune.
• Palette: LUT or palette quantization to enforce color language.
• Faces/hands: masked inpaint at low denoise; apply face restoration sparingly to keep linework.
• Wardrobe: create outfit variants as separate LoRA or prompt blocks.
• Backgrounds: reuse plates; lock horizon height and focal length across shots.

Visual QA

• Line discipline: avoid fuzzy halos or over-sharpening; check hair strands and eyelashes.
• Anatomy: hands, shoulders, neck length, ear alignment; fix via masked inpaint.
• Perspective and tangents: ensure clean overlaps; adjust with lineart control.
• Textures: avoid fabric noise; prefer cel-friendly gradients.

Post steps

• Upscale 1.5–2x with detail-preserving models; denoise small artifacts.
• De-flicker for sequences; match exposure and gamma across frames.
• Add subtle grain to hide banding; export with high-quality compression.

Checklist

• Art bible and character tokens finalized
• Control sources (pose/lineart/depth) prepared
• Seed and settings logged per shot
• Palette/LUT selected and applied
• QA pass done (anatomy, linework, continuity)

KPIs

• On-model rate (% of frames accepted on first pass)
• Revisions per shot (target <2)
• Avg gen time per accepted frame
• Consistency score (identity + palette match via quick visual rubric)

Brief A: School rooftop confession (4 panels)

• Look: soft cel, thin lines, dusk oranges and violets
• Assets: heroine_A_v2, background rooftop plate, pose refs
• Shots: medium two-shot, close-up eyes, wide establishing, hands detail

Brief B: Mecha hangar repair (6 keyframes + loop)

• Look: gritty cel with screen tones, cool palette with red accents
• Assets: engineer_C token, mech_07 plate, lineart control for machinery
• Shots: crane wide, over-shoulder tools, face close-up, motion loop on fans