Human-led direction is a structured art direction process applied to AI image and panel generation. Instead of relying on one-shot prompts, you define a creative brief, assemble style/character references, constrain composition, and iterate with specific feedback. The result is reliable, on-brief visuals across shots, pages, and sequences.

Use human-led direction whenever consistency, fidelity, or narrative intent matters.

• Series and episodic anime stills where characters must remain on-model
• Comic pages with recurring angles, lighting, and mood
• Marketing visuals that must match a style guide
• Style exploration with controlled A/B tests to pick a direction
• Multi-shot workflows (storyboards → keyframes → finals)

1. Define intent

• Creative brief: subject, tone, audience, deliverables
• Style rules: color palette, line quality, texture, lighting, camera behavior
• Acceptance criteria: what makes an image "done"

2. Assemble references

• Character: front/side turnarounds, key expressions
• Style: 4–8 target images for line, shading, color, backgrounds
• World: props, locations, mood boards

3. Constrain composition

• Shot notes: camera, lens, framing, action, focal hierarchy
• Layout guides: thirds, leading lines, silhouette checks
• For panels: read order, gutters, speech balloon safe areas

4. Scaffold generation

• Start with roughs: storyboard frames, pose/segmentation control
• Progress to clean passes: lighting, materials, fine line weight
• Lock style after approval; only then scale volume

5. Iterate with specific feedback

• Replace vague notes ("make it cooler") with targeted direction ("cooler rim light, 6500K, right side, 30% intensity")
• Change one variable at a time to learn causality

6. Final QA and delivery

• On-model check, consistency vs. references, text readability, artifact cleanup
• Export specs: resolution, color space, file formats

Use prompts as instructions layered over your brief and shot notes. Keep variables explicit and modular.

Anime character close-up (panel-ready):

[subject]: earnest high-school swordswoman
[style]: crisp anime line, cel shading, soft rim light, subtle film grain
[camera]: 85mm portrait, medium close-up, eye-level
[lighting]: key left 45°, cool rim right, dusk ambiance
[focus]: eyes sharp, shallow DOF, background bokeh
[consistency]: on-model face per ref A, uniform per ref B

Comic splash action:

[subject]: cyberpunk courier leaping over neon rooftops
[style]: bold inking, halftone textures, limited palette (cyan/magenta/yellow + black)
[camera]: wide 24mm, low angle, dynamic perspective
[motion]: speed lines, debris trails, motion blur only on background
[layout]: title-safe top, caption-safe lower third

Style-lock variant prompt:

Base prompt + "match line weight = ref_style_03, color palette = ref_style_03, shading depth = 2-step cel"
Negative: blurry, extra limbs, wonky hands, text artifacts, watermark

• Image references: character sheets, style boards, environment packs
• Pose/structure control: pose control/segmentation maps to lock anatomy and layout
• Style adapters: style-transfer adapters or LoRAs for line/paint fidelity
• Inpaint/outpaint: fix hands, signage, text, and extend canvases cleanly
• Batch + seed strategy: lock seeds for explorations; change one variable at a time
• Versioning: save prompt, seed, control weights, and references per iteration

• On-model: face, hair silhouette, costume details match reference
• Composition: clear focal point, readable silhouette, rule-of-thirds or intentional break
• Anatomy and hands: number/shape, joint alignment, foreshortening
• Perspective: horizon and vanishing points consistent across panels
• Lighting: key/fill/rim logic; shadows consistent with time of day
• Text and UI: balloon legibility, SFX placement, kerning and stroke
• Artifacts: remove extra fingers, warped logos, stray edges
• Consistency: line weight, palette, texture density across sequence

• Vague briefs → Inconsistent style: write explicit style rules and show 4–8 reference targets
• Character drift across shots: use on-model references + pose/face control and locked seeds
• Over-busy frames: reduce micro-detail, increase contrast hierarchy, simplify backgrounds
• Unreadable panels: enforce balloon safe areas; reserve quiet space for text
• Lighting mismatch: standardize a lighting schema per scene (time of day, color temp)
• Endless iterations: set acceptance criteria and cap to N rounds per shot

• Consistency score: % of frames passing on-model and style checks
• Readability score: panel legibility at target device size
• Turnaround time: concept-to-approve per shot/page
• Edit rate: average changes per round; aim to reduce by locking variables
• Stakeholder alignment: brief sign-off before style-lock and before batch generation