Consistency means a character’s identity remains recognizable across images, panels, or frames: same face structure, hair, colors, outfit details, and overall proportions—while allowing controlled variation in pose, camera, expression, and lighting. For AI anime/comics, this is achieved by combining identity anchors (prompts, references, or trained tokens) with deterministic generation (seeds, samplers) and structural control (pose/lineart/depth).

• Identity anchor: clear trait list (age, face shape, hair style/color, eye color, key outfit items, props, height/build).
• Style anchor: stable art style descriptor (e.g., “clean anime lineart, cel shading”).
• Reference anchor: one or more character reference images to guide appearance.
• Structure anchor: pose/lineart/depth control to hold proportions and silhouette.
• Determinism anchor: fixed seed, sampler, steps, CFG, and resolution.
• Scene anchors: camera (focal length, angle), palette, lighting, time of day.

Use at least three anchors at once (identity + reference + structure) for reliable results. Add more anchors for longer sequences.

1. Build a character sheet: front/3⁄4/side head, full body, 2–3 expressions, 1–2 outfits on neutral background.
2. Pick a consistent style prompt and finalize it on one strong portrait.
3. Save generation settings: model, seed, sampler, steps, CFG, resolution.
4. Use a reference-image module to inject identity (e.g., reference encoder or image adapter), strength ~0.5–0.8 for portraits, ~0.35–0.6 for full-body.
5. Add a structure control for pose/lineart/depth of your new shot. Keep strength moderate so identity isn’t overridden.
6. Reuse the same style prompt and seed. Change only pose/camera.
7. For each new scene, keep the seed fixed per angle; if you need small changes, adjust seed by tiny increments or use a low variation setting.
8. Export a panel set; inpaint minor mismatches (eyes, logos, seams) using the original reference intact.

Base template:

• Positive: "<char_id or reference>, female, early 20s, heart-shaped face, teal eyes, pink bob haircut with bangs, yellow bomber jacket, black pleated skirt, clean anime lineart, cel shading, 3/4 view, soft key light, studio background"
• Negative: "age shift, different hair color, extra fingers, mutated hands, off-model face, logo change"

Guidelines:

• Put immutable identity tokens early (hair style/color, eye color, signature clothing).
• Repeat palette and iconic items in every prompt (e.g., “yellow bomber jacket”).
• Keep style phrases constant across shots; vary camera/pose at the end of the prompt.
• If drift occurs, increase weight of identity traits and reduce CFG slightly.

• Define a continuity bible: character sheet, outfit rules, prop list, color palette, lighting per location.
• One seed per scene/angle: reuse for all panels within that setup.
• Storyboard first with rough poses; lock framing and pose with structure control; finalize with identity reference.
• Expressions: create a small bank (neutral, happy, angry, surprised) and reuse as refs.
• For action: drive poses from a consistent source (photos, 3D pose, or traced lineart) to maintain proportions.
• Keep resolution and aspect ratio identical across a page; resize at export, not per panel.

• Reference-only (no training): fastest; great for 1–30 images if you have a solid character sheet. Use identity + pose/lineart control.
• Textual inversion/embeddings: tiny, quick training; good for a personal tag (<char_id>) you can prompt later.
• LoRA: best for longer projects or complex identities/outfits; lets you control strength and combine with style LoRAs. Keep strength moderate to prevent overfitting.

Tip: Start reference-only. If you need 50+ consistent shots or complex wardrobe rules, train a lightweight identity (embedding or LoRA).

• Fix these per scene: model/checkpoint, seed, sampler, steps, CFG, resolution, style prompt, color palette, time of day.
• Adjust carefully: reference strength (too high → rigidity; too low → drift), structure strength (too high → identity loss), denoise strength in img2img (keep 0.3–0.55 for identity).
• Camera: specify focal length (e.g., 35mm close, 50–85mm portraits), angle (eye-level, low, high). Keep per-angle consistent.

• Hair/eye color drift: front-load hair/eye tokens; add palette terms; lower denoise; increase reference strength slightly.
• Outfit changes: repeat exact garment terms; use close-up outfit reference; inpaint logos/patches last.
• Face off-model: lower CFG 0.5–1.0; reduce steps slightly; increase identity weight; raise face detail inpaint at the end.
• Proportions shift: use consistent pose or lineart control; keep full-body references in the set.
• Style wobble: pin a single style phrase; avoid mixing many style tags per scene.
• Overfit look (too rigid): lower reference/LoRA strength; vary camera and lighting while keeping identity tokens constant.

• Character sheet exported (neutral background, multiple views, expressions, outfit variants)
• Saved settings JSON (model, seed, sampler, steps, CFG, resolution)
• Reference pack (top 3 identity refs, 1–2 outfit refs, expression refs)
• Style guide (line weight, shading, palette swatches, lighting setups)
• Scene kit (poses/lineart per panel, camera notes)
• Final QA pass: identity, colors, logos, continuity across panels