Consistent Characters 2026

Consistent Characters 2026

When to Use

Activate this skill whenever the job-to-be-done is identity persistence: the same recognizable subject must survive across more than one image. Trigger on requests like:

• "Make a comic / storyboard / picture book where the hero looks the same on every page."
• "Generate a product mascot in 12 marketing poses."
• "I have one photo of a person — put them in 5 different scenes / outfits / lighting setups."
• "Create a character turnaround (front / 3-4 / profile / back)."
• "Keep this face but change the background / clothing / age / expression."
• "Build a contact sheet of expressions for a game NPC."
• Any multi-shot narrative (manga panels, ad campaign, brand spokes-character, avatar set, dataset for a downstream LoRA).

Situational conditions that raise priority:

• The deliverable is a set (≥2 images), not a one-off hero shot.
• The subject is a named, reusable identity (recurring character, IP, real person with rights/consent).
• Downstream work depends on consistency (LoRA dataset curation, animation reference, sprite sheets).
• The client has already rejected a draft for "it doesn't look like the same person."

When NOT to use

• Single, disposable image with no recurrence requirement — just prompt directly; identity-locking machinery adds cost and rigidity for no benefit.
• Style consistency without identity (e.g., "same painterly look across a series" but different subjects) — that is a style-reference problem (--sref, style LoRA, USO style mode), covered by the style-lock skill, not character identity.
• Generic "a man," "a woman," "a robot" with no need to be the same one twice — locking a non-identity wastes effort and reduces variety.
• Real-person likeness without consent / rights, deepfakes of private individuals, or impersonation for deception. Decline and redirect to original or consented identities. Likeness of public figures for deceptive or harassing purposes is out of scope.
• Pure text-to-3D or rigged-model pipelines where the engine itself guarantees identity — there the model is the source of truth, not the diffusion sampler.

Prerequisites

• Access to at least one generation platform: Midjourney (v6 or v7), Flux (FLUX.1 Kontext, USO, PuLID-FLUX, or FLUX.1 LoRA training), or a ComfyUI/Automatic1111 environment with IP-Adapter/InstantID/PuLID/ControlNet nodes installed.
• For LoRA training (Tier 5): 15–30 curated images of the subject with varied angles, expressions, and lighting; a training environment (e.g., Kohya, ai-toolkit, or equivalent FLUX LoRA trainer).
• For real-person likeness: confirmed rights/consent documentation before proceeding.

Procedure

The universal arc is Establish → Build → Lock → Verify. Pick the platform track, but the gates are the same. Always produce a single canonical reference ("the source of truth") early, and reference it — not derivatives — for every shot.

The Consistency Ladder (loosest → tightest)

Start at the lowest tier that could plausibly work, verify, and climb only when verification fails. For a short series, Tier 2–3 usually suffices. For an ongoing franchise, invest once in Tier 5.

Phase A — Establish the Canonical Reference (all platforms)

1. Write the identity brief. Lock the immutable traits in words even if you'll use image refs: name, age range, sex/gender presentation, ethnicity, face shape, eye color, distinctive marks (scar, freckles, mole), hair (color/length/style), build/height, and 1–2 signature wardrobe/props only if they're part of the identity. Keep this brief verbatim and reuse it token-for-token.
2. Generate candidate hero frames. Produce 4–8 front-facing, neutral-lighting, clear-face options. Front, eye-level, unobstructed face is mandatory for a good reference — it is the highest-information view for every downstream tool.
3. Select and upscale ONE as the canonical reference. Favor a clean, well-lit, expression-neutral frame. Upscale it; this becomes the only image you cite as --cref/--oref/IP-Adapter input/Kontext base.
4. Optionally build a character sheet (turnaround + expression strip) from the canonical frame to widen the reference pool before scaling out.

Phase B — Midjourney Track (Build & Lock)

v6.x — Character Reference:

• --cref <image_url> with --cw <0–100>.
  • --cw 100 (default) = copy face and hair and clothing.
  • --cw 0 = copy the face only — the correct setting when you want to change the outfit/scene but keep the person.
  • Intermediate values blend. Drop --cw toward 0 whenever wardrobe or context must change.

v7 — Omni-Reference:

• --oref <image_url> with --ow <0–1000> (default 100). Omni-Reference generalizes beyond faces to any subject (characters, creatures, objects, props) and is the successor to --cref for v7 pipelines. Raise --ow for stronger adherence; lower it when omni-reference fights your prompt or causes pose-locking. Very high --ow over-copies the reference pose/crop — keep it moderate for new poses.

Style Reference:

• --sref <image_url> (with --sw style weight) controls look, not identity. Combine --cref/--oref (who) with --sref (how) to hold both identity and art style across a series.

Steps:

1. From the canonical upscaled image, copy its URL.
2. For each new shot compose: <scene + action prompt> --cref <canonical_url> --cw <value> (v6) or --oref <canonical_url> --ow <value> (v7).
3. Set the weight by intent:
   • Same outfit & look across shots → --cw 100 / higher --ow.
   • New outfit / new scene, keep the person → --cw 0 (face only) / moderate --ow.
4. Hold art style with --sref <style_url> --sw <value> reused across the whole series.
5. Keep the same --cref/--oref URL for every panel. If a panel is excellent, you may add it to your reference pool, but the original canonical frame stays primary to prevent drift.
6. Upscale keepers; reject and re-roll off-model results rather than "fixing" them downstream.

Phase C — Flux Track (Build & Lock)

Choose the sub-track by task:

"Keep this image, change one thing" → Flux Kontext:

1. Load the canonical frame as the Kontext base.
2. Issue a minimal, explicit edit instruction: "Change the background to a rainy Tokyo street at night. Keep the character's face, hairstyle, and outfit exactly the same."
3. Generate, verify identity, then branch new edits from the canonical frame again (star pattern), not serially from each edited output (chain pattern), to avoid cumulative drift.

"Same subject, new scenes/styles from scratch" → USO subject mode:

1. Provide the canonical frame as subject reference.
2. Prompt the new scene/pose; add a style reference only if a look change is also wanted.
3. Tune subject strength so the face holds without copying the original pose.

"Photoreal face from a few photos" → PuLID-FLUX (Tier 3):

1. Supply 1–4 clean face shots.
2. Set ID weight.
3. Prompt scene/pose freely.

"Production, many shots over time" → Character LoRA (Tier 5):

1. Curate 15–30 images: varied angles, expressions, lighting; consistent identity; clean backgrounds; minimal occlusion. Quality and variety beat quantity.
2. Caption with a unique trigger token (e.g., ch_arden).
3. Train the FLUX LoRA; validate on held-out prompts.
4. Generate all future shots with the trigger token + scene prompt. Combine with ControlNet for pose control as needed.

Phase D — IP-Adapter / Open-Source Track (Build & Lock)

1. Identity layer: load IP-Adapter FaceID / InstantID / PuLID with the canonical face; set adapter weight ~0.5–0.8.
2. Structure layer: add ControlNet for each shot — openpose for body pose, depth for scene volume, face landmarks for a specific expression. Drive these from pose references or a posed 3D dummy.
3. Look layer: apply a style LoRA or style reference if the series needs a unified art direction.
4. Generate; if identity is weak, raise the adapter weight or add a second face reference; if pose is wrong, strengthen/swap the ControlNet, not the identity weight.
5. Batch the shot list, then run the Verification Gate before delivery.

Mental model: Identity (face embedding/LoRA) + Structure (ControlNet) + Look (style ref/LoRA) are three independent axes. Lock the ones the brief requires; leave the rest free for variety.

Phase E — Lock the Recipe

Record the exact, reproducible recipe so the next session reproduces the character: model + version, canonical reference URL/file, all weights (--cw/--ow/--sw, adapter weights, LoRA name + trigger + strength), seeds where relevant, and ControlNet configs. This recipe is the deliverable's reproducibility guarantee.

Pitfalls

Cross-cutting heuristics:

• When identity and something else are both wrong, fix identity last — get pose/composition/scene right with structure tools, then dial identity up.
• Prefer fixing inputs over fixing outputs: a better canonical frame or tighter crop beats a downstream patch.
• If three re-rolls fail at the current tier, climb the ladder rather than fighting the knobs.

Verification

Do not deliver a character set until every applicable item passes. Treat this as a hard gate, not a suggestion.

Visual identity check (run on the full set side-by-side):

• Contact sheet built. All shots tiled into a single grid at uniform size for direct comparison.
• Face match across all shots. Eye color, eye shape/spacing, nose, jaw/chin, brow, lip shape are consistent. No shot reads as "a different person."
• Distinctive marks persist. Every signature trait from the identity brief (scar, freckles, mole, heterochromia, etc.) appears, in the correct place, on every relevant shot.
• Hair consistent in color, length, and style (allowing only intentional, briefed changes).
• Skin tone / ethnicity stable across all lighting conditions — no race or tone drift.
• Age presentation stable — no slips between child/adult/elderly.
• Build & proportions stable — head-to-body ratio and height read consistently.

Multi-angle / multi-expression validation:

• Angle coverage tested. Identity holds at front, 3-4, and profile (and back if required). Off-frontal views verified explicitly, not assumed.
• Expression range tested. At least neutral + 2 emotions (e.g., smile, anger/surprise) generated; identity survives expression changes.
• Pose variety tested. At least one non-portrait, full-body or action pose confirms identity holds beyond the headshot.

Wardrobe / scene isolation (when variation was requested):

• Outfit changed without identity change, and vice versa — wardrobe and identity are independently controllable.
• No background bleed from the reference into shots that should have new scenes.

Production / reproducibility:

• Canonical reference recorded (the single source-of-truth image/URL).
• Recipe recorded — model + version, all weights (--cw/--ow/--sw, adapter weights, LoRA name/trigger/strength), seeds, ControlNet configs — sufficient to reproduce the character in a future session.
• Rights / consent confirmed for any real-person likeness; no impersonation or deceptive use.
• Off-model rejects discarded, not shipped — every delivered frame passes the face-match check above.

If any box fails: identify the failure mode in Pitfalls, apply the fix or climb the consistency ladder, regenerate the affected shots, and re-run the gate. Ship only a fully passing set.