Consistent Character Sheet: Art Deco poster of 1920s Paris via IP-Adapter Plus (style only) + Depth (MiDaS)

You are building a consistent-character workflow — the holy grail of AI image generation. The user wants the same character to appear across multiple poses, expressions, and scenes without drift.

## Brief
- **Character subject:** Art Deco poster of 1920s Paris
- **Identity preservation:** IP-Adapter Plus (style only)
- **Pose control:** Depth (MiDaS) (expected: OpenPose for bodies)
- **Base model:** AlbedoBase XL
- **Target outputs:** A 4-view or 8-pose character sheet

## The Consistency Stack
No single technique delivers character consistency. You must layer:

1. **Identity anchor** — a reference image of the character (user-provided or pre-generated)
2. **Face preservation** — IP-Adapter FaceID Plus v2 OR InstantID (stronger)
3. **Style preservation** — IP-Adapter Plus (style only) with second reference
4. **Pose control** — OpenPose ControlNet
5. **Seed locking** — same seed across all outputs in the sheet
6. **Prompt anchor** — identical character-description text in every prompt

## Step 1 — Create the Anchor Image
Generate one hero frame of Art Deco poster of 1920s Paris on AlbedoBase XL using your best single-image prompt. This becomes the reference.

Reference prompt (60-80 words):
```
[Detailed character description: age, face shape, hair, eyes, skin, wardrobe — use 3-part wardrobe spec]
```

Save this image. It will be the IP-Adapter reference for every subsequent frame.

## Step 2 — Build the Pose Sheet Inputs
You need Depth (MiDaS) pose maps for each target pose. Options:

- **Hand-pose** in ComfyUI: use OpenPose editor nodes
- **Extract from stock** photos of people in desired poses (run OpenPose preprocessor on them)
- **Generated pose sheets** from Pose3D tools (like CharacterTool or PoseHelper)

Minimum pose sheet: front, 3/4 left, profile left, back, 3/4 right, profile right, action pose 1, action pose 2.

## Step 3 — Workflow Per Frame

For each pose, build a ComfyUI graph:

```
Anchor Image → IP-Adapter FaceID Plus v2 (weight 0.8)
            → IP-Adapter Plus style (weight 0.5)
Pose Map    → Depth (MiDaS) (strength 0.85)
Positive    → "[anchor text] in [pose description]"
Negative    → standard AlbedoBase XL negatives
Seed        → LOCKED (same number every frame)
Sampler     → DPM++ 2M Karras, 30 steps, CFG 5
Upscale     → 4x-UltraSharp 2x + FaceDetailer
```

## Step 4 — The Prompt Template
Every frame uses this template:

```
[SHARED ANCHOR TEXT — character description — copy-paste identically every time]

+ [FRAME-SPECIFIC ADDITION: pose, expression, environment]
```

Example anchor text structure:
`[Name], a [age] year old [gender] [ethnicity cue], [face detail], [hair detail], [eye detail], wearing [wardrobe 3-part spec]. [Signature prop if any].`

Keep the anchor text UNDER 60 words and COPY IT VERBATIM into every frame prompt. Any rewording causes drift.

## Step 5 — Verification Pass
After generating all frames:
1. Pin them side-by-side in an image viewer.
2. Check: face shape, eye color, hair exact shade, wardrobe colors, accessories.
3. Flag any frame with identity drift > 15% and regenerate with stronger IP-Adapter Plus (style only) weight (try 0.9).

## IP-Adapter Plus (style only)-Specific Consistency Notes

- **IP-Adapter FaceID Plus v2:** Best for face-only identity lock. Pair with IP-Adapter Plus style for full look.
- **IP-Adapter FaceID Portrait:** Strongest face lock; can feel stiff. Use weight 0.6-0.7.
- **IP-Adapter Composition:** Transfers composition, NOT identity — do not use for character work.
- **InstantID (separate, not IP-Adapter):** Often beats IP-Adapter FaceID for realistic humans. Use together for max lock.

## AlbedoBase XL Consistency Tips
- **SDXL family (RealVisXL, Juggernaut XL):** Strong with IP-Adapter FaceID Plus v2.
- **Flux.1 [dev]:** Best natural consistency. Needs Flux-compatible IP-Adapter (PuLID for Flux).
- **SD 3.5 Large:** Emerging consistency toolkit; use reference-only ControlNet.
- **Anime checkpoints (Pony, Animagine XL):** Use character LoRA if available instead of IP-Adapter.

## Deliverables to Produce
Return four blocks:

### 1. Character Anchor Specification
The frozen character description (anchor text) + the reference image prompt.

### 2. Pose Sheet Plan
A table of 8 frames with: pose name, pose description, environment, expression.

### 3. ComfyUI Workflow
Node-by-node graph (similar to the recipe format). Must include IP-Adapter, Depth (MiDaS), FaceDetailer.

### 4. QA Checklist
Specific checkpoints for visual consistency across the sheet.

## Pitfalls
- Changing ANY word in the anchor text between frames = drift. Lock it.
- Upscaling with too much denoise kills consistency. Keep upscale denoise under 0.25.
- Faces at distance (small in frame) lose identity. Use FaceDetailer to restore.
- Multiple characters in one frame break most consistency workflows — keep 1 char/frame.
- Seed randomness between frames is OK if IP-Adapter + ControlNet weights are high; fixed seed helps if weights are low.

Produce the complete 4-block deliverable.