This workflow is designed to quickly build a production-ready character reference sheet while keeping identity, wardrobe, and pose/layout as separate controllable inputs.
The workflow follows a simple three-stage process:
1. Wardrobe Reference — Optional
Provide a single outfit/attire image and generate a detailed wardrobe reference sheet. The workflow extracts the clothing, materials, colors, accessories, footwear, layering, and smaller costume details.
You can skip this step and provide your own wardrobe reference instead.
2. Character Face — Optional
Generate a photorealistic character face/headshot to establish the character's identity.
Already have a character? Skip generation and load your own face reference.
3. Final Character Sheet
The face reference and wardrobe reference are combined with a predefined mannequin/pose layout to generate a consistent three-panel character sheet:
Left: Full-body front view
Center: Full-body back view
Right: Large face/head-and-shoulders reference
The mannequin image is used only for composition, pose, scale, and framing. It should not influence the final character's appearance.
Why this workflow?
The main goal is to avoid asking the image model to invent identity + costume + layout simultaneously.
Instead, each component is established separately before the final generation:
Face → Identity
Wardrobe → Costume
Mannequin → Pose & Layout
This makes the final reference sheet much easier to control and gives you reusable assets for subsequent image/video generation.
The wardrobe prompt also uses neutral, bright studio lighting so darker costumes retain their actual colors, fabric textures, straps, seams, leather details, and construction instead of collapsing into black.
Flexible Input
The first two generation stages are completely optional.
You can use:
Generated Face + Generated Wardrobe
or
Your Own Face + Generated Wardrobe
or
Generated Face + Your Own Wardrobe
or
Your Own Face + Your Own Wardrobe
Then feed them into the final character-sheet stage.
Recommended Use
Useful for creating reference assets for:
AI filmmaking • character consistency • Ref2VA workflows • image-to-video • storyboards • costume development • fan films • cinematic character design
Description
v1.0 → v2.0 — Character Sheets
Summary: rebuilt the identity pipeline around a real reference photo instead of a generated one, and expanded one character-sheet output into three parallel variants.
Added
New "Load Reference Face + Clean Up" stage. Replaces the old text-to-image face generator with a
LoadImage+ a newCreate Portrait From Loaded Imagesubgraph, so the character sheet is now built from an actual uploaded photo of a specific person instead of a randomly generated "beautiful woman" headshot.Two additional character-sheet variants:
Grey Mannequin Template variant (new
LoadImage+ a secondImage Editsubgraph instance).Action Poses Sheet variant — no template image at all; builds a 10-panel walk/run/jump/punch/kick/etc. sheet from prompt text alone, using only the face + wardrobe outputs.
(v1 had exactly one character-sheet output; v2 has three.)
MarkdownNote— same one workflow-notes note carried over from v1, kept as-is at the top.
Changed
Character-sheet panel count: 3 → 5. v1's prompt built a 3-panel sheet (front, back, portrait). v2's prompt (and its color-coded mannequin template) builds 5 panels: 3/4-left, back, 3/4-right, frontal close-up, 3/4 close-up.
Character-sheet prompt rewritten from scratch. v1 used a long section-header style prompt ("PANEL 1 — LEFT:", "IDENTITY CONSISTENCY:", etc.) with heavy negative-prompt lighting/color-crush guarding. v2 uses a single continuous paragraph per the model's own prompt-rewriter spec, with explicit
<image1>/<image2>/<image3>tagging, viewer-relative pose language (frame-edge-relative angles), and dedicated hand-visibility rules that v1 didn't have.Wardrobe Sheet Generation node:
Renamed internally from generic "Image Edit" to "Wardrobe Sheet Creation."
switch:False→True.Output resolution:
16:9→1:1 (Square)(a flat garment breakdown doesn't need a widescreen canvas).Negative prompt gained an explicit
mannequinexclusion at the front.
Face resolution selector removed. v1's face step had its own
ResolutionSelector(16:9, increment 8); v2's replacement portrait-cleanup subgraph fixes output at 1024×1024 internally — one less control surface.Groups: 3 → 5, retitled and recolored to separate the two prep stages from the three sheet variants.
Fixed
"Face Generation (Optoinal)" typo in the group title.
Internal
TextEncodeQwenImage21input-slot ordering inside the shared Wardrobe subgraph was tidied (cosmetic only — no behavior change).
Removed
The Text-to-Image face generator subgraph and its "beautiful woman" prompt — no longer part of the pipeline.
FAQ
Comments (4)
What about full body reference
In STEP 2B we have ....."Identity lock: all five figures are one and the same man".... maybe better "the same person" :>
And second "fix" in 1B it's loosing orginal making more beauty face. Better prompt for keep orginal face:
Create a photorealistic identity-preserving shoulders-up portrait of the exact same person from <image1>.
The primary goal is maximum facial identity fidelity, not beautification, retouching, or reinterpretation.
Frame the subject from the top of the shoulders upward, including:
- full head
- full hairstyle
- full neck
- small portion of the upper shoulders
Do not show the torso, arms, hands, or full outfit.
Use a plain neutral grey studio background.
Preserve the exact identity from <image1> as faithfully as possible.
Preserve:
- exact face shape
- apparent age
- facial proportions
- forehead shape
- eyebrow shape and density
- eye shape
- eye spacing
- eyelids
- iris color
- nose shape
- nostril shape
- lip shape
- lip volume
- mouth width
- jawline
- chin shape
- cheek structure
- ear shape
- skin tone
- natural skin texture
- pores
- freckles
- moles
- fine lines
- under-eye detail
- natural facial asymmetry
- distinctive facial marks
- hairstyle
- hairline
- hair color
- hair texture
- hair density
- natural flyaway hairs
Do not beautify the face.
Do not smooth or airbrush the skin.
Do not remove natural skin texture, pores, freckles, moles, fine lines, subtle wrinkles, asymmetry, or distinctive facial details.
Do not change apparent age.
Do not make the face younger, older, slimmer, fuller, more symmetrical, or conventionally more attractive.
Do not reshape the eyes, nose, lips, jaw, cheeks, chin, or forehead.
Do not change makeup unless necessary to preserve the appearance of <image1>.
Preserve the original expression from <image1> as closely as possible.
Preserve the original gaze direction and eye openness unless minor adjustment is required for a natural portrait.
Preserve the hairstyle and hair placement from <image1> as closely as possible.
Do not restyle, recolor, straighten, curl, shorten, lengthen, or add volume to the hair.
Lighting should be neutral, soft, and realistic, but must preserve facial texture and small surface details.
Avoid beauty lighting that removes pores, wrinkles, shadows, or facial structure.
The neck must remain anatomically natural and proportional to the head and shoulders.
Clothing is secondary. Show only a small, simple, unobtrusive neckline or upper garment area. Do not invent elaborate clothing, jewelry, accessories, or wardrobe details.
Final priority:
1. exact identity preservation
2. exact facial geometry
3. natural skin texture and distinctive details
4. exact hair and hairline
5. original apparent age and expression
6. natural neck and shoulder anatomy
7. simple neutral portrait presentation
Thank you. Will update
The output images are very grainy and low quality, even setting high megapixels and 16bit outputs. 😔 Using the same models as in the workflow and an HD image as reference. any clues? thanks! 🙏🏻