NSFW-Character-Sheets-KREA2

Krea2 + Qwen Consistent Character Master Workflow

Workflow version: V11.4 Experiment B
Architecture base: V11.1 Full-Body Master Architecture
Platform: ComfyUI

This repository contains a multi-stage ComfyUI workflow for generating a consistent adult character across three coordinated outputs:

1. Bare Master — the canonical full-body identity source.
2. Portrait — a commercial-style head-and-shoulders portrait derived from the same identity.
3. Clothed Full Body — a dressed full-body version anchored to both the Bare Master and the final Portrait.

The workflow is designed around one core idea:

Generate the identity once, then derive every later representation from that same master character instead of independently regenerating the person.

This helps reduce identity drift between portrait and full-body outputs.

Content notice: The supplied workflow is configured for adult / NSFW-capable character generation in the Bare Master stage. It is intended only for clearly adult fictional or consenting subjects. Do not use it to create sexual imagery of minors, non-consensual real-person deepfakes, or other unlawful or abusive content.

---

Workflow Overview

TEXT PROMPT
    |
    v
[1] BARE MASTER
Krea2 Turbo FP8
+ skin realism LoRA
+ NSFW LoRA
+ optional / subject-specific anatomy LoRAs
    |
    | 1024x1824
    | 12-step base
    | 1.25x latent upscale
    | 8-step hires refinement
    v
BARE MASTER IMAGE
    |
    +--------------------------+
    |                          |
    v                          |
[2] PORTRAIT STRUCTURE         |
Qwen Rapid AIO SFW v23         |
uses Bare Master as reference  |
    |                          |
    v                          |
Krea2 Identity Edit v1.2       |
quality / identity pass        |
    |                          |
    v                          |
FINAL PORTRAIT ----------------+
    |
    v
[3] CLOTHED FULL STRUCTURE
Qwen Rapid AIO SFW v23
uses:
- Bare Master
- Final Portrait
    |
    v
Krea2 Identity Edit v1.2
quality / identity pass
    |
    v
FINAL CLOTHED FULL BODY

---

Why This Architecture

A common problem in AI character pipelines is that separately generated portrait, nude/bare reference, and clothed full-body images slowly become different people.

This workflow instead uses a master-first identity architecture:

• The Bare Master establishes the canonical body and identity.
• The Portrait is created from the Bare Master.
• The Clothed Full Body is created from both the Bare Master and final Portrait.
• Krea2 Identity Edit is used as a quality-preserving reconstruction stage for the SFW portrait and clothed branches.
• The identity-edit quality stages intentionally focus on reconstruction and photographic detail rather than redesigning the character.

The exported workflow contains only three final SaveImage nodes:

• Bare Master
• Portrait
• Clothed Full Body

---

Required Models

The workflow references the following filenames exactly.

Diffusion Model

darkBeast30BF16INT8_darkBeastKREA2FP8.safetensors

Loaded through UNETLoader as the Krea2 Turbo FP8 base.

Expected location:

ComfyUI/models/diffusion_models/

---

Krea2 Text Encoder

Qwen3-VL-4B-Instruct-Uncensored.safetensors

Loaded using:

CLIPLoader
type: krea2

Expected location:

ComfyUI/models/text_encoders/

---

VAE

sharp_vae_bf16.safetensors

Expected location:

ComfyUI/models/vae/

---

Qwen Structure Checkpoint

Qwen/Qwen-Rapid-AIO-SFW-v23.safetensors

Used for both:

• Portrait structure generation
• Clothed full-body structure generation

Expected location:

ComfyUI/models/checkpoints/Qwen/

---

Required LoRAs

Shared Realism LoRA

skin_human.safetensors

Strength:

0.80

Expected location:

ComfyUI/models/loras/

This LoRA feeds the shared Krea2 model used by the Bare Master and both later Krea2 quality branches.

---

Bare Master NSFW LoRA

Krea2/Krea2_HMNSFW_AIO.safetensors

Strength:

0.55

Expected location:

ComfyUI/models/loras/Krea2/

This LoRA is applied only to the Bare Master model chain.

---

Krea2 Identity Edit

krea2_identity_edit_v1_2.safetensors

Strength:

1.0

Expected location:

ComfyUI/models/loras/

Used by both the Portrait and Clothed Full quality passes.

---

Male-Only Anatomy LoRAs

The exported workflow also contains these two active LoRAs:

Krea 2 - Penis Lora v2.safetensors
strength: 0.65

detailed_penis_krea_2.safetensors
strength: 0.35

Expected location:

ComfyUI/models/loras/

Important

These nodes are labeled MALE ONLY, but in the exported workflow they are currently active in the model chain rather than automatically bypassed.

If your controller or API does not dynamically disable them, manually bypass or mute them for characters where they are not applicable.

---

Required Custom Nodes

ComfyUI-Krea2Edit

The workflow uses:

Krea2EditModelPatch
Krea2EditGroundedEncode

Install:

cd ComfyUI/custom_nodes
git clone https://github.com/lbouaraba/comfyui-krea2edit

Restart ComfyUI afterward.

The Krea2Edit node pack requires a ComfyUI build with native Krea2 support and the matching Krea2 Identity Edit LoRA.

Repository:

https://github.com/lbouaraba/comfyui-krea2edit

---

ComfyUI Core Qwen Image Edit Plus

The workflow also uses:

TextEncodeQwenImageEditPlus

This is a current ComfyUI core node used for Qwen Image Edit Plus conditioning.

If the node is missing, update ComfyUI before installing additional third-party node packs.

Documentation:

https://docs.comfy.org/built-in-nodes/TextEncodeQwenImageEditPlus

---

Recommended Model Directory Layout

ComfyUI/
└── models/
    ├── diffusion_models/
    │   └── darkBeast30BF16INT8_darkBeastKREA2FP8.safetensors
    │
    ├── text_encoders/
    │   └── Qwen3-VL-4B-Instruct-Uncensored.safetensors
    │
    ├── vae/
    │   └── sharp_vae_bf16.safetensors
    │
    ├── checkpoints/
    │   └── Qwen/
    │       └── Qwen-Rapid-AIO-SFW-v23.safetensors
    │
    └── loras/
        ├── skin_human.safetensors
        ├── krea2_identity_edit_v1_2.safetensors
        ├── Krea 2 - Penis Lora v2.safetensors
        ├── detailed_penis_krea_2.safetensors
        └── Krea2/
            └── Krea2_HMNSFW_AIO.safetensors

---

Stage 1 — Bare Master

The Bare Master is the canonical identity source for the entire workflow.

Base Generation

Resolution:

1024 x 1824

Sampler:

Sampler: Euler
Scheduler: Simple
Steps: 12
CFG: 1.0
Denoise: 1.0
Seed: 424242424242

Negative conditioning is produced with:

ConditioningZeroOut

This means the workflow is using zeroed negative conditioning rather than a conventional text negative prompt.

Hires Refinement

Latent upscale:

Method: nearest-exact
Scale: 1.25x

Approximate final latent size:

1280 x 2280

Second sampler:

Sampler: Euler
Scheduler: Simple
Steps: 8
CFG: 1.0
Denoise: 0.25
Seed: 424242424242

The resulting image becomes the reference source for the later branches.

---

Stage 2 — Portrait

Qwen Structure Pass

The Portrait structure pass uses:

Qwen-Rapid-AIO-SFW-v23.safetensors

Reference:

image1 = Bare Master

Target latent:

1024 x 1365

Sampler:

Sampler: Euler Ancestral
Scheduler: Beta
Steps: 4
CFG: 1.0
Denoise: 1.0
Seed: 65454653

The default instruction is intentionally simple:

Create a head-and-shoulders portrait of the same person.

This stage establishes composition and crop before the Krea2 reconstruction pass.

---

Portrait Krea2 Identity / Quality Pass

Model:

Krea2 base
+ skin_human @ 0.80
+ krea2_identity_edit_v1_2 @ 1.0

Identity patch:

ref_boost: 5
ref_boost_a: 1
fit_mode: fit

Grounding resolution:

1024

Target:

1224 x 1632

Base quality sampler:

Euler
Simple
12 steps
CFG 1
Denoise 1

Hires refinement:

1.25x latent upscale
Euler
Simple
8 steps
CFG 1
Denoise 0.25

Approximate final latent size:

1530 x 2040

The quality prompt explicitly instructs the model to preserve identity, face, crop, hair, expression, and other core character traits while improving photographic detail.

---

Stage 3 — Clothed Full Body

Qwen Structure Pass

The Clothed Full structure pass also uses:

Qwen-Rapid-AIO-SFW-v23.safetensors

References:

image1 = Bare Master
image2 = Final Portrait

Target latent:

1024 x 1824

Sampler:

Sampler: Euler Ancestral
Scheduler: Beta
Steps: 4
CFG: 1.0
Denoise: 1.0
Seed: 65454653

Default instruction:

Add clothing to the same adult character.

Using both the Bare Master and Portrait gives Qwen body information plus a higher-quality facial identity reference.

---

Clothed Full Krea2 Identity / Quality Pass

Model:

Krea2 base
+ skin_human @ 0.80
+ krea2_identity_edit_v1_2 @ 1.0

Identity patch:

ref_boost: 5
ref_boost_a: 1
fit_mode: fit

Grounding resolution:

1024

Base target:

1024 x 1824

Base sampler:

Euler
Simple
12 steps
CFG 1
Denoise 1

Hires refinement:

1.25x latent upscale
Euler
Simple
8 steps
CFG 1
Denoise 0.25

Approximate final latent size:

1280 x 2280

The reconstruction prompt instructs Krea2 to preserve:

• identity
• face
• hairstyle
• pose
• body proportions
• outfit
• framing
• background

while improving photographic detail and texture.

---

API / Automation Injection Points

This workflow is structured well for external API control because the primary text instructions are concentrated in three nodes.

Useful final output nodes:

This makes it relatively straightforward for an external controller to inject character-specific prompts and collect the three outputs independently.

---

Default Output Location

All three SaveImage nodes currently use:

WannaWatch/characters

Therefore ComfyUI will save them beneath its normal output directory using the same filename prefix.

Important

The node titles say:

Save _bare
Save portrait
Save _full

but the exported filename_prefix value is the same for all three nodes.

The node title does not automatically add _bare, _portrait, or _full to the actual filename.

If you want human-readable output names without relying on node IDs or an external controller, give each SaveImage node a distinct filename prefix.

Example:

WannaWatch/characters_bare
WannaWatch/characters_portrait
WannaWatch/characters_full

---

Important Configuration Notes

1. Male-Only LoRAs Are Active in the Export

Both nodes labeled MALE ONLY are currently in the active Bare Master chain.

They are not automatically conditional inside the workflow JSON.

If gender-specific switching is performed by an external API/controller, document that behavior separately. Otherwise manually bypass them when not appropriate.

---

2. Seeds Are Fixed

The workflow currently uses fixed seeds:

Bare / Krea2 quality:
424242424242

Qwen structure:
65454653

This is useful for reproducible testing.

For production character creation, you may want your controller to randomize the Bare Master seed while keeping downstream derivation deterministic for a given character.

---

3. Final Krea2 Hires Passes Exceed ~2 MP

The Krea2Edit project recommends generating at approximately 2 megapixels or less because higher resolutions may increase source bleed or subject duplication.

The workflow's base Krea2 quality latents are close to that range:

Portrait quality base:
1224 x 1632 ≈ 2.00 MP

Clothed quality base:
1024 x 1824 ≈ 1.87 MP

However, the 1.25x latent hires passes produce approximately:

Portrait:
1530 x 2040 ≈ 3.12 MP

Clothed full:
1280 x 2280 ≈ 2.92 MP

If you experience:

• duplicate facial features
• identity drift
• reference-image bleed
• duplicated subjects
• unexpected composition changes

test one of the following:

1. Disable the final 1.25x latent upscale.
2. Reduce the latent upscale factor.
3. Keep Krea2 generation near 2 MP and perform a separate pixel-space upscale after decoding.

---

4. target_latent Wiring Is Correct

Both Krea2 quality branches connect the target latent into Krea2EditModelPatch.

This is recommended by the Krea2Edit node author because it allows the source reference to be encoded before sampling rather than forcing VAE work during the sampler loop.

The workflow also uses:

fit_mode: fit

which is the recommended geometry mode for Krea2 Identity Edit v1.2.

---

5. V11.4 Experiment B Uses Qwen SFW v23 for Both Derived Branches

The workflow metadata identifies this build as:

V11.4 experiment B

with:

Portrait branch: Qwen SFW v23
Clothed full branch: Qwen SFW v23

This is described as the baseline for comparison against a ClothingV19 variant.

The embedded architecture note still references the earlier V11.1 fixed architecture, so the most accurate description is:

V11.4 Experiment B built on the V11.1 master-first architecture.

---

Suggested Usage

1. Install/update ComfyUI.
2. Install comfyui-krea2edit.
3. Place all required models and LoRAs in the expected folders.
4. Restart ComfyUI.
5. Load the workflow JSON.
6. Verify no nodes are red or missing.
7. Configure or inject the Bare Master character prompt.
8. Enable/bypass gender-specific LoRAs as appropriate.
9. Queue the workflow.
10. Collect:

• Bare Master
• Portrait
• Clothed Full Body

1. Compare identity consistency across all three images before using them downstream.

---

Troubleshooting

Missing TextEncodeQwenImageEditPlus

Update ComfyUI. This is a ComfyUI core node.

---

Missing Krea2EditModelPatch or Krea2EditGroundedEncode

Install:

cd ComfyUI/custom_nodes
git clone https://github.com/lbouaraba/comfyui-krea2edit

Restart ComfyUI.

---

Model Not Found

Confirm the filename and directory exactly match the values in the workflow.

Pay special attention to subfolders:

models/checkpoints/Qwen/Qwen-Rapid-AIO-SFW-v23.safetensors
models/loras/Krea2/Krea2_HMNSFW_AIO.safetensors

---

Identity Changes During Portrait Refinement

Try:

• keeping ref_boost high
• lowering output resolution
• disabling the final latent hires pass
• reducing denoise on the hires pass
• preserving fit_mode: fit
• keeping the same source images attached to both positive and negative grounded conditioning

---

Clothing Pass Changes the Face

The intended reference order is:

Bare Master
+
Final Portrait

Confirm both references are still connected to the Clothed Full structure and quality branches.

---

Responsible Use

This workflow includes adult-content-capable components.

Use it only in accordance with the licenses and acceptable-use terms of every model and LoRA in the stack.

Do not use this workflow to create:

• sexual content involving minors or age-ambiguous subjects
• non-consensual sexual imagery of real people
• deceptive real-person deepfakes
• unlawful abusive content
• content that violates a model's individual license

For real-person likenesses, obtain appropriate consent and rights before generation or publication.

The Krea2 Identity Edit project itself states that its identity-edit LoRA was trained on SFW data and does not endorse non-consensual, harmful, or sexual imagery of real people.

---

Licenses

This repository contains a workflow configuration and does not automatically grant redistribution rights for the third-party models referenced by it.

The following components may each have separate licenses:

• Krea2 base / derivative model
• Qwen3-VL text encoder
• Qwen Rapid checkpoint
• Krea2 Identity Edit LoRA
• realism LoRA
• NSFW LoRA
• anatomy LoRAs
• VAE

Verify each model's source repository and license before redistributing weights or publishing a bundled package.

ComfyUI-Krea2Edit is distributed separately and its repository documents the license for its node code and identity-edit weights.

---

Credits

This workflow combines technology from the ComfyUI ecosystem, including:

• ComfyUI
• Krea2
• Qwen3-VL
• Qwen Image Edit / Qwen structure editing
• ComfyUI-Krea2Edit
• Krea2 Identity Edit v1.2
• the respective creators of the additional checkpoints, VAEs, and LoRAs referenced by filename

Please credit the original model and node authors according to their individual licenses.

---

Version Notes

V11.4 Experiment B

• Uses the V11.1 master-first character architecture.
• Bare Master generated first.
• Portrait derived from Bare Master.
• Clothed Full derived from Bare Master + final Portrait.
• Portrait and Clothed Full structure branches both use Qwen Rapid AIO SFW v23.
• Krea2 Identity Edit v1.2 used for SFW identity/quality reconstruction.
• ConditioningZeroOut negative connection restored.
• Model paths use the root/subfolder layout stored in the exported workflow.
• Only three final image outputs remain.

---

References

• ComfyUI: https://github.com/Comfy-Org/ComfyUI
• ComfyUI Krea2Edit: https://github.com/lbouaraba/comfyui-krea2edit
• TextEncodeQwenImageEditPlus documentation: https://docs.comfy.org/built-in-nodes/TextEncodeQwenImageEditPlus