Caption + SDXL aDetailer + Upscale

ChaosCreationAIChaosCreationAI
11
230
upscale도ꡬcaptionadetailersdxl
v1.0
model-image-0

μ„ΈλΆ€ 정보

파일 λ‹€μš΄λ‘œλ“œ (1)

λͺ¨λΈ μ„€λͺ…

πŸ“Œ Caption + SDXL aDetailer + Upscale β€” Workflow Overview

This workflow is designed as an end-to-end enhancement pipeline for improving existing images using three major stages:

  1. Automatic Captioning (JoyCaption)

  2. Multi-stage SDXL aDetailer refinement

  3. High-quality 2Γ—/4Γ— upscaling

It is especially useful for batch-processing portrait images, NSFW or SFW, while keeping full control over detail restoration (faces, hands, skin, body parts), composition integrity, and final resolution.

πŸ”Ή 1. Automatic Captioning (JoyCaption)

The pipeline begins with JoyCaption:

  • Loads images from a directory via ImageBatchPath

  • Applies the user-defined Extra Options block (lighting, camera info, text removal, etc.)

  • Produces a clean SDXL-style prompt

  • The prompt is shown via ShowText and sent directly into the conditioning stack

This step is ideal for:

  • Captioning large datasets

  • Creating prompts from real photos

  • Generating consistent SDXL prompts for further enhancement or LoRA training

πŸ”Ή 2. Multi-Stage SDXL aDetailer Refinement

The core of the workflow is a chain of aDetailer passes, each dedicated to a specific target area:

  • Face

  • Hands

  • Teeth

  • Armpits

  • Breasts

  • Vagina

  • General body segmentation

Each pass includes:

  • YOLOv8 detectors from the Impact Pack

  • SAM for accurate mask generation

  • SDXL inpaint refinement with positive/negative conditioning

  • Full previews for each refinement stage

This allows extremely fine-grained control and produces:

  • Cleaner face geometry

  • Corrected hands

  • Better skin continuity

  • Consistent anatomy in NSFW work

  • Reduced artifacts from low-quality images

The workflow is modular: users can disable or remove specific detailers as needed.

πŸ”Ή 3. High-Resolution Upscaling (2Γ— or 4Γ—)

Upscaling is handled through:

  • UpscaleModelLoader (4x_NMKD-Siax_200k)

  • ImageUpscaleWithModel for AI-based enhancement

  • ImageScaleBy for additional clean integer scaling

  • Two final outputs:

    • 2Γ— upscaled

    • 4Γ— upscaled

This results in:

  • Clean, crisp edges

  • Preservation of style

  • Better texture definition

  • High-resolution images suitable for posting or further processing

The Set/Get image node pair provides a lightweight way to reuse the intermediate image cleanly across the upscale branch.

이 λͺ¨λΈλ‘œ λ§Œλ“  이미지