🔱 Trinity SDXL Workflow — V2.0

A Modular SDXL Workflow for Speed, Quality, Img2Img & Inpainting

A powerful, flexible, multi-stage workflow designed primarily for Illustrious-based models, with support for other SDXL model families, including compatible Pony and NoobAI checkpoints.

Your Workflow, Your Balance: Speed or Quality.

Trinity lets you choose between a lighter, faster generation process and a more complete, quality-focused pipeline with additional refinement, detailing, and upscaling. Instead of forcing every user to run every stage, Trinity provides a modular structure that lets you adapt the workflow to your hardware, model, and creative goals.

Version 2.0 builds on the original Trinity workflow with an optional SageAttention Speed Boost, refined generation settings, a redesigned layout, improved metadata prompt formatting, and a simplified custom-node dependency list.

🖥️ Tested System

Tested on:

• GPU: NVIDIA GeForce RTX 3070 Ti

• RAM: 32 GB

• CPU: Intel Core i5-9400F

• OS: Windows 11

Note: Performance and generation times may vary depending on your hardware, base resolution, selected model, and workflow settings. These specifications represent the system used for testing, not the minimum system requirements.

✨ What's New in Version 2.0?

• ⚡ Optional SageAttention Speed Boost: Added a dedicated performance-optimization group for supported systems. It is bypassed by default and can be enabled after installing and verifying a compatible SageAttention setup.

• 🎯 Refined Generation Settings: Selected Detailer settings and the 3rd Generation stage have been adjusted to improve the workflow's overall refinement process.

• 🎨 Redesigned Layout: Reorganized nodes and refreshed the overall visual structure for a cleaner, more convenient workflow experience.

• 📝 Improved Metadata Prompt Formatting: Refined how prompts are formatted and written by the Civitai Metadata Saver.

• 🧩 Simplified Dependencies: Removed the comfy_gps_Spplements custom-node pack from the workflow's dependency list.

• 🧪 Expanded Model Testing: Tested with multiple Illustrious-based checkpoints, including WAI-illustrious-SDXL, oneObsession, Ultimate Hentai Anime RX, Nova 3DCG XL, Nova Anime XL, Nova Anime3D XL, and AniMayhem – Pale Rider.

The original core features remain: multi-stage generation, optional detailers, upscaling, Skin Contrast, independent seed controls, Empty Latent, Img2Img, Inpainting, and Civitai-compatible metadata saving.

🚀 Key Features

• Multi-stage generation: 1st Generation → 2nd Generation → 3rd Generation.

• Flexible refinement: Independent controls for the main generation stages.

• Optional Person, Face, and Hand Detailers: Refine anatomy, facial features, and hands when needed.

• High-quality upscaling: Increase output resolution and improve visual clarity.

• Skin Contrast enhancement: Add subtle contrast and definition.

• New Image, Img2Img, and Inpainting: Switch between different image-generation modes.

• Independent seed control: Control randomness at each generation stage.

• Fast Group Bypasser: Quickly switch between lighter and more complete workflow configurations.

• Optional SageAttention Speed Boost: Test an alternative attention implementation for potential performance improvements.

• Civitai Metadata Saver: Save generated images with embedded generation information.

• Modular node layout: Enable or bypass optional sections according to your needs.

⚡ Speed vs. Quality — Choose Your Workflow

Trinity is designed around flexibility. You do not always need to run every stage to create an image.

⚡ Fast Generation

Use a lighter configuration when iteration speed matters most. Bypassing optional refinement and upscaling stages can significantly reduce the total processing time.

⚖️ Balanced Generation

Keep the stages that contribute most to your desired result while bypassing less essential processing. This is a practical starting point for everyday use.

💎 Quality-Focused Generation

Use the full multi-stage pipeline, optional detailing, and upscaling when you want more refinement and a higher-resolution final image.

Important: These are workflow configurations, not guaranteed quality presets. The best balance depends on the checkpoint, prompt, resolution, hardware, and personal preference. A lighter configuration may produce excellent results, but bypassing refinement or upscaling can reduce certain details.

⚙️ Hardware and Resolution Matter

Trinity can be adapted to different systems, but it cannot eliminate hardware limitations. GPU performance, available VRAM, system memory, base resolution, model architecture, sampling steps, and enabled stages all affect generation time and resource usage.

A powerful GPU can generally handle more demanding configurations comfortably, while a more limited system may benefit from lower base resolutions and fewer active stages.

Higher base resolutions can substantially increase computational and memory requirements. For a fair speed comparison, keep the model, seed, resolution, sampler, steps, and other settings unchanged.

⚡ Speed Boost — SageAttention

Version 2.0 introduces an optional SageAttention optimization through the ComfyUI-KJNodes Patch Sage Attention KJ node.

SageAttention can make attention computation more efficient and may reduce generation time on compatible systems. The actual improvement varies according to the GPU architecture, model, resolution, sampling settings, installed software versions, and selected backend.

Default State

The entire Speed Boost group is bypassed by default.

You can use Trinity without enabling SageAttention. The standard attention path remains available by bypassing the group.

To enable Speed Boost:

1. Install a compatible SageAttention package into the Python environment used by ComfyUI.

2. Install any additional backend dependency required by your GPU and chosen implementation.

3. Run the recommended SageAttention test and confirm that it succeeds.

4. Restart ComfyUI completely.

5. Remove the Speed Boost group from Bypass.

6. Start with sage_attention = auto and allow_compile = false.

Do not enable the group before verifying the installation.

Recommended Settings

• sage_attention = auto

• allow_compile = false

Automatic backend selection is the recommended starting point. Manual backends and compilation-based optimizations should only be tested when you understand their compatibility requirements.

SageAttention is a performance optimization, not a model enhancement. It does not replace the checkpoint, LoRAs, VAE, or refinement stages, and it does not guarantee better image quality.

If You Encounter Problems

If enabling Speed Boost causes errors, instability, or unexpected output differences, bypass the entire group first. If the workflow works normally afterward, investigate the SageAttention installation, backend, or software compatibility.

The KJNodes interface and the SageAttention Python library are separate requirements. Depending on your GPU and backend, Triton may also be needed.

For installation instructions, environment checks, compatibility guidance, and troubleshooting, follow the SageAttention Setup Guide included with the workflow.

📦 Required Custom Nodes

Install the custom-node packages needed by the workflow before use.

• Comfyroll Custom Nodes

• rgthree-comfy

• ComfyUI-Essentials

• comfyui_controlnet_aux

• Crystools

• ComfyUI-Image-Saver

• ComfyUI-Impact-Pack

• ComfyUI-Impact-Subpack

• ComfyUI-KJNodes

• ComfyUI_UltimateSDUpscale

❗The comfy_gps_Spplements pack used in the previous version is no longer listed as a dependency in V2.0.

Important: This list covers the workflow's custom-node packs, not the model files themselves. You must also have the appropriate checkpoints, VAE, LoRAs, ControlNet models, and upscalers installed for the features you intend to use.

📘 Workflow Stage Guide

Optional sections can be bypassed when they are not needed. Avoid bypassing core generation stages unless you intentionally want a lighter workflow or understand how the selected switches affect the pipeline.

🚀 Speed Boost (Optional):

May reduce generation time using optimized SageAttention kernels on compatible systems. Actual speed improvements and output differences depend on your hardware, model, resolution, backend, and software environment.

1. Loaders

Loads the base checkpoint, CLIP, VAE, LoRAs, embeddings, and other primary resources used throughout the workflow.

AniMayhem – Pale Rider is the default recommended checkpoint, but compatible alternatives can be selected.

2. Prompt Processing

Processes positive and negative prompts and prepares the conditioning data used for generation.

3. Tools — Resolution & Randomize

Contains resolution controls, the Fast Group Bypasser, and independent seed controls for the 1st, 2nd, and 3rd Generation stages.

Use fixed seeds when comparing settings and randomize them when exploring new variations.

4. ControlNet — Optional

Provides pose guidance through compatible ControlNet models, helping guide character positioning and structural consistency.

Its usefulness depends on the selected model, input pose, and desired result. Bypass it when pose guidance is unnecessary.

5. 1st Generation

Creates the initial image, establishing the overall composition, characters, environment, and main visual elements.

This stage forms the foundation for subsequent refinement.

6. 2nd Generation

Refines selected image details using a secondary generation stage and its dedicated prompt and settings. Depending on the configuration, this stage can improve facial features, eyes, hands, skin appearance, and expression.

The final effect depends on denoise strength, prompts, LoRAs, and the checkpoint being used.

7. 3rd Generation

Provides an additional refinement stage designed to reinforce the original prompt and character identity while preserving useful improvements from earlier stages.

Its settings have been refined in V2.0. The degree of change depends on the selected configuration and denoise strength.

Bypass this stage when prioritizing speed, or retain it when its contribution benefits your particular image.

8. Person Detailer — Optional

Detects and refines people in the image, with an emphasis on full-body appearance and anatomy.

It can be useful when a character needs additional refinement beyond the face alone.

9. Face Detailer — Optional

Detects and enhances visible faces, helping refine facial features and other localized details.

Its effectiveness depends on face detection, image resolution, and the selected settings.

10. Hand Detailer — Optional

Detects and refines hands to help correct finger and hand anatomy while preserving the overall pose.

Results can vary depending on hand visibility, occlusion, resolution, and the source image.

11. Upscaler — Optional

Increases the output resolution and can improve sharpness, perceived detail, and overall clarity.

Upscaling is more computationally demanding than simply generating a smaller image. The final result also depends on the chosen upscaler and its settings.

12. Add Skin Contrast — Optional

Applies a contrast-enhancement stage intended to improve skin definition and local tonal separation.

The effect is subtle and can be bypassed if it does not suit the image or model.

13. Civitai Metadata Saver — Optional

Saves the final image with embedded generation metadata, including the available prompt, model, LoRA, workflow, and settings information.

The metadata-bearing output is the preferred file to keep or upload when you want to preserve generation details. Preview images or files saved through other routes may not contain the same metadata.

📐 Recommended Resolutions

The following are starting points, not mandatory settings.

Use case | Base resolution

🖥️🖥️Dual-monitor ultrawide wallpaper | 1704 × 480

🖥️Standard desktop | 1280 × 720

❎Square image | 1280 × 1280 or 720 × 720

📱Phone wallpaper — 19.5:9 | 544 × 1176

The selected Empty Latent resolution is the starting resolution, not necessarily the final saved resolution. When the upscaler is enabled, the final image can be substantially larger than the base image.

Choosing the Right Resolution

• Start with a lower base resolution if you need faster iteration or have limited VRAM.

• Use higher resolutions when the hardware can handle them and the image benefits from the additional detail.

• Remember that changing the aspect ratio or resolution can change composition, character proportions, and the way a model interprets the prompt.

• Do not assume that two systems will produce identical generation times or results at the same resolution.

⚡ Speed vs. Quality Guide

For a substantially lighter workflow, consider the following configuration changes.

1. ControlNet: Bypass if pose guidance is unnecessary.

2. 2nd Gen Tools: Set all linked switches to Input 2 when using the lightweight configuration.

3. CR Image Input Switch: Set to Input 2 as required by the selected generation path.

4. 3rd Generation: Bypass when its additional refinement is not needed.

5. Person Detailer: Bypass if full-body detailing is unnecessary.

6. Face Detailer: Bypass if additional face refinement is unnecessary.

7. Hand Detailer: Bypass if additional hand refinement is unnecessary.

8. Upscaler: Bypass when a lower-resolution output is sufficient.

9. Add Skin Contrast: Bypass if this finishing step is not needed.

Important: Steps 2, 3, and 4 are linked and should be changed together when using the corresponding lightweight configuration.

Two balanced configurations from the original workflow are useful starting points:

• 2, 3, 4, 8

• 2, 3, 4, 5, 8

• 1, 2, 3, 4, 5, 8

These refer to the numbered changes above, not generation-stage numbers.

For Civitai uploads, keeping the Civitai Metadata Saver enabled is recommended.

The SageAttention Speed Boost is a separate optional optimization. It can be tested independently of these workflow simplifications, provided it has been installed and verified correctly.

🔀 Latent Mode Guide — Img2Img & Inpainting

Trinity supports creating new images, transforming existing images, and inpainting masked areas.

Black Switch

• 1 = Empty Latent — create a new image.

• 2 = Img2Img — use an existing image as the starting point.

Purple Switch

• 1 = Empty Latent / Img2Img path.

• 2 = Inpainting path.

Typical Configurations

New image 👉 ⬛Black Switch : 1 🟪Purple Switch : 1

Img2Img 👉 ⬛Black Switch : 2 🟪Purple Switch : 1

Inpainting with a mask 👉 ⬛Black Switch: As required by the input path 🟪Purple Switch: 2

For inpainting, provide the appropriate image and mask. The switches select the processing path; they do not create a mask automatically.

When using Img2Img or Inpainting, the source image, mask, denoise strength, and selected model all influence the result.

⚙️ Recommended Generation Settings

The following values are starting points based on the original workflow. Different checkpoints may benefit from different settings.

• Steps: 20–30

• CFG Scale: 3.0–5.0

• Sampler + Scheduler: Euler + Simple or DPM++ 2M SDE + Karras

• 1st Generation Denoise: 1.0 for a standard new-image generation path.

• Other-stage Denoise: 0.18–0.40 as a starting range.

• Seed: Use a fixed seed for testing and comparison; randomize it for new variations.

Understanding Denoise Strength

Higher denoise values allow a stage to make stronger changes to its input, but may alter composition, character appearance, or details established in earlier stages.

Lower values generally preserve more of the existing image, although the exact behavior depends on the model and stage.

Avoid changing every denoise value at once. Adjust one stage at a time so you can identify which settings actually improve your result.

When changing the base checkpoint, you may also need to adjust the sampler, CFG, steps, and denoise values.

📦 Models & LoRAs

Trinity is primarily optimized around Illustrious-based checkpoints, with additional testing across several compatible models.

Default Recommended Base Model

AniMayhem – Pale Rider

https://civarchive.com/models/1984400/animayhem-pale-rider

📂 directory:

ComfyUI/models/checkpoints/

Additional Tested Checkpoints

The following models have been tested with Trinity V2.0:

• WAI-illustrious-SDXL

• oneObsession

• Ultimate Hentai Anime RX

• Nova 3DCG XL

• Nova Anime XL

• Nova Anime3D XL

• AniMayhem – Pale Rider — default recommended checkpoint

Testing with these models does not mean that every model uses identical settings or produces identical results. Different checkpoints may respond differently to prompts, LoRAs, samplers, CFG, and denoise strength.

Pale Rider remains the default reference model for the workflow. If you switch to another checkpoint, use the recommended settings as a starting point and adjust them when necessary.

ControlNet

Illustrious-XL ControlNet OpenPose

Use a compatible OpenPose ControlNet model for the pose-guidance stage.

📂 directory:

ComfyUI/models/controlnet/

Upscalers

• 4x-UltraSharp.pth

• 1xSkinContrast-High-SuperUltraCompact

📂 directory:

ComfyUI/models/upscale_models/

The first is used for image upscaling, while the second is associated with the Skin Contrast enhancement stage.

Selectable VAE Models

• FTastic VAE

• HDR Effect VAE (V3 FP32)

• XL_VAE_C (G10.2)

📂 directory:

ComfyUI/models/vae/

The appropriate VAE can depend on the checkpoint and the desired output. An incompatible or unsuitable VAE may affect colors, contrast, and image decoding.

Recommended LoRAs

The following LoRAs were included among the original workflow's recommended resources:

• Scenery Enhancer - illustrious (New Rec)

• Breast Implants – PonyXL

• Ri-mix – Style LORA [Illustrious + Anima]

• Detailer | Tool / Concept LoRA IllustriousXL

• Face Enhancer Illustrious

• Eyes for Illustrious Lora

• Smooth Detailer Booster

• Well-Drawn Hands

📂 directory:

ComfyUI/models/loras/

These are recommendations, not mandatory dependencies for every generation. LoRA compatibility and effectiveness vary by checkpoint. Pony-focused LoRAs, in particular, should be evaluated with the model family they were designed for.

For the best results, ensure that any LoRA intended to contribute consistently across multiple generation stages is enabled at the appropriate stages with compatible weights.

📝 Civitai Metadata Saver — Important Notes

The Metadata Saver is intended to preserve the generation information associated with the final image.

For the best results:

• Keep the Metadata Saver enabled if you want to upload the resulting image to Civitai with its generation metadata.

• Use the image saved by the Metadata Saver rather than relying on a preview image to preserve all available information.

• Check the prompt formatting and model selection when switching checkpoints.

⚠️ Missing Model Warning — How to Fix It

When opening Trinity, you may occasionally see a Missing Models warning in ComfyUI's Workflow Overview, especially after changing the default checkpoint or opening the workflow for the first time.

In this workflow, the warning may refer to the model selection stored in the Civitai Metadata Saver rather than a missing checkpoint required by the generation pipeline.

If generation resources are installed correctly and the warning identifies the Metadata Saver's model selection, follow these steps:

1. Click Show Missing Models to open the missing-model list.

2. Locate the model entry associated with the Metadata Saver.

3. Open the Use from Library dropdown.

4. Select the checkpoint you actually intend to use.

5. Confirm the selection by clicking the checkmark.

This updates the saved model selection in the Metadata Saver to match your chosen checkpoint.

Important: This explanation applies when the missing entry is only the Metadata Saver's model reference. If ComfyUI reports that a checkpoint or another resource required by the actual generation pipeline is missing, install or select the correct resource separately.

After correcting the Metadata Saver's model selection, the warning should no longer appear for that same saved selection. If you change checkpoints again, you may need to update the metadata selection accordingly.

📌 Important Notes

• Model compatibility: Trinity is primarily designed for Illustrious-based checkpoints. Other SDXL-family models may require adjustments to sampling settings, CFG, denoise strength, and LoRA selection.

• Hardware differences: Generation speed, VRAM usage, and practical output resolution depend on the GPU, available memory, and active workflow stages.

• Resolution differences: Higher base resolutions increase computational demands and may produce different compositions or details.

• Quality vs. speed: Bypassing stages can save time but may remove refinements that matter to a particular image. Compare results before deciding which stages to keep.

• SageAttention: The Speed Boost group is optional and bypassed by default. Install and verify a compatible setup before enabling it.

• Metadata: The image saved through the Civitai Metadata Saver is the recommended output when preserving generation information matters.

• Seeds: Fixed seeds are useful for controlled comparisons, but different models, settings, or software implementations can still produce different results.

• LoRA consistency: For LoRAs intended to influence multiple generation stages, use compatible settings and weights across those stages where appropriate.

• Custom nodes and models: Installing the custom-node packages does not automatically install checkpoints, LoRAs, VAEs, ControlNet models, or upscalers.

🔱 Enjoy Creating with Trinity SDXL!

Trinity SDXL Workflow V2.0 — Flexible by design. Built for your balance of speed and quality.

If you use or share this workflow, please credit Trinity SDXL Workflow.

Your support helps the Trinity project continue to grow and evolve. 💫