A unified, modular SDXL/Pony/Illustrious powerhouse designed to consolidate professional-grade tools into a single, cohesive environment. This workflow eliminates "tab fatigue" by integrating CFG/Noise engineering, multi-LoRAs management, ControlNet, IPAdapter, InstantID, advanced Upscaling, and Detailers into ONE interface.
Designed for speed and high-precision control, this workflow is essentially a "power workbench" overflowing with switches, sliders, and logic gates. It is undeniably complex, complex enough to make a beginner’s brain do a backflip, but for the pros and technical geeks - it’s a total playground. Think of it as the cockpit of a fighter jet: intimidating at first, but incredibly powerful once you know what the buttons do.
Don’t panic! I’ve left a breadcrumb trail of notes and mini-tutorials in every single section. Please, I’m begging you read them. If you skip the manual and things get weird, don't say I didn't warn you! Devour the notes before you hit that "Queue Prompt" button; your VRAM (and your sanity) will thank you.
Core Philosophy: Process in clear layers; fully configure the workflow in Row 1, generate and review a draft in Row 2, then produce the final high-quality render in Row 3. The completed image returns to Row 2 for side-by-side comparison before saving.
The workflow is intentionally organized into three horizontal zones for logical progression:
Row 1 (Top) – Setup & Configuration
Row 2 (Middle) – Draft Generation & Review
Row 3 (Bottom) – Final Processing & Output
🛠️ System Requirements & Hardware
GPU: NVIDIA (30xx/40xx Series).
VRAM: 12GB Minimum / 24GB Recommended.
System RAM: 32GB Minimum / 64GB+ Recommended.
🎯 Software Environment
ComfyUI Version: 0.15.0+
ComfyUI Frontend: 1.39.16+
💻 Core Software Environment
Python 3.12.x: (Highly recommended). While versions as low as 3.10, 3.11 may work, 3.12 provides the best compatibility for recent dependency updates). This version provides native support for nested f-strings and delivers improved execution speed.
PyTorch: 2.10.0+cu130 - (Minimum 2.6.0+cu124) specialized build specifically engineered for full compatibility with CUDA 13.0.
TorchVision: 0.25.0+cu130
TorchAudio: 2.10.0+cu130
📦 Essential Python Dependencies
These libraries are crucial for the stability of your Custom Nodes. Versions are strictly locked to prevent conflict.
Core: numpy==1.26.4 , setuptools>=70.0.0 , pydantic>=2.10.0 .
Imaging: pillow==10.x , opencv-python , blendmodes==2025 , glitch_this , rembg .
Identity & Face: insightface==0.7.3 , onnxruntime-gpu .
Miscellaneous: tensorflow
⚡ High-Performance Attention
Please read this article for more information: https://civitai.com/articles/21516/install-sageattention-2-notes .
SageAttention: Version 2.2.0 ( Mandatory optimization).
Flash Attention 2: (Fallback/Legacy support via Torch SDPA).
Flash Attention 3: Version 3.0.0+cu130 (Optimized for RTX 40 series).
Triton: 3.+ Windows-compatible build for FP8/NF4 dequantization.
🧩 Required Custom Node Packs
Core Infrastructure: rgthree-comfy , ComfyUI-Manager , ComfyUI_essentials , Use Everywhere (UE Nodes) , ComfyUI-LogicUtils , ComfyMath , ComfyUI-JNodes , ComfyLiterals .
CFG & Sampling: efficiency-nodes-comfyui , sd-perturbed-attention , Skimmed_CFG , RES4LYF , pre_cfg_comfy_nodes_for_ComfyUI , ComfyUI-Detail-Daemon .
Identity & Perception: ComfyUI_IPAdapter_plus , comfyui_controlnet_aux , ComfyUI_ZenID , ComfyUI-WD14-Tagger , ComfyUI-AdvancedLivePortrait , Extra Models for ComfyUI .
Upscaling & Motion: ComfyUI Impact Pack , ComfyUI Impact Subpack , UltimateSDUpscale , ComfyUI-SeedVR2_VideoUpscaler .
Creative Suite: JPS Custom Nodes , Comfyroll Studio , ComfyUI-KJNodes , ComfyUI_Primere_Nodes , ComfyUI-JakeUpgrade , ComfyUI-pause , ComfyUI-mxToolkit .
VFX & Realism: ComfyUI-Optical-Realism , CRT-Nodes , Virtuoso Nodes , ComfyUI_SKBundle , ComfyUI-FBCNN , comfyui_fill-nodes .
📥 Model Requirement
SDXL Checkpoint: Any SDXL/Pony/Illustrious
SDXL VAE: Baked VAE (or any SDXL VAE)
Upscale/ESRGAN: Any 4x upscale models.
ControlNet: xinsir/controlnet-tile-sdxl-1.0 ; ttplanet_sdxl_controlnet_tile ; diffusion_pytorch_model-instantID ; diffusion_pytorch_model_union_promax (or you can use separated models for every preprocesses)
IPAdapter: ip-adapter_sdxl_vit-h or ip-adapter-plus_sdxl_vit-h ; ip-adapter-faceid-plusv2_sdxl ; ip-adapter-faceid-portrait_sdxl ; ip-adapter-faceid-portrait_sdxl_unnorm
CLIP-Vision: CLIP-ViT-H-14 or CLIP-ViT-bigG-14
Instant-ID: ip-adapter_instant_id_sdxl
Detailers: Any detailer you prefer, I use person (#1), face (#2), eyes (#3), hair (#4), hands (#5), skin (#5), fashion (non-cond #1), furry (non-cond #2)
Others: depth_anything_v2_vitl (depth map); ema_vae_fp16 (SeedVR2); seedvr2_ema_7b_fp8_e4m3fn_mixed_block/seedvr2_ema_3b_fp8_e4m3fn (SeedVR2); clipseg-rd64-refined-fp16 (noise map)
🚩 Optimized Execution Flags (Optional)
Use these flags in your .bat file to leverage the full power of your GPU
--fast fp16_accumulation
--dont-upcast-attention
--use-sage-attention
--disable-xformers (because xformers only support python
--preview-method auto
--highvram/--normalvram/--lowvram
If you love my works, please review them with images and consider donating me some yellow buzz. This would be my great motivation to release more future works.
