SDXL Omni IMG Gen

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 <= 3.10.11)

• --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.