If you’re using ComfyUI and want to try Hydream, this guide will help you install, configure, and optimize it for the best results. I recently set it up myself, and in this article, I’ll share the exact steps, performance tips, and workflow optimizations that worked for me.
Why Use Hydream in ComfyUI?
Hydream is gaining popularity because of its strong prompt adherence and image structure control. Unlike some AI models that ignore parts of your text input, Hydream follows instructions more accurately. Plus, with native ComfyUI support, you don’t need extra custom nodes—just download the right files and start generating.
What You’ll Need
Before starting, make sure you have:
✅ ComfyUI installed (updated to the latest version)
✅ At least 16GB VRAM for the full FP8 model (lower VRAM options available)
✅ Basic knowledge of ComfyUI workflows
Step 1: Download the Right Model Files
Hydream offers multiple versions:
- FP8 (Best Quality) – 17GB file, ideal for high-end GPUs
- FP16 – Smaller size, good balance between speed and quality
- GGUF (Low VRAM Option) – Quantized models for weaker hardware
🔗 Where to Download:
- Official Hugging Face repository (links in description)
Step 2: Install & Organize Model Files
I recommend this folder structure inside your ComfyUI/models/ directory:
📂 models/
├── 📂 clip/ (For text encoders)
├── 📂 diffusers/hydream/ (Main model goes here)
└── 📂 vae/ (Reuse Flux VAE if available) 💡 Pro Tip: If you’ve used Flux before, you can reuse:
- VAE files (same as Flux)
- CLIP/T5XL text encoders (just add the new Llama 3.1B file)
Step 3: Load Hydream in ComfyUI
Once files are in place, setting up the workflow is simple:
- Use the native “Load Diffusion Model” node (no custom nodes needed).
- Load CLIP text encoders (supports multiple files).
- Connect to a VAE loader (use your existing Flux VAE if available).
I’ve included a ready-to-use workflow JSON in the description—just drag and drop it into ComfyUI!
Step 4: Optimizing Performance
From my tests:
- Best Sampler for Full Model: UniPC + Simple scheduler
- Best for Fast Generations: LCM + Normal scheduler (no negative prompts needed)
- Upscaling: SD Ultimate Upscale works well (denoise 0.2, 30 steps)
⚠️ VRAM Warning:
- The FP8 model uses ~16GB VRAM at 1024x resolution.
- If you’re on a weaker GPU, try the GGUF quantized versions instead.
Step 5: Testing & Comparing Results
I ran multiple tests comparing Hydream with Flux and SDXL. Here’s what stood out:
✅ Better prompt adherence – Follows detailed descriptions well.
✅ Stronger structure control – Keeps compositions intact.
❌ Image-to-image not working yet – Stick to text-to-image for now.
Final Thoughts
Hydream is a powerful addition to ComfyUI, especially if you need precise AI-generated images. The setup is straightforward, and performance is solid—just make sure you pick the right model version for your hardware.
🔗 Download the workflow & models here
city96/HiDream-I1-Full-gguf at main
city96/HiDream-I1-Dev-gguf at main
city96/HiDream-I1-Fast-gguf at main
city96/t5-v1_1-xxl-encoder-gguf at main
bartowski/Meta-Llama-3.1-8B-Instruct-GGUF at main
add-detail-xl.safetensors · trieult/lora_add_details_xl at main
RealVisXL_V3.0_Turbo.safetensors · SG161222/RealVisXL_V3.0_Turbo at main
Comfy-Org/HiDream-I1_ComfyUI at main
