ComfyUI Workflow 2026
ComfyUI hit 106K GitHub stars in 2026. Beginner-friendly setup guide, model recommendations for 2026, and 5 production-ready workflow templates (text-to-image, inpaint, upscale, video, character consistency).
- ⭐ 6056
- ComfyUI
- Stable Diffusion
- Python
- CUDA
- GPL-3.0
- Updated 2026-05-25
{{< resource-info >}}
Meta Description: ComfyUI hit 106K stars in 2026. Setup guide + 5 production-ready workflow templates (text-to-image, inpaint, upscale, video, character consistency).
ComfyUI became the default tool for serious AI image generation in 2026. Node-based, reproducible, automatable. This guide gets you from zero to 5 working production workflows in an afternoon.
⚡ TL;DR #
Why ComfyUI: workflow reproducibility, automation, complex pipelines. 106K stars in 2026.
Hardware: 8GB VRAM minimum, 16GB+ comfortable.
Setup time: 1 hour to first generation.
5 templates below: text-to-image, inpaint, upscale chain, video, character consistency.
Setup (1 Hour) #
Step 1: Install (15 min) #
# Clone + setup venv
git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
python main.py
Browser opens at http://localhost:8188.
Step 2: Download models (30 min) #
Drop into ComfyUI/models/checkpoints/:
- SDXL base + refiner (most versatile, ~13GB total)
- Flux.1 Schnell (fast prototyping, ~24GB)
- SD 3.5 Large (best photorealism, ~17GB)
Optional but useful:
- 2-3 LoRAs for your style (Civitai, search “2026 SDXL trending”)
- ControlNet models (OpenPose, Depth, Canny — each ~1.5GB)
Step 3: First generation (15 min) #
- Drag the default workflow from
ComfyUI/workflows/ - Load SDXL checkpoint
- Type prompt
- Queue prompt
Done. You’re now generating. The hard part starts now: building reusable workflows.
5 Production-Ready Templates #
Template 1: Text-to-Image (SDXL base + refiner) #
Use: standard generation, daily workhorse. Nodes: Load Checkpoint → CLIP Text Encode (prompt) → KSampler (base 70%) → KSampler (refiner 30%) → VAE Decode → Save Image. Time per image: 8-15 sec on RTX 4090.
Template 2: Inpaint Mask (selective editing) #
Use: change a specific area without regenerating the whole image. Nodes: Load Image → MaskEditor → CLIP Text Encode (new content) → InpaintModelConditioning → KSampler → Composite back. Time per edit: 5-10 sec.
Template 3: 4x Upscale Chain (4K output) #
Use: take 1024×1024 generation → 4096×4096 production-ready output. Nodes: Generate at 1024 → Upscale Latent 2x → KSampler refine pass → Upscale Latent 2x again → Final refine. Time: 30-45 sec per image at 4K.
Template 4: Image-to-Video (5 sec clip) #
Use: animate a still into 5-second motion clip. Nodes: Load SVD model → Load Image → Image to Video (24 frames @ 8fps) → VAE Decode → Save Video. Time: 60-90 sec on RTX 4090. 2026 model: Stable Video Diffusion XT or LTX Video.
Template 5: Character Consistency (LoRA + IPAdapter) #
Use: generate same character across many scenes with consistent face. Nodes: Load LoRA (character-trained) + IPAdapter (reference image) → CLIP Text Encode → KSampler → output. Time per image: 12-20 sec. Trick: train your own LoRA on 15-20 source images of one character — IPAdapter handles the rest.
Workflow Sharing #
All five templates are saveable as .json. Drag onto ComfyUI canvas to load. Share with team via git or Discord.
The community publishes thousands of workflows at:
- ComfyUI subreddit
- OpenArt.ai workflow library
- Civitai (look for “ComfyUI workflow” filter)
Bring 2-3 community workflows in and customize for your style. That’s how most production artists work — not building from scratch.
Recommended Infrastructure #
For serious ComfyUI work:
- DigitalOcean — $200 credit, GPU droplets (H100/L40S/A100)
- HTStack — Hong Kong VPS for low-latency Asia generation
Affiliate links — same price, supports dibi8.com.
Conclusion #
ComfyUI’s learning curve is real but the payoff is real. Once you have 5 reusable workflows, you’re shipping faster than any single-shot tool. The 2026 ecosystem (Flux, SD 3.5, IPAdapter, ControlNet improvements) is the most capable image-generation stack ever assembled — and ComfyUI is the only tool that orchestrates it cleanly.
Start with the 5 templates above. Customize. Share. The compound returns of reusable workflows show up after week 2 — when you realize you’re combining nodes faster than you’d write code.
Related: Stable Diffusion WebUI Setup · Top AI Image Generators 2026 · Local-First AI Stack 2026
💬 Discussion