File size: 9,440 Bytes

---
license: mit
---

# 🚀 ComfyUI Workflows by wizdroid

[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
[![ComfyUI](https://img.shields.io/badge/ComfyUI-Workflows-blue)](https://github.com/comfyanonymous/ComfyUI)
[![Flux](https://img.shields.io/badge/Flux-Accelerated-orange)](https://blackforestlabs.ai/)
[![Nunchaku](https://img.shields.io/badge/Nunchaku-Fast%20Inference-purple)](https://github.com/nunchaku-tech)

**High-quality, production-ready ComfyUI workflows** focused on **character consistency**, fast generation, and automated dataset creation for LoRA training.

These workflows power consistent character generation, multi-reference image-to-image pipelines, lightning-fast turbo inference, state-of-the-art upscaling, and intelligent dataset preparation.

---

## ✨ Highlights

- **Flux2-Klein Series** — Premium quality with Nunchaku acceleration for blazing fast inference
- **Multi-Reference i2i (1/2/3 images)** — Industry-leading character identity preservation using Reference Latent injection
- **Qwen & Z-Image-Turbo** — Versatile text-to-image and fast generation options
- **SeedVR2 Upscaler** — One of the best image upscalers available (DiT + VAE with advanced tiling & color correction)
- **WizdroidLoRADataset** — Custom node for automated vision-LLM captioning, dataset structuring, and validation — perfect companion for training consistent characters
- **LoRA-ready** — Built-in toggleable LoRA loading across most pipelines
- **Square 1:1 + Flexible Resolutions** — Optimized for character reference work

> **Pro Tip:** Pair these workflows with the **[Consistent Character Reference Prompts (PROMPTS.md)](https://huggingface.co/wizdroid/comfyui-workflows/blob/main/PROMPTS.md)** (40+ square 1:1 variations covering angles, expressions, beauty, upper/full body, and profiles) for unbeatable training data.

---

## 📁 Workflows

| Workflow                  | Type          | Size    | Key Features                                      | Best For |
|---------------------------|---------------|---------|---------------------------------------------------|----------|
| `flux2-klein-t2i.json`    | Text-to-Image | ~23 MB  | Nunchaku Flux2-Klein 9B, advanced Flux sampler, LoRA support | High-quality prompt-driven generation |
| `flux2-klein-1i2i.json`   | 1-Image Ref   | ~29 MB  | Single reference image + Reference Latent         | Strong character consistency from one photo |
| `flux2-klein-2i2i.json`   | 2-Image Ref   | ~35 MB  | Dual reference images                             | Even stronger identity with two angles/expressions |
| `flux2-klein-3i2i.json`   | 3-Image Ref   | ~41 MB  | Triple reference images + advanced conditioning   | Maximum character fidelity (recommended for LoRA training) |
| `qwen-t2i.json`           | Text-to-Image | ~14 MB  | Qwen-based, flexible resolution, LoRA             | Alternative aesthetic / artistic control |
| `qwen-aio-t2i.json`       | Text-to-Image | ~12 MB  | Qwen All-in-One variant                           | Quick Qwen generations |
| `qwen-aio-1i2i.json`      | 1-Image Ref   | ~14 MB  | Qwen image reference                              | Qwen-powered character consistency |
| `qwen-aio-2i2i.json`      | 2-Image Ref   | ~15 MB  | Dual reference                                    | Multi-ref Qwen |
| `qwen-aio-3i2i.json`      | 3-Image Ref   | ~16 MB  | Triple reference                                  | Maximum Qwen consistency |
| `z-image-turbo-t2i.json`  | Fast T2I      | ~18 MB  | Z-Image-Turbo (very few steps), Nunchaku          | Lightning-fast prototyping & iteration |
| `seedvr2-i2i.json`        | Upscaler      | ~13 MB  | SeedVR2 DiT + VAE upscaler, tiling, color correction | Best-in-class 2×–4× upscaling & detail recovery |
| `dataset-generator.json`  | Dataset Tool  | ~3 MB   | WizdroidLoRADataset + Ollama vision (Moondream)   | Automated LoRA training dataset creation with captions |

---

## 🛠️ Installation & Setup

### 1. ComfyUI
Make sure you have a recent [ComfyUI](https://github.com/comfyanonymous/ComfyUI) installation.

### 2. Required Custom Nodes
Install these via **ComfyUI Manager** (recommended) or git clone into `custom_nodes/`:

- **ComfyUI-nunchaku** — For accelerated Flux inference
- **ComfyUI-SeedVR2_VideoUpscaler** (ainvfx) — For the SeedVR2 upscaler workflow
- **[Wizdroid Character](https://github.com/wizdroid/wizdroid-character)** — Custom nodes for character prompting, multi-angle generation, animation adapters, I2V latent patches, and the `WizdroidLoRADataset` for automated captioning & dataset creation
  ```bash
  git clone https://github.com/wizdroid/wizdroid-character custom_nodes/wizdroid-character
  ```

Restart ComfyUI after installing custom nodes.

### 3. Models
Place models in the standard ComfyUI folders (or update paths inside workflows):

**Flux2-Klein (main quality engine):**
- `flux/2/flux2-klein-9B.safetensors`
- `flux/2/vae.safetensors`

**Qwen:**
- `qwen/qwen-image-edit-sfw.safetensors`
- `qwen/qwen3-8B.safetensors` (or 4B variant)
- `qwen/qwen3-4B.safetensors`

**Z-Image Turbo:**
- `z-image/z-image-turbo.safetensors`
- `z-image/vae.safetensors`
- `z-image/turbo/4nup4m4.safetensors` (example turbo LoRA)

**SeedVR2:**
- `seedvr2_ema_7b_sharp_fp16.safetensors` (DiT)
- `ema_vae_fp16.safetensors` (VAE)

**CLIP / Other:**
- Various Qwen CLIP models as referenced in the workflows

### 4. Ollama (for Dataset Generator)
The `dataset-generator.json` workflow uses a local vision model via Ollama:

```bash
# Install Ollama and pull a good vision model (Moondream recommended in the workflow)
ollama pull moondream:latest
```

Start Ollama (`ollama serve`) before running the dataset workflow. You can change the model/URL inside the node.

---

## 🚀 How to Use

1. **Load a workflow**: Drag & drop any `.json` file directly onto your ComfyUI canvas, or use **Load** → select the file.
2. **Queue Prompt** as usual.
3. Most workflows expose clean proxy widgets on the right for:
   - Prompt / negative prompt
   - Seed, steps, CFG, sampler
   - Width / Height
   - LoRA toggle + strength
   - Reference images (for i2i workflows)

### Character Consistency Workflow Recommendations

- Use **flux2-klein-3i2i.json** (or 2i2i) with 2–3 high-quality, varied reference photos of your character.
- Generate using the same character across many angles/expressions using the reference prompts collection.
- Feed the resulting images into `dataset-generator.json` to auto-caption and structure a ready-to-train dataset (with your character trigger tag).

---

## 💡 Tips for Best Results

- **Reference Images**: Clean, well-lit, minimal makeup/jewelry/background. Square 1:1 references work great.
- **Prompting**: Start simple ("Photo of a woman", "beautiful detailed face") and let the reference images do the heavy lifting for identity.
- **LoRAs**: The LoRA switch is disabled by default in many workflows. Enable it and point to character or style LoRAs.
- **Speed vs Quality**: Use `z-image-turbo-t2i.json` for rapid iteration, then switch to Flux2-Klein for final outputs.
- **Upscaling**: Always run important generations through `seedvr2-i2i.json` for that extra polish.
- **Dataset Quality**: Use the dataset generator + good reference images + Ollama captions, then manually review the `validation_report`.

---

## � Full Guide: Mastering Consistent Characters

For the **complete step-by-step tutorial** (problem → solution → exact pipeline using these workflows + PROMPTS.md + the WizdroidLoRADataset node → pro tips + visual examples), see the dedicated CivitAI article:

**→ [Mastering Consistent Characters for LoRA Training](https://civitai.com/articles/31442/mastering-consistent-characters-for-lora-training)**

The article covers:
- Why single-reference methods fail for modern models (Flux, SD3, Aurora, etc.)
- How Reference Latent injection + clean square 1:1 prompts solve identity drift
- Generating 40–100+ varied training images using the included PROMPTS.md
- Using the WizdroidLoRADataset node for perfect auto-captions, validation reports, and trigger word injection
- Training recommendations, best practices, and real-world results

This README is the technical reference. The CivitAI guide is the teaching + visual companion. Use both.

---

## �📜 License

MIT License — feel free to use, modify, and share these workflows.

---

## 🤝 Contributing

Pull requests and workflow improvements are welcome! If you create amazing results with these, tag **@wizdroid** — I'd love to see what you build.

---

**Made with ❤️ for the ComfyUI community by wizdroid**

*Consistent characters start with great references and great tools.*

---

**Additional Resources**
- [PROMPTS.md — 40 Consistent Character Reference Prompts](https://huggingface.co/wizdroid/comfyui-workflows/blob/main/PROMPTS.md)
- [CivitAI Guide: Mastering Consistent Characters for LoRA Training](https://civitai.com/articles/...) *(update this link after posting the article)*
- Full workflows and updates in this repository31442/mastering-consistent-characters-for-lora-training)

**For HF users:** The `HF_README_UPDATED.md` file in this folder is a clean, concise version you can copy directly into the repo root as `README.md` if you prefer a shorter model card. The current README is the rich version recommended for the repo.