| --- |
| license: mit |
| --- |
| |
| # π ComfyUI Workflows by wizdroid |
|
|
| [](https://opensource.org/licenses/MIT) |
| [](https://github.com/comfyanonymous/ComfyUI) |
| [](https://blackforestlabs.ai/) |
| [](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. |
|
|
| --- |
|
|
| ## β¨ Highlights |
|
|
| - **Flux2-Klein Series** β Premium quality with Nunchaku acceleration for blazing fast inference |
| - **Multi-Reference i2i (1/2/3 images)** β Industry-leading character identity preservation using Reference Latent injection |
| - **Qwen & Z-Image-Turbo** β Versatile text-to-image and fast generation options |
| - **SeedVR2 Upscaler** β One of the best image upscalers available (DiT + VAE with advanced tiling & color correction) |
| - **WizdroidLoRADataset** β Custom node for automated vision-LLM captioning, dataset structuring, and validation β perfect companion for training consistent characters |
| - **LoRA-ready** β Built-in toggleable LoRA loading across most pipelines |
| - **Square 1:1 + Flexible Resolutions** β 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. |
|
|
| --- |
|
|
| ## π 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 2Γβ4Γ upscaling & detail recovery | |
| | `dataset-generator.json` | Dataset Tool | ~3 MB | WizdroidLoRADataset + Ollama vision (Moondream) | Automated LoRA training dataset creation with captions | |
|
|
| --- |
|
|
| ## π οΈ 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** β For accelerated Flux inference |
| - **ComfyUI-SeedVR2_VideoUpscaler** (ainvfx) β For the SeedVR2 upscaler workflow |
| - **[Wizdroid Character](https://github.com/wizdroid/wizdroid-character)** β 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. |
|
|
| --- |
|
|
| ## π How to Use |
|
|
| 1. **Load a workflow**: Drag & drop any `.json` file directly onto your ComfyUI canvas, or use **Load** β 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 2β3 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). |
|
|
| --- |
|
|
| ## π‘ 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`. |
|
|
| --- |
|
|
| ## οΏ½ Full Guide: Mastering Consistent Characters |
|
|
| For the **complete step-by-step tutorial** (problem β solution β exact pipeline using these workflows + PROMPTS.md + the WizdroidLoRADataset node β pro tips + visual examples), see the dedicated CivitAI article: |
|
|
| **β [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 40β100+ 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. |
|
|
| --- |
|
|
| ## οΏ½π License |
|
|
| MIT License β feel free to use, modify, and share these workflows. |
|
|
| --- |
|
|
| ## π€ Contributing |
|
|
| Pull requests and workflow improvements are welcome! If you create amazing results with these, tag **@wizdroid** β I'd love to see what you build. |
|
|
| --- |
|
|
| **Made with β€οΈ for the ComfyUI community by wizdroid** |
|
|
| *Consistent characters start with great references and great tools.* |
|
|
| --- |
|
|
| **Additional Resources** |
| - [PROMPTS.md β 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. |
|
|
