A clean, ready-to-use ComfyUI workflow for FLUX.2 Klein 9B supporting 0, 1, or 2 reference images in a single graph. Powered by the Q4_K_M GGUF quantization with the text encoder offloaded to CPU, it runs smoothly on 8 GB VRAM in just 4 sampling steps (CFG 1) — no LoRA training required.
---
## 🎯 What It Can Do (6 Modes in One Graph)
Switch between modes instantly by toggling the two ReferenceLatent nodes REF 1 and REF 2) with Ctrl+B (or Right-Click → Bypass):
1. Text-to-Image (0 references): Leave both REF nodes bypassed (default) for standard prompt generation.
2. Face Reference (REF 1): Keep the exact identity of a person across any new scene, outfit, or pose.
3. Face + Body (REF 1 + REF 2): Combine the face from Image 1 with the body build and proportions from Image 2.
4. Face + Location (REF 1 + REF 2): Place your character into the environment from Image 2 while preserving its layout and lighting.
5. Face Swap (REF 1 + REF 2): Replace only the face in a ready-made shot (Image 2), keeping pose, clothes, and background intact.
6. Full-Person Swap (REF 1 + REF 2): Replace the entire person in a ready-made shot (Image 2) while keeping the camera angle, lighting, and scene.
> 💡 Built-in Prompt Templates: Copy-paste prompt templates for all 6 modes are included right inside the workflow inside the "Prompt templates" note.
---
## 👥 Who It's For
* 8 GB VRAM GPU owners (RTX 3060 Ti, 3070, 4060, etc.) looking to run FLUX.2 Klein 9B locally without OOM errors.
* Creators & Storytellers who need consistent characters for Shorts, Reels, TikTok, or comics without training a LoRA.
* Beginners: Every group is color-coded, every node is clearly labeled, and step-by-step instructions are built directly into the canvas.
---
## 📥 Required Files & Setup
| Model File | Download Source | Target Folder |
| :--- | :--- | :--- |
| flux-2-klein-9b-Q4_K_M.gguf | [Unsloth HuggingFace](https://huggingface.co/unsloth/FLUX.2-klein-9B-GGUF/blob/main/flux-2-klein-9b-Q4_K_M.gguf) | ComfyUI/models/unet/ [source: 1, 2, 1.3.1] |
| qwen_3_8b_fp8mixed.safetensors | [HuggingFace](https://huggingface.co) (or ComfyUI Manager) | ComfyUI/models/text_encoders/ (or clip/) [source: 1, 2, 1.4.9] |
| flux2-vae.safetensors | [Comfy-Org HuggingFace](https://huggingface.co/Comfy-Org/Flux-2-Dev/blob/main/split_files/vae/flux2-vae.safetensors) | ComfyUI/models/vae/ [source: 1, 2, 1.5.1] |
* Required Custom Node: [ComfyUI-GGUF](https://github.com) (install via ComfyUI Manager).
* ComfyUI Version: Make sure your ComfyUI is updated so built-in FLUX.2 nodes Flux2Scheduler, EmptyFlux2LatentImage, ReferenceLatent) are available.
* **System RAM:** 16 GB+ RAM is recommended since the Qwen 3 8B text encoder runs on CPU.
---
## 🚀 How to Use
1. Select Models: Pick the 3 downloaded model files in the blue "Models ⚙" group on the left.
2. Choose Your Mode:
For *Text-to-Image**, leave both REF 1 and REF 2 nodes bypassed (purple).
For *Reference modes**, load your images into REF 1 (face) and/or REF 2 (body/location/shot), click the corresponding REF node in the Prompt & mode group, and press Ctrl+B (Right-Click → Bypass) to enable it.
3. Write Your Prompt: Refer to the inputs as "image 1" (REF 1) and "image 2" (REF 2). Grab a ready-made template from the "Prompt templates" note.
4. Set Resolution: Adjust Width and Height on the right (default is 720×1248 vertical 9:16; use 1248×720 for landscape or 1024×1024 for square).
5. Generate: Click Run.
---
## 💡 Pro Tips
* Best Likeness: Use a clear, well-lit, front-facing portrait for REF 1.
* Shirtless Body References: If your REF 2 body reference is shirtless, always explicitly describe the clothing in your prompt so the model doesn't copy the bare torso.
* Running Out of Memory? Lower the megapixels value in the "Resize REF 2" node from 1 to 0.5.
* Ready-Made Shot Swaps: Match the output Width and Height to the exact dimensions of your REF 2 image.
* Fine-Tuning: If the composition is close but needs a slight tweak, keep the prompt and try changing the seed first.
---
Please use only images you have the rights to, and respect the underlying model licenses.
If this workflow saved you time or VRAM, please leave a ❤️ Like and post your generations in the gallery below — it helps more creators discover it!
Description
Initial release: 6-in-1 reference workflow for 8GB VRAM (4 steps, CFG 1).




