What You'll Build
Generate images locally with Flux.1 Dev on an RTX 5060 Ti. The model runs on 16GB VRAM but has a known VRAM reservation bug in ComfyUI on Windows that requires a workaround.
Hardware data: RTX 5060 Ti (16GB VRAM) · Known issue: 4777 MB reserved after first generation · See benchmark data
⚠️ Known issue: On Windows + ComfyUI, Flux.1 Dev leaves ~4.7GB of VRAM reserved after the first generation, which can cause a hang on the second run. The workaround is to restart the ComfyUI process between sessions or use a memory management node.
Requirements
| Component | Minimum | Tested |
|---|---|---|
| GPU | 16GB VRAM | RTX 5060 Ti (16GB) |
| RAM | 16GB | — |
| Storage | 35GB | 35GB (model weights) |
Installation
1. Install ComfyUI
git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI
pip install -r requirements.txt
2. Download Flux.1 Dev Weights
# Using huggingface-cli (requires HuggingFace account + license agreement)
huggingface-cli login
huggingface-cli download black-forest-labs/FLUX.1-dev \
--local-dir models/unet/
Place model files:
flux1-dev.safetensors→ComfyUI/models/unet/ae.safetensors→ComfyUI/models/vae/- Text encoders →
ComfyUI/models/clip/
Running
Start ComfyUI:
python main.py --listen
Navigate to http://localhost:8188 and load a Flux.1 Dev workflow.
Known Issue: VRAM Reservation on Windows
The RTX 5060 Ti benchmark shows a persistent VRAM reservation bug on Windows + ComfyUI:
Symptom: After the first generation, ~4777 MB of VRAM stays reserved. The second generation may hang.
Workaround options:
- Restart ComfyUI between sessions
- Use
--lowvramflag:python main.py --listen --lowvram - Restart the Python process to fully clear VRAM
This appears to be a ComfyUI memory management issue, not a hardware limitation. The model itself runs correctly on the first generation.
Performance
Community benchmark data for RTX 5060 Ti + Flux.1 Dev is limited. If you benchmark your setup, consider submitting your results.
For reference, the RTX 4090 runs Flux.1 Dev at approximately 2.1 it/s with FP16. The RTX 5060 Ti should be slower due to lower VRAM bandwidth, though exact numbers are not yet confirmed. See compare RTX 5060 Ti vs RTX 4090 for hardware specs.
Troubleshooting
- OOM on first run: Reduce resolution or enable
--lowvramflag - VRAM leak (4.7GB reserved): Restart ComfyUI or use memory management node
- Slow generation: Ensure you're using CUDA (not CPU): verify
torch.cuda.is_available()returnsTrue - CFG = 1 negative prompts not working: Install ComfyUI-ppm (see Installation step 3)