I've been reading threads here for months, trying to get a local setup working on my older card (6GB VRAM). The advice is everywhere. Tried Automatic1111, but it OOMs on anything over 512px.
Finally got something stable. The key for me was:
- Using SDXL Turbo models, not SD 1.5. They seem to need less VRAM for decent output.
- Forcing `--medvram` and `--xformers` in the webui bat file.
- Switching to ComfyUI, which seems to manage memory better for workflows.
But I'm still hitting walls with control nets or upscaling. Is there a definitive guide now? What's the actual best practice for under 8GB?
I'm a data engineer at a mid-sized e-commerce company, and we run image generation for product mockups locally on a mix of dev machines, some of which have older 8GB cards. We've tested several setups for cost and stability.
Here are the specifics from our testing:
1. **VRAM management**: ComfyUI consistently uses 1-1.5GB less VRAM than Automatic1111 for an equivalent SDXL pipeline. It's the difference between a 6GB card running a 768px image and hitting an OOM error.
2. **Model choice impact**: You're right about SDXL Turbo. A 1.5-base model with a control net can spike to 7GB on a 512px image, while SDXL Turbo with a similar workflow holds around 5.5GB at 768px. The quality trade-off for the headroom is real.
3. **The upscaling wall**: This is the biggest limitation. Even with ComfyUI, adding a face detailer and a 2x upscaler to a workflow will push you past 8GB total memory. The only reliable workaround we found is splitting the workflow and using an ESRGAN upscaler on the CPU, which is about 4x slower.
4. **Tool maturity**: Automatic1111 has better one-click install and broader community guides, but ComfyUI's node-based system lets you isolate and cache heavy components like VAE decode, which saves 20-30% memory on subsequent generations in a batch.
My pick is ComfyUI with SDXL Lightning models, but only for generating base images under 1024px. For your use case, if you need control nets or high-res fixes regularly, you'll have to accept small batch sizes of one and a very incremental workflow. Tell us exactly what resolution you need for final output and if you're batching, and I can give a clearer config.