Table of Contents
ComfyUI VRAM Overview
ComfyUI is the leading node-based interface for image generation workflows. Unlike single-model inference, ComfyUI workflows often load multiple models simultaneously: a base diffusion model, ControlNet adapters, LoRA weights, upscalers, and post-processing models. This means total VRAM usage can be significantly higher than running the base model alone. Planning your dedicated GPU server configuration around your most complex workflow is essential.
Base Model VRAM in ComfyUI
| Model | FP16 VRAM | FP8 VRAM | NF4 VRAM |
|---|---|---|---|
| SD 1.5 | ~2.5 GB | N/A | N/A |
| SDXL Base | ~6.5 GB | ~3.5 GB | ~2 GB |
| SDXL Base + Refiner | ~12 GB | ~7 GB | N/A |
| Flux.1 Dev | ~18-20 GB | ~13-15 GB | ~8-10 GB |
| Flux.1 Schnell | ~18-20 GB | ~13-15 GB | ~8-10 GB |
ComfyUI manages model loading dynamically: it loads models into VRAM when a node needs them and can offload to system RAM when not in use. However, peak VRAM occurs when the diffusion model, VAE, and text encoders are all active during generation. For base model VRAM details, see our Stable Diffusion VRAM and Flux.1 VRAM guides.
Extension and Node VRAM Overhead
| Extension | Additional VRAM | Notes |
|---|---|---|
| ControlNet (single) | +1-3 GB | Depends on ControlNet model size |
| IP-Adapter | +2-4 GB | Plus CLIP vision model |
| LoRA (single) | +0.1-0.5 GB | Merged into base model weights |
| Multiple LoRAs (3-5) | +0.3-1.5 GB | Cumulative per LoRA |
| Upscale model (4x) | +0.3-1 GB | Loaded during upscale node |
| Face restoration | +0.5-1 GB | CodeFormer or GFPGAN |
ControlNet is the largest VRAM consumer among common extensions. A single SDXL ControlNet adapter adds 1.5-3 GB. Stacking two ControlNets with an IP-Adapter can add 5-8 GB on top of the base model.
Common Workflow VRAM Profiles
| Workflow | Total VRAM (FP16) | Minimum GPU |
|---|---|---|
| SD 1.5 + LoRA + upscale | ~4 GB | RTX 3050 |
| SDXL + ControlNet + LoRA | ~10-12 GB | RTX 4060 Ti |
| SDXL + 2x ControlNet + IP-Adapter | ~14-18 GB | RTX 3090 |
| Flux.1 Dev FP16 + ControlNet | ~22-24 GB | RTX 3090 |
| Flux.1 Dev FP8 + ControlNet | ~16-18 GB | RTX 4060 Ti / RTX 3090 |
| Flux.1 NF4 + LoRA | ~10-12 GB | RTX 4060 Ti |
Complex Flux workflows with ControlNet at FP16 require 22-24 GB, making the RTX 3090 the minimum practical GPU. FP8 quantisation brings Flux + ControlNet workflows within reach of 16 GB cards.
GPU Recommendations
| GPU | VRAM | Best Workflow Tier |
|---|---|---|
| RTX 3050 | 6 GB | SD 1.5 basic workflows |
| RTX 4060 | 8 GB | SD 1.5 complex, SDXL basic |
| RTX 4060 Ti | 16 GB | SDXL + extensions, Flux NF4/FP8 |
| RTX 3090 | 24 GB | Flux FP16, complex multi-model |
Memory Optimisation Tips
- Use FP8 for Flux in ComfyUI to halve base model VRAM with minimal quality loss.
- Enable model offloading in ComfyUI settings to unload models not currently in use to system RAM.
- Use VAE tiling for high-resolution outputs to prevent VRAM spikes during VAE decoding.
- Load ControlNets selectively. Only keep the ControlNets your current workflow needs in VRAM.
- Use SDXL Turbo for preview workflows. It uses the same VRAM as SDXL but generates in 1-4 steps. See our SDXL Turbo VRAM guide.
Compare GPU options with the GPU comparisons tool. Estimate costs with the cost calculator. Browse all image generation guides in the model guides section.
Run ComfyUI on Dedicated GPUs
Deploy ComfyUI workflows on dedicated GPU servers from 6 GB to 24 GB VRAM. Full root access and no generation limits.
Browse GPU Servers
