AI Image Model VRAM Calculator (Stable Diffusion & Flux)

Every figure is computed in bytes from published component sizes, then shown in GiB (1 GiB = 2³⁰ bytes). Diffusion models have a different memory profile from text LLMs — there is no autoregressive KV cache. Peak memory is the model weights plus an activation/latent peak that grows with resolution, plus a fixed CUDA-context floor.

1. Model weights

Weight memory for each component is params × bytes-per-param, summed over the backbone (UNet for SD 1.5/SDXL, MMDiT for SD 3.5/Flux), the text encoder(s), and the VAE. Bytes per parameter: fp32 = 4, fp16/bf16 = 2, fp8/int8 = 1, and NF4 = 0.5. So SDXL's 2.567 B UNet is ~4.8 GiB at fp16 but ~1.2 GiB at NF4. (Diffusers quantization guide.)

2. What quantization touches

The chosen precision applies to the backbone. Text encoders are kept at fp8 minimum — a T5-XXL rarely runs at 4-bit — and the VAE is kept at fp16 because quantizing it to 4-bit visibly degrades output. That is why quantizing a Flux model to NF4 shrinks the 11.9 B transformer but leaves the VAE and (fp8) text encoders roughly fixed.

3. Activation / latent peak

The activation term is k × (N / 16384) × batch, where N = (H/8) × (W/8) is the latent token count (the VAE downsamples 8×) and kis a per-family coefficient calibrated to the model cards' reference VRAM figures (SD 1.5 ≈ 1.0, SDXL ≈ 2.5, SD 3.5 and Flux ≈ 3.0 GiB at the 1024² reference of N = 16384). This is the term that carries the ± margin, and it is why doubling resolution — quadrupling N — is such a common out-of-memory cause.

4. Optimizations & offload

Attention slicing computes attention in chunks (≈ ½ the activation peak); VAE tiling decodes in tiles to cap the decode spike (a further ≈ ½). Text-encoder CPU offload subtracts the encoder weights from GPU peak by running them on the CPU before denoising. Sequential CPU offload streams one component at a time, so peak VRAM falls to roughly the largest single component — much lower memory, much slower. A flat ~0.8 GiB covers the CUDA context. (Diffusers “Reduce memory usage”.)

5. GPU fit & cross-check

The total is compared against each of the 6/8/12/16/24 GB tiers. If a card is too small, the tool escalates a fixed optimization ladder (slicing → tiling → offload → quantization → sequential offload) and reports the minimum set that fits. As an independent check, each preset is run at a widely-benchmarked reference config and the estimate is confirmed to land inside the model card's published VRAM band — the same idea as cross-checking a tax figure against the regulator's own formula.

• HuggingFace Diffusers — Reduce memory usage (attention slicing, VAE tiling, CPU offload)
• HuggingFace Diffusers — Quantization (bitsandbytes NF4/int8, fp8) overview
• Black Forest Labs — Flux.1 [dev] model card (transformer, T5-XXL, VAE sizes)
• Stability AI — Stable Diffusion XL base 1.0 model card
• Stability AI — Stable Diffusion 3.5 Large model card

Component parameter counts, precision byte-sizes, and optimization effects were last verified against the sources above on 2026-07-09. This tool estimates inference (image generation), not training — LoRA and full fine-tuning need more memory for gradients and optimizer states. Figures target NVIDIA/CUDA VRAM; AMD (ROCm) and Apple silicon (MPS) differ.