Home → Image generation
VRAM for AI image generation
Image workloads are more forgiving than LLMs — models fit in far less memory — but the step from SD 1.5 to SDXL/Flux roughly triples the practical floor.
Tiers by model family
| Workload | Practical minimum | Comfortable |
|---|---|---|
| SD 1.5 | 4GB | 6–8GB |
| SDXL | 12GB | 16–24GB |
| Flux · SD 3.5 | 16GB baseline | — |
| LoRA training | 16GB | 24GB |
| Modern multi-step workflows | 12–16GB | — |
Source: Hivenet community guide (June 2026), tier definitions as published.
What the tiers mean in practice
- 4–8GB (older cards). SD 1.5 only. Fine for learning the workflow; the model ceiling is the limit, not your patience.
- 12GB. The entry point for current-generation image work — SDXL runs, Flux is tight, and heavy multi-model workflows will swap or OOM.
- 16–24GB. The current comfort zone: SDXL with headroom, Flux and SD 3.5 at their baseline, and LoRA training of your own adapters. This is where 24GB consumer cards (3090/4090 class) earn their keep.
If your card is under-spec
Quantized/offloaded variants can squeeze bigger models onto smaller cards at a speed cost, and renting a larger card by the hour is the other standard lever — image batch jobs are short enough that hourly rental often beats upgrading hardware for occasional use.