Building for Image Models: FLUX and SDXL VRAM Budgets in BuildSmith
Image models play by different rules than chat models, and BuildSmith's image mode plans for it. The weights are small next to a 70B LLM: FLUX.1 dev is a 12B model, SDXL about 3.5B. The memory pressure comes from activations instead, which grow with resolution and batch size, plus the text encoders that ride along with the transformer.
So the planner budgets in two parts. The weights line uses the same parameters times bytes math as the chat flow. Then it adds an activation allowance scaled to the resolution you pick, and it warns in plain terms when a 12 GB card will fit the weights at INT8 but not the 1024 pixel batch you asked for. That warning, boring as it sounds, is the difference between a plan and a surprise.
The weight budgets the app starts from
| Model | Params | FP16 weights line | INT8 weights line | INT4 weights line | Watch out |
|---|---|---|---|---|---|
| FLUX.1 dev | 12B | 28.8 GB | 14.4 GB | 7.2 GB | Plus text encoders |
| SDXL | 3.5B | 8.4 GB | 4.2 GB | 2.1 GB | Activations grow with resolution |
| Llama 3.1 8B (for prompt work) | 8B | 19.2 GB | 9.6 GB | 4.8 GB | Often shares the card |
Weight memory at INT8
The practical output: for SDXL work, the app steers most people to 12 and 16 GB cards and spends the savings on storage and RAM. For FLUX at volume, it starts at 24 GB and treats anything less as a compromise to sign off on, not a default. Both flows end in the same store step as every BuildSmith plan.
Batch size is the quiet multiplier. One 1024 pixel image is a plan; a batch of eight for a contact sheet is a different memory line, and the planner asks which one you mean before it sizes the card. The text encoders get their own line for the same reason: FLUX carries two of them, and on small cards they are the difference between fitting cleanly and swapping to system memory mid generation, which is where the mysterious slowdowns in community threads usually come from. Budget them as residents, not guests, and the mystery goes away.
One last habit the app encourages: save the plan even when you do not buy yet. Image hardware ages in dog years, and a saved budget from three months ago, re-run against today's directory, shows exactly which line moved: the card, the drive, or your own ambition. Watching a plan get cheaper, or more honest, is its own kind of progress.
Try the underlying math in the VRAM Fit Checker, which includes both image models. BuildSmith is in development; waitlist on the homepage.