Private SDXL: Judge Your Local Image Rig by Measured Benchmarks
In short: A privacy-local SDXL rig is a Stable Diffusion XL image-generation setup that runs entirely on hardware you control, so prompts, reference images, and outputs never leave the machine — and it should be judged not by cloud throughput claims but by measured VRAM headroom, data-residue behavior, and log-retention policy.
A privacy-local SDXL rig is a Stable Diffusion XL image-generation setup that runs entirely on hardware you control, so prompts, reference images, and outputs never leave the machine — and it should be judged not by cloud throughput claims but by measured VRAM headroom, data-residue behavior, and log-retention policy.
최대 VRAM 상주(스냅샷) 84.8 GB
What did Hax measure on its own stack?#
Reference numbers Hax measured directly on its own infrastructure (measured, sourced).
| Dataset item | Measured value | Date | Source |
|---|---|---|---|
| 최대 VRAM 상주(스냅샷) | 84.8 GB | 2026-07-04 | bench_harness.probe_comfy_gpus (bc_comfy_gpus 실측) |
| 카드당 총 VRAM | 95.6 GB | 2026-07-04 | bench_harness.probe_comfy_gpus (bc_comfy_gpus 실측) |
| 최대 GPU 사용률 | 95 % | 2026-07-04 | bench_harness.probe_comfy_gpus (bc_comfy_gpus 실측) |
- 표본
- 3 measured metrics (Hax /data curated)
- 측정 환경
- bench_harness.probe_comfy_gpus (bc_comfy_gpus 실측)
- 수집일
- 2026-07-04
How can you reproduce these numbers?#
Follow the source column above and our open dataset at /data.
Max VRAM resident snapshot 84.8 GB
| Metric | Hax local | Cloud SDXL API |
|---|---|---|
| Prompt/image leaves device | No, stays local | Yes, uploaded (estimated) |
| Max VRAM resident (snapshot) | 84.8 GB (measured) | n/a |
| Total VRAM per card | 95.6 GB (measured) | not disclosed (estimated) |
| Min free VRAM (pool low) | 10.2 GB (measured) | n/a |
| GPU cards | 4 (measured) | shared, unknown (estimated) |
| Max GPU utilization | 95% (measured) | unknown (estimated) |
| Installed checkpoints / LoRA | 32 / 63 (measured) | fixed menu (estimated) |
| Throughput (images/min) | not measured / 측정대기 | 8-20 (estimated) |
| Log retention | operator-controlled | provider policy (estimated) |
Why data policy beats raw speed#
When the goal is that personal photos and prompts never travel, the honest benchmark is not images per minute — it is whether the pipeline can produce zero-residue runs. On a local SDXL rig, the request path is a loop back to the same box: no upload, no third-party queue, no retained copy on someone else's disk. That is the property a cloud API cannot structurally offer, because the image must reach their servers to be processed.
Hax's rig was probed on 2026-07-04. The GPU pool holds 4 cards (measured) with 95.6 GB total VRAM per card (measured). Under load the snapshot showed 84.8 GB resident (measured) and a pool low of 10.2 GB free (measured), with peak utilization at 95% (measured). That headroom is what lets SDXL, its refiner, and multiple ControlNets sit resident without swapping to cloud offload — the swap step is exactly where data would otherwise leave the device.
What is actually installed#
Model breadth decides whether you ever feel tempted to reach for a hosted service. The rig carries 32 checkpoints (measured), 63 LoRAs (measured), 44 samplers (measured), and 15 ControlNets (measured). For most SDXL tasks — portraits, product shots, pose transfer — that catalog removes the usual reason people upload to a cloud tool.
The number we do NOT have yet#
Throughput in images/min is not measured / 측정대기 for this rig. Public SDXL-on-consumer-GPU figures land roughly 8-20 images/min (estimated) at 1024px with 25-30 steps, but we will not attach a measured label to a number we did not run. Treat any images/min claim here as estimated until a probe records it.
How to judge one yourself#
Ask three questions in order. First, does any request byte leave the device? Second, what is the log-retention default, and can you set it to zero? Third, is there enough VRAM headroom to keep everything resident so no cloud-offload path activates. Only after those pass does speed matter. A rig that is fast but uploads your data fails the actual test for privacy-local work.
Note: figures labeled measured are from the 2026-07-04 probe; VRAM and utilization shift with concurrent jobs, and throughput remains unmeasured. See Hax data for the underlying probe records.
Related reading: 처음 SDXL, 실측 VRAM과 설치 실패 지점으로 판단하기, 개인정보 안 나가는 BGE-M3 다국어 검색 5분 퀵스타트
Full guide: 노트북에서 돌리는 AI 모델, 흔한 함정과 해결법
Responses
No responses yet. Be the first to respond.