| # BF16 / INT8 comparison |
|
|
| Additional adult Chinese woman portrait pairs are reported separately: [supplement](young_woman/report.md). |
|
|
| 14 paired outputs on NVIDIA GeForce RTX 5090; 1024×1024, 40 steps, offload=model, CFG=1, KV cache enabled. |
|
|
| Warmup excluded. Pixel metrics measure drift, not semantic quality. The same BF16 portrait is used as input for every editing pair. The suite is small and does not establish a general quality ranking. |
|
|
| | Case | Seed | BF16 s | INT8 s | BF16 peak GiB | INT8 peak GiB | RGB MAE (white) | |
| |---|---:|---:|---:|---:|---:|---:| |
| | portrait | 42 | 29.77 | 19.65 | 16.40 | 22.86 | 0.0134 | |
| | portrait | 123 | 26.55 | 19.61 | 16.40 | 22.86 | 0.0163 | |
| | english_text | 42 | 25.95 | 19.42 | 16.41 | 22.86 | 0.0263 | |
| | english_text | 123 | 25.88 | 19.40 | 16.41 | 22.86 | 0.0118 | |
| | chinese_text | 42 | 25.47 | 19.54 | 16.41 | 22.86 | 0.0597 | |
| | chinese_text | 123 | 25.85 | 19.86 | 16.41 | 22.86 | 0.0246 | |
| | composition | 42 | 25.38 | 19.13 | 16.41 | 22.86 | 0.0227 | |
| | composition | 123 | 25.35 | 19.30 | 16.41 | 22.86 | 0.0196 | |
| | texture | 42 | 25.85 | 19.22 | 16.40 | 22.86 | 0.0698 | |
| | texture | 123 | 25.76 | 19.57 | 16.40 | 22.86 | 0.0168 | |
| | rgba | 42 | 26.04 | 19.39 | 16.40 | 22.86 | 0.0183 | |
| | rgba | 123 | 25.07 | 19.29 | 16.40 | 22.86 | 0.0465 | |
| | edit | 42 | 31.16 | 25.40 | 19.08 | 23.08 | 0.0097 | |
| | edit | 123 | 31.03 | 25.24 | 19.08 | 23.08 | 0.0039 | |
|
|
| ## Summary |
|
|
| Mean latency: BF16 26.79s; INT8 20.29s. |
|
|
| [Interactive-sized side-by-side gallery](comparison.html). Raw data: comparison.csv, bf16/records.jsonl, int8/records.jsonl. Environment records include package versions and loading overhead. |
|
|
| [Qualitative observations](qualitative.md). |
|
|
| ## Interpretation limits |
|
|
| - No CLIP, OCR, human preference, FID or benchmark leaderboard score is claimed. |
| - Peak CUDA allocated/reserved memory excludes other processes and display usage. |
| - Measured call latency includes transfers; disk writing and model loading are excluded. |
| - Paired images may diverge with quantization even when both remain plausible. |
| - Editing and transparency should be inspected in the gallery, including preserved details. |
|
|
