Informal release evaluation: BF16 / INT8 comparison

This community evaluation is provided for reference only and does not represent any official evaluation.

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	28.59	19.62	16.40	9.49	0.0134
portrait	123	26.07	19.48	16.40	9.49	0.0163
english_text	42	25.42	19.54	16.41	9.50	0.0263
english_text	123	25.66	19.68	16.41	9.50	0.0118
chinese_text	42	25.62	19.61	16.41	9.50	0.0597
chinese_text	123	25.46	19.61	16.41	9.50	0.0246
composition	42	25.39	19.32	16.41	9.49	0.0227
composition	123	25.48	19.15	16.41	9.49	0.0196
texture	42	25.32	19.42	16.40	9.49	0.0698
texture	123	25.18	19.16	16.40	9.49	0.0168
rgba	42	25.22	19.22	16.40	9.48	0.0183
rgba	123	25.54	19.22	16.40	9.48	0.0465
edit	42	30.79	25.07	19.08	12.09	0.0097
edit	123	30.46	24.81	19.08	12.09	0.0039

Summary

Mean latency: BF16 26.44s; INT8 20.21s.

Interactive-sized side-by-side gallery. Raw data: comparison.csv, bf16/records.jsonl, int8/records.jsonl. Environment records include package versions and loading overhead.

Qualitative observations.

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.