LAXMAYDAY/Qwen_image_edit_2511_int4

Qwen-Image-Edit-2511 Int4 Checkpoints

Overview

This repository provides Nunchaku int4 safetensors checkpoint files for Qwen-Image-Edit-2511. It packages multiple quantized variants in one place so you can choose between footprint and quality tradeoffs without pulling local-only reports or build metadata into the public release.

The repo now includes both standard 50-step checkpoints and selected 4-step / 8-step lightning checkpoints. This README is the current published index of files in the repo, including some newer lightning benchmark results that have not yet been folded back into the consolidated JSON reports.

Recommended Starting Points

• quality-r64: default quality-focused starting point for the standard 50-step family
• fast-r128: best FID among the consolidated 50-step variants
• mid-r128-lightning8steps: strongest currently benchmarked 8-step lightning balance
• quality-r128-b15-lightning8steps: best currently benchmarked 8-step lightning FID
• balanced-r32-lightning8steps: smaller-footprint 8-step lightning option
• Performance evaluations are not conducted separately for Lightning Variants, as their degradation is generally considered negligible and can be ignored

Standard 50-Step Variants

Variant	File	Size (GiB)	PSNR	SSIM	LPIPS	FID	Notes
balanced-r32	nunchaku_qwen_image_edit_2511_balanced_r32_int4.safetensors	10.73	19.1581	0.8058	0.2019	77.3707	smaller footprint
balanced-r32-i64	nunchaku_qwen_image_edit_2511_balanced_r32_i64_int4.safetensors	10.73	19.2460	0.8134	0.2059	76.1155	-
balanced-r32-b125	nunchaku_qwen_image_edit_2511_balanced_r32_b125_int4.safetensors	10.73	19.2856	0.8103	0.2005	78.7961	-
fast-r32	nunchaku_qwen_image_edit_2511_fast_r32_int4.safetensors	10.73	19.1576	0.8110	0.2058	77.9059	-
fast-r64	nunchaku_qwen_image_edit_2511_fast_r64_int4.safetensors	11.08	19.4508	0.8130	0.2024	71.5310	-
fast-r128	nunchaku_qwen_image_edit_2511_fast_r128_int4.safetensors	11.79	19.5990	0.8153	0.1897	70.1136	best FID
mid-r32	nunchaku_qwen_image_edit_2511_mid_r32_int4.safetensors	10.73	19.4421	0.8075	0.2006	82.5143	-
mid-r128	nunchaku_qwen_image_edit_2511_mid_r128_int4.safetensors	11.79	19.6128	0.8259	0.1828	76.2305	-
quality-r32	nunchaku_qwen_image_edit_2511_quality_r32_int4.safetensors	10.73	19.4316	0.8136	0.1954	82.3566	-
quality-r64	nunchaku_qwen_image_edit_2511_quality_r64_int4.safetensors	11.08	19.6273	0.8158	0.1856	73.9762	default quality
quality-r96	nunchaku_qwen_image_edit_2511_quality_r96_int4.safetensors	11.43	19.5872	0.8158	0.1935	74.5920	-
quality-r96-b175	nunchaku_qwen_image_edit_2511_quality_r96_b175_int4.safetensors	11.43	19.3046	0.8167	0.1939	78.3015	-
quality-r96-i128	nunchaku_qwen_image_edit_2511_quality_r96_i128_int4.safetensors	11.43	19.6934	0.8189	0.1915	74.2818	-
quality-r128-b15	nunchaku_qwen_image_edit_2511_quality_r128_b15_int4.safetensors	11.79	19.7177	0.8202	0.1857	80.3767	-

Lightning Variants

Variant	File	Size (GiB)	PSNR	SSIM	LPIPS	FID	Notes
balanced-r32-lightning4steps	nunchaku_qwen_image_edit_2511_balanced_r32_lightning4steps_q128o64m64_int4.safetensors	11.26	-	-	-	-	4-step lightning, metrics pending
balanced-r32-lightning8steps	nunchaku_qwen_image_edit_2511_balanced_r32_lightning8steps_q128o96m96_int4.safetensors	11.52	20.3516	0.8262	0.1955	70.8734	8-step lightning, smaller footprint
mid-r128-lightning4steps	nunchaku_qwen_image_edit_2511_mid_r128_lightning4steps_q128o128m128_int4.safetensors	11.79	-	-	-	-	4-step lightning, metrics pending
mid-r128-lightning8steps	nunchaku_qwen_image_edit_2511_mid_r128_lightning8steps_q128o128m128_int4.safetensors	11.79	21.1254	0.8349	0.1668	69.9338	8-step lightning, best PSNR / SSIM / LPIPS among benchmarked lightning rows
quality-r64-lightning4steps	nunchaku_qwen_image_edit_2511_quality_r64_lightning4steps_q128o64m64_int4.safetensors	11.26	-	-	-	-	4-step lightning, metrics pending
quality-r64-lightning8steps	nunchaku_qwen_image_edit_2511_quality_r64_lightning8steps_q128o64m64_int4.safetensors	11.26	-	-	-	-	8-step lightning, metrics pending
quality-r96-lightning4steps	nunchaku_qwen_image_edit_2511_quality_r96_lightning4steps_q128o96m112_int4.safetensors	11.63	-	-	-	-	4-step lightning, metrics pending
quality-r96-b175-lightning4steps	nunchaku_qwen_image_edit_2511_quality_r96_b175_lightning4steps_q128o96m112_int4.safetensors	11.63	-	-	-	-	4-step lightning, metrics pending
quality-r96-b175-lightning8steps	nunchaku_qwen_image_edit_2511_quality_r96_b175_lightning8steps_q160o128m128_int4.safetensors	11.87	-	-	-	-	8-step lightning q160, metrics pending
quality-r128-b15-lightning4steps	nunchaku_qwen_image_edit_2511_quality_r128_b15_lightning4steps_q128o128m128_int4.safetensors	11.79	-	-	-	-	4-step lightning, metrics pending
quality-r128-b15-lightning8steps	nunchaku_qwen_image_edit_2511_quality_r128_b15_lightning8steps_q128o128m128_int4.safetensors	11.79	20.6295	0.8311	0.1745	69.5480	8-step lightning, best FID among benchmarked lightning rows

Metrics

• Higher is better for PSNR and SSIM.
• Lower is better for LPIPS and FID.
• The standard 50-step rows above come from the consolidated local metrics report.
• The three benchmarked 8-step lightning rows above come from the same 24-prompt holdout setup against the same BF16 50-step reference set, but they have not yet been folded into the consolidated JSON report in this repo.
• Rows marked metrics pending are published checkpoints without a consolidated quality report in the repo yet.

Limitations

• Quantized checkpoints can shift quality, latency, and memory behavior depending on runtime, hardware, scheduler, LoRA choice, and inference settings.
• Lightning checkpoints operate in a different step regime than the standard 50-step variants, so treat the table as practical guidance rather than a single global leaderboard.