Image21-INT4

Built with Qwen. An independent 4-bit derivative of Qwen-Image-2.1, prepared by ixim. This is a community conversion, not an official Qwen release.

The stored weights are SDNQ UINT4 (unsigned 4-bit integers) plus a rank-32 SVD residual. "INT4" in the repository name means this 4-bit integer checkpoint. It is not bitsandbytes NF4, GPTQ, or a pure integer pipeline. Quantized matmul is left off, so the CUDA and Apple Silicon paths dequantize with ordinary PyTorch operators.

What runs where

• CUDA, about 8GB: group offload keeps one transformer block, or one text-encoder linear layer, on the GPU. VAE encode and decode run on CPU so a 1024 activation does not sit beside the denoiser cache. A 1024×1024, 40-step generation completed under a 7.2 GiB PyTorch allocator cap; see the measurement below. That cap was applied on a 32GB RTX 5090, so it is a memory-ceiling test rather than a log from a physical 8GB card. The installed PyTorch build still has to support the GPU.
• Larger CUDA GPUs: the loader uses model CPU offload, one pipeline component at a time. The paired comparison below used this path.
• Apple Silicon: the same files load on MPS. This Windows host has no Apple GPU, so there is no MPS timing or memory measurement here. Unified memory has to hold the packed checkpoint plus activations; an 8GB Mac is not the target. 24GB of unified memory is the comfortable size for a resident load. Set IMAGE21_DTYPE=float16 if a particular MPS operator rejects bfloat16.

Quantization

Base revision: b3179ad355be050328e483a9dfdd9e60cd62adfa. Eligible linear layers in the transformer and the Qwen3-VL text encoder are UINT4. Sensitive projections, normalization, embeddings, the vision tower, the output head, and the VAE stay in floating point. Exact module names are in the component quantization JSON files. No calibration set and no fine-tuning were used.

Installation

pip install torch torchvision
pip install -r requirements.txt

On NVIDIA machines, install a CUDA wheel of PyTorch 2.10 that matches the GPU before requirements.txt. The conversion was made with torch 2.10.0+cu128.

import torch
import sys
from huggingface_hub import snapshot_download

model_dir = snapshot_download("ixim/Image21-INT4")
sys.path.insert(0, model_dir)
from scripts.runtime import load_pipeline
pipe = load_pipeline(model_dir, local_files_only=True)
image = pipe(
    prompt='A neon sign reading "CREATE WITH LIGHT", rainy night',
    width=1024, height=1024, num_inference_steps=40,
    true_cfg_scale=1.0, use_kv_cache=True,
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]
image.save("output.png")

load_pipeline selects CUDA, MPS, or CPU. On an 8GB-class CUDA board it selects group offload by itself. Do not re-quantize these files at load time.

For an edit, keep the edit seed different from the seed that created the source image. The published edit pairs use source seed 42 and edit seeds 1000042 and 1000123. Save PNG when you need the alpha channel.

Informal evaluation

Measured comparison

14 pairs, 1024×1024, 40 steps, CFG 1, KV cache on, model CPU offload, one excluded warmup. Host: NVIDIA GeForce RTX 5090.

	BF16	INT4
Weight files (decimal GB)	33.116	13.441
Mean call latency (s)	26.98	29.40
Maximum allocated CUDA memory (GiB)	19.08	9.99

Case	Seed	BF16 s	INT4 s	BF16 GiB	INT4 GiB	RGB MAE
chinese_text	42	25.55	28.67	16.41	7.71	0.0791
chinese_text	123	25.61	28.79	16.41	7.71	0.0819
composition	42	25.21	28.72	16.41	7.71	0.0297
composition	123	25.51	28.67	16.41	7.71	0.0594
edit	1000042	38.67	33.81	19.08	9.99	0.0112
edit	1000123	31.10	33.51	19.08	9.99	0.0107
english_text	42	25.40	28.35	16.41	7.71	0.0839
english_text	123	25.23	28.67	16.41	7.71	0.0547
portrait	42	28.05	29.41	16.40	7.71	0.0528
portrait	123	26.06	28.21	16.40	7.71	0.0639
rgba	42	25.02	28.55	16.40	7.70	0.0444
rgba	123	25.76	28.54	16.40	7.70	0.0760
texture	42	25.40	28.98	16.40	7.70	0.0775
texture	123	25.13	28.76	16.40	7.70	0.0489

Package versions

Package	Version
torch	2.10.0+cu128
diffusers	0.41.0.dev0
transformers	5.17.0
accelerate	1.15.0
sdnq	0.2.7
safetensors	0.8.0
huggingface-hub	1.32.0
tokenizers	0.23.2
numpy	2.2.6
pillow	12.2.0
psutil	7.0.0

Visual inspection

Visual inspection

These notes describe images I looked at. They are not a score, and they do not cover every seed. Pixel drift for all 14 pairs is in comparison.csv.

portrait, seed 42: Both are waist-up photographs of an older woman in a blue wool sweater beside a window. The face, the hair, and how much of the window is in frame differ. Skin and knit texture stay photographic in both.

english_text, seed 42: Both render the headline CREATE WITH LIGHT and the subtitle September 2026. The orange lamp, books, and notebook are arranged differently. I did not see an extra headline.

chinese_text, seed 42: Both render 慢下来，喝杯咖啡 and 小店今日营业 above a latte. Both also add a footer of unwanted, partly illegible text, so neither follows “不添加其他文字” completely.

composition, seed 42: Both show exactly three cups, red then blue then yellow, and one green apple in front of the blue cup. Cup shape and spacing differ. The INT4 cups have handles; the BF16 cups do not.

texture, seed 42: Both are plausible kingfishers on a mossy branch over water. The INT4 bird faces left and the BF16 bird faces right. Feather and bark detail are visible in both.

rgba, seed 42: Both are full-body green cartoon dragons with orange wings, saved as RGBA. Measured alpha runs from 0 to 255. Pixels with alpha at or below 5 are 63.5% of the BF16 image and 73.0% of the INT4 image. The BF16 dragon has a white sticker outline.

edit, seed 1000042: The input is the BF16 portrait from seed 42, and the edit seed is not 42. Both change the sweater from blue to red and keep the same face, window light, and pose. This pair is not visibly oversharpened. That does not show that every edit will preserve identity.

The 8GB-cap portrait (group offload, CPU VAE encode and decode) is a separate run from this paired set. On that image the woman, blue sweater, and window are all present. The same cap also completed both edit seeds, 1000042 and 1000123, at 1024 for 40 steps. Allocated peaks were 3.48 GiB and reserved peaks were 4.08 and 4.24 GiB.

目视记录

下面只记录看过的图，不是评分，也没有逐张覆盖第二个种子。14 对的像素差在 comparison.csv。

portrait，种子 42：两种精度都是窗边、蓝毛衣、半身的老年女性照片。脸、头发和窗户入画的多少不同。皮肤和针织纹理都还像照片。

english_text，种子 42：两种精度都写出了 CREATE WITH LIGHT 和 September 2026。橙色台灯、书和笔记本的摆放不同。没有看到多出来的主标题。

chinese_text，种子 42：两种精度都写出了“慢下来，喝杯咖啡”和“小店今日营业”，中间是拿铁。两种精度的页脚都出现了多余且部分无法辨认的文字，都没有完全遵守“不添加其他文字”。

composition，种子 42：两种精度都是左红、中蓝、右黄三个杯子，蓝杯前有一个青苹果。杯子形状和间距不同。INT4 的杯子有把手，BF16 没有。

texture，种子 42：两种精度都是水边苔枝上的翠鸟。INT4 的鸟朝左，BF16 的鸟朝右。羽毛和树皮细节都还在。

rgba，种子 42：两种精度都是全身、绿鳞、橙翼的卡通龙，文件为 RGBA。alpha 范围是 0 到 255。alpha 小于等于 5 的像素，BF16 占 63.5%，INT4 占 73.0%。BF16 的龙有一圈白色贴纸描边。

edit，种子 1000042：输入是 BF16 种子 42 的肖像，编辑种子不是 42。两种精度都把蓝毛衣改成了红毛衣，脸、窗光和姿态还在。这一对没有明显过锐。这不能说明每次编辑都会保住身份。

8GB 分配上限下的肖像是另一次运行（分组卸载，VAE 的编码和解码都在 CPU）。那张图里仍有人物、蓝毛衣和窗户。同一次上限设置也完成了两个编辑种子 1000042 和 1000123，1024、40 步。已分配峰值都是 3.48 GiB，保留峰值分别是 4.08 GiB 和 4.24 GiB。

8GB CUDA allocator proof

The saved INT4 checkpoint generated portrait seed 42 at 1024×1024 for 40 steps while PyTorch's caching allocator was limited to 7.2 GiB on this NVIDIA GeForce RTX 5090. Offload mode: group: one transformer block or one text-encoder leaf on GPU, VAE encode and decode on CPU. Allocated peak 1.22 GiB, reserved peak 2.37 GiB, call 84.32s after an excluded warmup. This is a ceiling test on a larger board, not a run on a physical 8GB card.

Paired images

chinese_text / seed 42

设计一张现代咖啡馆海报，米白背景，中央一杯拉花拿铁。顶部准确写上中文大标题“慢下来，喝杯咖啡”，下方准确写上“小店今日营业”。文字清晰，留白充分，暖色摄影，不添加其他文字。

BF16	INT4

chinese_text / seed 123

设计一张现代咖啡馆海报，米白背景，中央一杯拉花拿铁。顶部准确写上中文大标题“慢下来，喝杯咖啡”，下方准确写上“小店今日营业”。文字清晰，留白充分，暖色摄影，不添加其他文字。

BF16	INT4

composition / seed 42

A studio photograph on a light gray tabletop: exactly three ceramic cups in a row, a red cup on the left, a blue cup in the middle, and a yellow cup on the right. A single green apple sits in front of the blue cup. Soft shadows, no text.

BF16	INT4

composition / seed 123

A studio photograph on a light gray tabletop: exactly three ceramic cups in a row, a red cup on the left, a blue cup in the middle, and a yellow cup on the right. A single green apple sits in front of the blue cup. Soft shadows, no text.

BF16	INT4

edit / seed 1000042

Change only the blue sweater to a red sweater. Preserve the same person, face, pose, lighting and background.

BF16	INT4

edit / seed 1000123

Change only the blue sweater to a red sweater. Preserve the same person, face, pose, lighting and background.

BF16	INT4

english_text / seed 42

A clean editorial poster with a deep blue background. Large exact headline at the top: "CREATE WITH LIGHT". Smaller exact subtitle: "September 2026". A realistic orange desk lamp occupies the lower half. Elegant balanced typography, no additional text.

BF16	INT4

english_text / seed 123

A clean editorial poster with a deep blue background. Large exact headline at the top: "CREATE WITH LIGHT". Smaller exact subtitle: "September 2026". A realistic orange desk lamp occupies the lower half. Elegant balanced typography, no additional text.

BF16	INT4

portrait / seed 42

A natural documentary portrait of an elderly woman with silver hair and freckles, wearing a blue wool sweater, standing beside a window in soft morning light. Realistic skin texture, gentle expression, waist-up composition.

BF16	INT4

portrait / seed 123

A natural documentary portrait of an elderly woman with silver hair and freckles, wearing a blue wool sweater, standing beside a window in soft morning light. Realistic skin texture, gentle expression, waist-up composition.

BF16	INT4

rgba / seed 42

This is an RGBA image with transparency. A cute cartoon dragon sticker, full body, green scales and small orange wings. The image has alpha channel and the background is transparent.

BF16	INT4

rgba / seed 123

This is an RGBA image with transparency. A cute cartoon dragon sticker, full body, green scales and small orange wings. The image has alpha channel and the background is transparent.

BF16	INT4

texture / seed 42

Macro photography of a small kingfisher perched on a mossy branch beside clear water, detailed blue feathers, droplets, natural sunlight, softly blurred forest background, realistic textures.

BF16	INT4

texture / seed 123

Macro photography of a small kingfisher perched on a mossy branch beside clear water, detailed blue feathers, droplets, natural sunlight, softly blurred forest background, realistic textures.

BF16	INT4

The Qwen Research License keeps this derivative limited to non-commercial research and evaluation. License, Notice, and CHANGES.md are in the repository.