ixim/Image21-INT8

Image21-INT8

Built with Qwen. An independent bitsandbytes INT8 derivative of Qwen-Image-2.1, prepared by ixim. This is a community conversion, not an official Qwen release.

Quantization

Base revision: b3179ad355be050328e483a9dfdd9e60cd62adfa. Eligible linear layers in the generation transformer and the Qwen3-VL text encoder use bitsandbytes LLM.int8, with outlier threshold 6.0. This is mixed-precision inference: outlier handling uses floating point; the VAE, normalization, embeddings, vision encoder, output head, and excluded conditioning/projection layers remain floating point. See the two component quantization JSON files for exact coverage. No fine-tuning or calibration dataset was used. This is not a pure W8A8 pipeline, FP8 conversion, or GGUF checkpoint. Lower storage does not imply faster inference.

Installation and loading

Use Python 3.13 and a CUDA-capable PyTorch installation. The tested runtime uses PyTorch 2.10.0+cu128 on an RTX 5090. Install the accompanying requirements.txt; it pins Diffusers to the official commit used for this conversion.

pip install torch==2.10.0 torchvision==0.25.0 --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt

import torch
import sys
from huggingface_hub import snapshot_download

model_dir = snapshot_download("ixim/Image21-INT8")
sys.path.insert(0, model_dir)  # bundled, inspectable scripts/runtime.py
from scripts.runtime import load_int8_pipeline
pipe = load_int8_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")

The saved quantization configuration is detected automatically. Do not re-quantize these weights at load time. For editing, use the 2048px editing recipe below, setting width, height and output_resolution together and enabling VAE tiling. For transparency, use the upstream RGBA prompt format and save PNG files.

Use the bundled loader on the pinned runtime: it loads/offloads the two INT8 components sequentially and moves their auxiliary tensors during model offload. Calling only enable_model_cpu_offload() leaves INT8 CUDA tensors resident in this tested stack. The helper does not change the checkpoint or quantization.

Informal release evaluation

This community evaluation accompanies this model release and is provided for reference only. It does not represent any official evaluation by Qwen, Hugging Face, ModelScope, or their teams. Results reflect a small, locally selected sample on one hardware/software configuration; they are not a standardized ranking or a guarantee of quality, speed, or memory requirements.

14 paired outputs across 7 cases at 1024×1024 on an RTX 5090.

Observed limitations: both precisions over-sharpen the 1024px sweater edit and alter its contrast; both add unwanted footer text to the inspected Chinese poster. The 2048px snowy-forest background replacement also fails in both precisions. These failures are retained below. Pixel differences measure drift, not semantic quality; no official score, blinded preference result, OCR, CLIP or FID score is claimed. Multi-reference editing at scale, other GPUs/attention backends and ComfyUI compatibility were not evaluated.

All paired samples, prompts and observations

Left: BF16; right: saved and reloaded INT8. Same prompt and seed within each pair. The main seven-case observations primarily inspect seed 42; both supplementary seeds were inspected. This is not a blinded preference study or an OCR/CLIP/FID score.

Young Chinese adult woman portrait (supplement)

Exact measured prompt:

一张写实摄影风格的中国年轻成年女性肖像，人物明确为25岁成年人。黑色及肩直发，穿着米白色圆领针织衫，胸部以上构图，正面看向镜头，神情自然放松，轻微微笑。窗边柔和的自然日光，浅灰色简洁背景，真实自然的皮肤纹理和少量雀斑，淡妆，不过度磨皮。85mm人像摄影，浅景深，色彩自然，画面中没有文字。

Both seeds retain the broad framing, clothing and window lighting. Expression, facial contours, skin texture and individual hair strands change. No obvious large facial distortion was seen in these four outputs; this is not an identity-preservation score.

Seed 42

SHA256 prefixes: BF16 00716ad4721f / INT8 5ffeb0f35bf0.

Seed 123

SHA256 prefixes: BF16 a2f28b40c897 / INT8 350e800c857e.

Older adult portrait

Exact measured prompt:

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.

Both depict the requested older woman, blue sweater and window. Face shape, wrinkles, framing and fabric details shift.

Seed 42

SHA256 prefixes: BF16 01e37e178fb0 / INT8 e37187a2530d.

Seed 123

SHA256 prefixes: BF16 cff376e50c41 / INT8 09391672b7aa.

English lettering

Exact measured prompt:

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.

Both inspected outputs visibly render CREATE WITH LIGHT and September 2026. Layout and desk objects differ.

Seed 42

SHA256 prefixes: BF16 ba31b0c8a312 / INT8 98ea66acd842.

Seed 123

SHA256 prefixes: BF16 cf5e76dee9f0 / INT8 b55551e7c179.

Chinese lettering

Exact measured prompt:

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。

Both inspected outputs render the requested main title and subtitle, but also add unwanted, partly nonsensical footer text despite the no-extra-text instruction.

Seed 42

SHA256 prefixes: BF16 7082cb71d997 / INT8 dce2de1e3b8b.

Seed 123

SHA256 prefixes: BF16 eae74256d7ed / INT8 7ff8508a9454.

Counting and spatial arrangement

Exact measured prompt:

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.

Both show three cups in red/blue/yellow order and one green apple in front of the blue cup. Object sizes and spacing differ.

Seed 42

SHA256 prefixes: BF16 91e967685fbb / INT8 5704252c99e7.

Seed 123

SHA256 prefixes: BF16 bc11612f260a / INT8 c030c007a9c3.

Feather and branch texture

Exact measured prompt:

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.

Both depict a plausible kingfisher; head angle, branch orientation and feather texture differ. Pixel drift alone does not establish worse semantic quality.

Seed 42

SHA256 prefixes: BF16 40a0f73bb407 / INT8 37042d165914.

Seed 123

SHA256 prefixes: BF16 18c4403f388d / INT8 82fc88d5ab3c.

RGBA transparency

Exact measured prompt:

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.

Both have a real varying alpha channel and a white sticker outline. Sampled background alpha is 0–1/255 and the subject is opaque. Page background colors are not part of the PNG; click through to inspect the original RGBA file.

Seed 42

SHA256 prefixes: BF16 6f5dfea6ba02 / INT8 8624d495fe40.

Seed 123

SHA256 prefixes: BF16 6f1abbb1a64a / INT8 4871aab8f9d4.

Image editing: sweater recoloring

Exact measured prompt:

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

Both shift the sweater toward red/burgundy, but both also show conspicuous oversharpening and contrast/texture changes across the image. Faithful preservation of lighting and skin texture is not achieved. The cause is not identified or attributed specifically to INT8.

Both variants use this identical BF16 portrait input.

Seed 42

SHA256 prefixes: BF16 8713f6ef2ec0 / INT8 79725e243ccf.

Seed 123

SHA256 prefixes: BF16 7d00c5bb7a2b / INT8 d8023e34ba10.

2048px image editing comparisons

Four paired editing examples are shown with their shared input, generated by BF16 in this evaluation. Outputs use 2048×2048, 40 steps, seed 42, CFG=1, KV cache and model CPU offload. Both precisions enable default VAE tiling (256px minimum tiles, 192px strides). The 1024px input PNGs are resized inside the pipeline with output_resolution=2048. All images link to the full original PNG.

The sweater example at 1024px changes the requested color but fails to preserve image texture. At 2048px with tiling, both precisions better preserve the soft light and skin texture. RGB input, KV-cache-off and 1024px tiling checks did not visibly resolve that failure; an untiled 2048px call was stopped during memory-constrained decoding and is excluded. These few checks do not establish a root cause or universal editing reliability.

Older woman: blue sweater to red

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

Both outputs make the sweater red and retain soft lighting and broadly natural skin texture. Fine facial and fabric details still shift, and faint colored marks are visible in both; this is not lossless preservation.

Adult woman: cream sweater to navy blue

Prompt: Change only the cream sweater to a navy blue sweater. Preserve the same woman, face, hairstyle, pose, lighting and background.

Both outputs recolor the sweater to navy blue while broadly retaining the woman, pose, window light and background. Facial, hair and knitted-fabric details still shift slightly; this is not an identity-preservation score.

Product edit: ceramic cup to brushed metal

Prompt: Change only the middle blue ceramic cup into a brushed silver metal cup. Preserve its shape and position, the red and yellow cups, the green apple, the tabletop, lighting and background.

Both make the middle cup look like brushed silver metal while broadly preserving the arrangement of three cups and the apple. Cup rims, reflections and apple texture shift slightly; faint line artifacts are visible in the background.

Background replacement: snowy forest (unsuccessful)

Prompt: Replace only the blurred forest background with a softly blurred snowy winter forest. Preserve the same kingfisher, its colors and pose, the mossy branch, framing and foreground lighting.

Both fail the requested edit: the mossy branch disappears, the background becomes largely white instead of a snowy forest, and the bird loses photographic detail. This failure is present in both BF16 and INT8 and cannot be attributed to quantization alone.

Editing measurements (three-case timed set)

The clothing-color, cup-material and bird-background cases each have one timed call per precision after one full-settings warmup. The older-woman sweater example is a separate cold, instrumented call and is shown for visual comparison only; its latency is not pooled here. These timings must not be compared directly with the 1024px mean.

Every measured output from the three declared cases is included above. Visual observations are subjective and are not identity, material-recognition or instruction-following scores.

Raw records and reproduction: BF16, INT8, summary, process isolation, cases, measured case-file bytes, runner.

Editing example using the pipeline loaded above:

from PIL import Image
pipe.vae.enable_tiling()
image = pipe(
    prompt="Change only the blue sweater to a red sweater. Preserve the same person, face, pose, lighting and background.",
    image=Image.open("input.png"),
    width=2048, height=2048, output_resolution=2048,
    num_inference_steps=40, true_cfg_scale=1.0, use_kv_cache=True,
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]
image.save("edit.png")

Measurement methods and detailed metrics

The 1024px main suite has seven cases and two seeds per case; the adult portrait set and 2048px editing set are reported separately. Each precision runs in a fresh process, with the preceding process exited before the next starts. Process IDs, GPU process listings, zero-allocation startup and per-image CUDA baselines are recorded. Both use the bundled CPU-offload helper.

Paired calls use identical inputs, prompts, CPU random generators, seeds, dimensions, 40 steps, CFG=1 and KV cache. Timed suites exclude a full-settings warmup. BF16 runs before INT8; each case/seed has one timed call, with no confidence interval. Desktop background activity is not strictly controlled, so single-call timings are indicative only. Timing includes encoding, denoising, decoding and CPU transfers, but excludes model loading and file writing. Peak allocated/reserved memory is process-local PyTorch CUDA usage, not total board memory or a minimum GPU capacity. Only the RTX 5090 was tested.

Supplementary portrait summary

Per-image timing and memory

Pixel drift (not a semantic quality score)

RGB MAE and PSNR use a white composite so invisible RGB values in transparent pixels do not dominate the comparison. Alpha MAE is measured separately.

Additional memory telemetry

Loading is measured separately: BF16 peaks at 0.00 GiB CUDA allocated (loaded to CPU), INT8 at 9.34 GiB with the bundled sequential loader. Neither value is a minimum supported GPU capacity.

Reserved memory is the PyTorch CUDA allocator peak. CPU RSS is the process snapshot after inference, not a peak.

Evaluation scope and runtime

The main suite has 14 pairs; the separate portrait supplement has 2 pairs. All 16 pairs / 32 original PNG images from these suites are shown below. Supplementary measurements are kept separate from the main summary average.

All previews link to their full-resolution original PNG. Click an image to open it; no crop, retouching or regenerated substitute is used.

Machine-readable evidence

The measured data and all images above are embedded in this card. Original files remain available for independent checking:

• —Core CSV · Portrait CSV
• —Core BF16 JSONL · Core INT8 JSONL
• —Portrait BF16 JSONL · Portrait INT8 JSONL
• —Core environment · INT8 environment
• —File checksums · Quantization provenance

License and modifications

The weights remain subject to the Qwen Research License Agreement: non-commercial research and evaluation only; commercial use requires a separate license from the original licensor. Redistribution must preserve that agreement, attribution and modification notices. See Notice and CHANGES.md.

Source: Hugging Face | ModelScope.