Text-to-Image
Diffusers
Safetensors
English
Chinese
QwenImage21Pipeline
bitsandbytes
int8
image-generation
image-editing
rgba
8-bit precision
Instructions to use ixim/Image21-INT8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use ixim/Image21-INT8 with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("ixim/Image21-INT8", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
- DiffusionBee
Download evaluation/audit/summary.json from ixim/Image21-INT8: direct link, hf CLI and curl.
- Browser
- Download file 6.66 kB
-
https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/audit/summary.json
- Command line
-
hf download hf://ixim/Image21-INT8/evaluation/audit/summary.json
-
curl -L -o summary.json https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/audit/summary.json
6.66 kB
| { | |
| "process_isolation": [ | |
| { | |
| "name": "bf16", | |
| "launcher_pid": 7376, | |
| "inference_pid": 25920, | |
| "exit_code": 0, | |
| "start_allocated_bytes": 0, | |
| "inference_pid_absent_from_gpu_after_exit": true | |
| }, | |
| { | |
| "name": "int8", | |
| "launcher_pid": 35376, | |
| "inference_pid": 29636, | |
| "exit_code": 0, | |
| "start_allocated_bytes": 0, | |
| "inference_pid_absent_from_gpu_after_exit": true | |
| }, | |
| { | |
| "name": "young_bf16", | |
| "launcher_pid": 31516, | |
| "inference_pid": 35212, | |
| "exit_code": 0, | |
| "start_allocated_bytes": 0, | |
| "inference_pid_absent_from_gpu_after_exit": true | |
| }, | |
| { | |
| "name": "young_int8", | |
| "launcher_pid": 5400, | |
| "inference_pid": 11368, | |
| "exit_code": 0, | |
| "start_allocated_bytes": 0, | |
| "inference_pid_absent_from_gpu_after_exit": true | |
| } | |
| ], | |
| "original_and_corrected_identical_images": [ | |
| [ | |
| "", | |
| "bf16", | |
| "portrait", | |
| 42 | |
| ], | |
| [ | |
| "", | |
| "bf16", | |
| "portrait", | |
| 123 | |
| ], | |
| [ | |
| "", | |
| "bf16", | |
| "english_text", | |
| 42 | |
| ], | |
| [ | |
| "", | |
| "bf16", | |
| "english_text", | |
| 123 | |
| ], | |
| [ | |
| "", | |
| "bf16", | |
| "chinese_text", | |
| 42 | |
| ], | |
| [ | |
| "", | |
| "bf16", | |
| "chinese_text", | |
| 123 | |
| ], | |
| [ | |
| "", | |
| "bf16", | |
| "composition", | |
| 42 | |
| ], | |
| [ | |
| "", | |
| "bf16", | |
| "composition", | |
| 123 | |
| ], | |
| [ | |
| "", | |
| "bf16", | |
| "texture", | |
| 42 | |
| ], | |
| [ | |
| "", | |
| "bf16", | |
| "texture", | |
| 123 | |
| ], | |
| [ | |
| "", | |
| "bf16", | |
| "rgba", | |
| 42 | |
| ], | |
| [ | |
| "", | |
| "bf16", | |
| "rgba", | |
| 123 | |
| ], | |
| [ | |
| "", | |
| "bf16", | |
| "edit", | |
| 42 | |
| ], | |
| [ | |
| "", | |
| "bf16", | |
| "edit", | |
| 123 | |
| ], | |
| [ | |
| "", | |
| "int8", | |
| "portrait", | |
| 42 | |
| ], | |
| [ | |
| "", | |
| "int8", | |
| "portrait", | |
| 123 | |
| ], | |
| [ | |
| "", | |
| "int8", | |
| "english_text", | |
| 42 | |
| ], | |
| [ | |
| "", | |
| "int8", | |
| "english_text", | |
| 123 | |
| ], | |
| [ | |
| "", | |
| "int8", | |
| "chinese_text", | |
| 42 | |
| ], | |
| [ | |
| "", | |
| "int8", | |
| "chinese_text", | |
| 123 | |
| ], | |
| [ | |
| "", | |
| "int8", | |
| "composition", | |
| 42 | |
| ], | |
| [ | |
| "", | |
| "int8", | |
| "composition", | |
| 123 | |
| ], | |
| [ | |
| "", | |
| "int8", | |
| "texture", | |
| 42 | |
| ], | |
| [ | |
| "", | |
| "int8", | |
| "texture", | |
| 123 | |
| ], | |
| [ | |
| "", | |
| "int8", | |
| "rgba", | |
| 42 | |
| ], | |
| [ | |
| "", | |
| "int8", | |
| "rgba", | |
| 123 | |
| ], | |
| [ | |
| "", | |
| "int8", | |
| "edit", | |
| 42 | |
| ], | |
| [ | |
| "", | |
| "int8", | |
| "edit", | |
| 123 | |
| ], | |
| [ | |
| "young_woman", | |
| "bf16", | |
| "young_chinese_woman", | |
| 42 | |
| ], | |
| [ | |
| "young_woman", | |
| "bf16", | |
| "young_chinese_woman", | |
| 123 | |
| ], | |
| [ | |
| "young_woman", | |
| "int8", | |
| "young_chinese_woman", | |
| 42 | |
| ], | |
| [ | |
| "young_woman", | |
| "int8", | |
| "young_chinese_woman", | |
| 123 | |
| ] | |
| ], | |
| "diagnostic_identical_images": [ | |
| "edit_fresh", | |
| "portrait", | |
| "edit_repeat", | |
| "edit_rgb", | |
| "edit_no_cache" | |
| ], | |
| "initial_peak": { | |
| "bf16": 19.081242561340332, | |
| "int8": 23.080156326293945 | |
| }, | |
| "corrected_peak": { | |
| "bf16": 19.081242561340332, | |
| "int8": 12.087220668792725 | |
| }, | |
| "load_peak": { | |
| "bf16": 0.0, | |
| "int8": 9.335469722747803 | |
| }, | |
| "native": { | |
| "bf16": { | |
| "name": "edit_native", | |
| "seed": 42, | |
| "input_mode": "original", | |
| "kv_cache": true, | |
| "width": 2048, | |
| "height": 2048, | |
| "steps": 40, | |
| "cfg": 1.0, | |
| "vae_tiling": true, | |
| "seconds": 233.8040861999998, | |
| "peak_allocated_bytes": 28467959296, | |
| "image_sha256": "c551ac809ccb48146d1ad842d75975c0e9663d388a368010d2bfccb7883eeeec", | |
| "input_sha256": "01e37e178fb0d35d5c706b7edd76bc1a6c5b50ec3536901a467fd45295d097aa" | |
| }, | |
| "int8": { | |
| "name": "edit_native", | |
| "seed": 42, | |
| "input_mode": "original", | |
| "kv_cache": true, | |
| "width": 2048, | |
| "height": 2048, | |
| "steps": 40, | |
| "cfg": 1.0, | |
| "vae_tiling": true, | |
| "seconds": 213.56440760000078, | |
| "peak_allocated_bytes": 23243435008, | |
| "image_sha256": "b6af18d8688bb2669632adf0cfae962bc34431e30e935014997010cde5c2ca59", | |
| "input_sha256": "01e37e178fb0d35d5c706b7edd76bc1a6c5b50ec3536901a467fd45295d097aa" | |
| } | |
| }, | |
| "editing_conclusion_en": "At 1024px, enabling VAE tiling alone still produces excessive sharpening. At 2048px with the same tiling configuration, both BF16 and INT8 produce a red sweater while substantially preserving the soft lighting and natural skin texture. This controlled comparison supports resolution sensitivity in this editing case (including reference-image preprocessing). It does not identify a deeper model/pipeline cause or establish universal editing reliability.", | |
| "editing_conclusion_zh": "1024px 下仅开启 VAE 分块仍过度锐化;2048px 配合同样的分块设置时,BF16 和 INT8 均生成红色毛衣,并较好保持柔和光线与自然皮肤质感。这一受控对照支持该编辑用例对分辨率敏感(包括参考图预处理分辨率),但尚未确定更深层的模型/流水线原因,也不能证明所有编辑任务均可靠。", | |
| "images": [ | |
| { | |
| "path": "input.png", | |
| "en": "Shared input", | |
| "zh": "两种精度共同输入" | |
| }, | |
| { | |
| "path": "int8-original/edit_fresh.png", | |
| "en": "Fresh-process INT8, 1024px (failure reproduced)", | |
| "zh": "新进程 INT8,1024px(复现失败)" | |
| }, | |
| { | |
| "path": "int8-original/edit_rgb.png", | |
| "en": "1024px RGB input (not improved)", | |
| "zh": "1024px RGB 输入(未改善)" | |
| }, | |
| { | |
| "path": "int8-original/edit_no_cache.png", | |
| "en": "1024px KV cache off (not improved)", | |
| "zh": "1024px 关闭 KV 缓存(未改善)" | |
| }, | |
| { | |
| "path": "int8-fixed/vae_roundtrip.png", | |
| "en": "VAE-only round trip (no diffusion editing)", | |
| "zh": "仅 VAE 编解码(未进行扩散编辑)" | |
| }, | |
| { | |
| "path": "int8-1024-tiled/edit_fresh.png", | |
| "en": "1024px with VAE tiling (not improved)", | |
| "zh": "1024px 加 VAE 分块(未改善)" | |
| } | |
| ], | |
| "aborted_native_untiled": "Stopped owned diagnostic process during prolonged VAE decoding after board memory saturated; no completed output or timing.", | |
| "profiler_note": "The early after_delete event retained bound-method references and is not evidence of process-exit cleanup. All process isolation claims use fresh startups and explicit subprocess exits." | |
| } |