Skip to page content
SAVRN
Search Plan a project

Open-weight model · Image to image

Real-ESRGAN-OpenVINO-INT4

by Max Sun HelloSun/Real-ESRGAN-OpenVINO-INT4

Real-ESRGAN-OpenVINO-INT4 is an open-weight model for image to image from Max Sun, released under Apache License 2.0. Its published files total 105.5 MB.

把 Real-ESRGAN 的 RealESRGANx4plus (RRDBNet,23 個 RRDB block,16.7 M 參數)轉成 OpenVINO IR,並用 NNCF 做成 INT4 / INT8 / FP16 多種精度版本。所有模型都是純 OpenVINO 推論, 執行時不需要 PyTorch,只要 openvino + opencv + numpy(首次使用會自動從本 repo 下載權重)。 測試機:Intel Xeon Platinum 8559C(Emerald Rapids,16…

Parameters—
Context—
Weights79.7 MB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads—

Model Card

By Max Sun, published under apache-2.0, revision 29e3ef0ad3ef.

把 Real-ESRGAN 的 RealESRGANx4plus (RRDBNet,23 個 RRDB block,16.7 M 參數)轉成 OpenVINO IR,並用 NNCF 做成 INT4 / INT8 / FP16 多種精度版本。所有模型都是純 OpenVINO 推論, 執行時不需要 PyTorch,只要 openvino + opencv + numpy(首次使用會自動從本 repo 下載權重)。 測試機:Intel Xeon Platinum 8559C(Emerald Rapids,16 vCPU,含 AMX-INT8 / AVX512-VNNI / AVX512-FP16)、 OpenVINO 2026.4.1、tile=128、tilepad=16。 畫質(3 張 inputs/ 測試圖,PSNR / SSIM 對 fp16 版本;fp16 與 PyTorch FP32 差異 直接下載 00sourcelr.mp4 你的瀏覽器不支援 video 標籤,請改用 直接下載 20sidebyside1920x780.mp4 若播放器不支援 HTML video,可直接看這支 GIF: 原始 PyTorch FP32(放大後) 你的瀏覽器不支援 video 標籤,請改用 直接下載 10pytorchfp324x.mp4 OpenVINO FP16 你的瀏覽器不支援 video 標籤,請改用 直接下載 11openvinofp164x.mp4 OpenVINO INT8(W8A8) 你的瀏覽器不支援 video 標籤,請改用 直接下載…

Read Max Sun's full model card

把 Real-ESRGAN 的 RealESRGAN_x4plus (RRDBNet,23 個 RRDB block,16.7 M 參數)轉成 OpenVINO IR,並用 NNCF 做成 INT4 / INT8 / FP16 多種精度版本。所有模型都是純 OpenVINO 推論, 執行時不需要 PyTorch,只要 openvino + opencv + numpy(首次使用會自動從本 repo 下載權重)。

模型 key 檔案 精度 大小 (xml+bin) 說明
int4 models/realesrgan_x4plus_int4_w4a16.* W4A16(權重 u4 per-channel asym,activation FP16) 11.0 MB 純 4-bit,351 個 conv 全部 4-bit,體積最小
int4-mixed models/realesrgan_x4plus_int4_mixed_w4a16.* W4/W8 混合 15.2 MB 約一半權重 4-bit、最敏感的一批保留 INT8,畫質明顯較好
int4-w4a8 models/realesrgan_x4plus_int4_w4a8.* W4A8(權重 4-bit、activation INT8) 11.8 MB 純 4-bit 權重 + INT8 計算路徑
int8 models/realesrgan_x4plus_int8_w8a8.* W8A8 19.5 MB CPU 上最快,畫質/速度最佳平衡點
fp16 models/realesrgan_x4plus_fp16.* FP16 35.1 MB 幾乎無損(對 FP32 約 48 dB),當作 reference 基準
fp32 用 export_openvino_int4.py 產生 FP32 66.8 MB 未壓縮,僅供對照/量測 PSNR 用

原始 PyTorch 權重檔:RealESRGAN_x4plus.pth(67 MB,來自 Real-ESRGAN v0.1.0 release), checkpoint 內的 key 是 params_ema。本 repo 的轉換管線已驗證 與 BasicSR RRDBNet 逐 bit 相同 (max|diff| = 0.0)。


1. 實測結果

測試機:Intel Xeon Platinum 8559C(Emerald Rapids,16 vCPU,含 AMX-INT8 / AVX512-VNNI / AVX512-FP16)、 OpenVINO 2026.4.1、tile=128、tile_pad=16。

畫質(3 張 inputs/ 測試圖,PSNR / SSIM 對 fp16 版本;fp16 與 PyTorch FP32 差異 < 0.3 dB):

模型 大小 PSNR ↑ SSIM ↑ PSNR vs. 原圖 ↑(bicubic LR 輸入)
fp16 35.1 MB 基準 1.0000 18.90 dB
int8 (W8A8) 19.5 MB 29.62 dB 0.927 19.25 dB
int4-mixed 15.2 MB 24.73 dB 0.897 17.77 dB
int4 (W4A16) 11.0 MB 18.41 dB 0.679 16.83 dB
int4-w4a8 11.8 MB 18.72 dB 0.618 16.95 dB

速度(合成 512×256 輸入、tile=128、取 3 次最佳):

模型 秒 / 張 輸入 MPix/s 備註
int8 (W8A8) 1.30 s 0.10 AMX/VNNI INT8 路徑,比 FP32 快約 1.9×
fp16 2.48 s 0.05
int4-mixed 2.28 s 0.06
int4 (W4A16) 2.32 s 0.06
int4-w4a8 2.49 s 0.05 INT8 quantize/dequantize 節點反而稍慢

為什麼 INT4 沒有比較快?

OpenVINO CPU plugin 執行 u4 權重時會在 runtime 解壓回 FP32/FP16 再做卷積, 所以 4-bit 的收益是 模型體積 / 記憶體(4×)而不是速度;要速度請用 int8(W8A8)。 4-bit 權重大多數時候適合「記憶體受限 / 想把模型塞進邊緣裝置 / 減少下載量」的場景。

為什麼 INT4 畫質掉得比較多?

Real-ESRGAN 是純卷積的深層殘差網路(23 個 RRDB,每層輸出 x*0.2 + conv(x)), 誤差會沿 351 層累積。4-bit 每層權重 SQNR 只有 ~17 dB,累積後 PSNR 掉到 ~18 dB。 因此本 repo 同時提供:

  • int4-mixed:由 NNCF 敏感度分析挑出「最不能壓」的層保留 INT8,其餘 4-bit → 幾乎無損的畫質(SSIM 0.90+),大小只多 38%。
  • int8:想要又快又準就用這個。

1.5 效果對比(同一張圖:原始 PyTorch vs 各個 OpenVINO 版本)

測試照片是 NASA 肯尼迪太空中心的公有領域白頭海雕照片 (KSC-20230425-PH-JBS02_0003), 裁切 1024×768 當 Ground Truth,bicubic 降 4 倍成 256×192 餵進模型, 所有模型都用 tile=128 推論成 768×1024。

細節並排比較(左上 bicubic、右上原始 PyTorch FP32、中上 FP16、左下 INT8、下中 INT4 混合、右下純 INT4):

模型 大小 PSNR vs PyTorch FP32 ↑ SSIM ↑ PSNR vs 原圖 ↑ 推論時間
pytorch_fp32(原始 PyTorch,基準) 67.0 MB 基準 1.0000 20.12 dB 4.46 s
bicubic x4(對照) – 22.54 dB 0.7049 21.70 dB –
openvino_fp16.jpg 35.1 MB 48.38 dB 0.9981 20.14 dB 0.60 s
openvino_int8_w8a8.jpg 19.5 MB 27.76 dB 0.8675 20.39 dB 0.39 s
openvino_int4_mixed_w4a16.jpg 15.2 MB 22.84 dB 0.8160 18.68 dB 0.61 s
openvino_int4_w4a16.jpg 11.0 MB 18.02 dB 0.4879 17.19 dB 0.60 s
openvino_int4_w4a8.jpg 11.8 MB 17.89 dB 0.4624 17.09 dB 0.69 s

重現方式、完整檔案清單與說明見 samples/eagle_4x/README.md, 照片授權為 NASA 公有領域(Public Domain)。


1.6 影片對比(放大前 vs 各個量化版本,repo 內可直接播放)

除了靜態圖,repo 也放了一段公有領域的 NASA 影片(ISS 氣輝夜景,GSFC_20181022_ICON_m12902_Airglow), 用原始 PyTorch 模型與所有 OpenVINO 版本做 640×360 → 2560×1440 的 4 倍放大,逐影格記錄時間與畫質。

① 放大前(模型輸入,640×360,1.5 s / 45 影格)

② 並排比較影片(bicubic / PyTorch FP32 / FP16 / INT8 / INT4 mixed / INT4,含標籤)

若播放器不支援 HTML video,可直接看這支 GIF:

③ 各個模型的 4x 放大影片(2560×1440)

原始 PyTorch FP32(放大後)

OpenVINO FP16

OpenVINO INT8(W8A8)

OpenVINO INT4 混合精度(W4/W8)

OpenVINO 純 INT4(W4A16)

④ 城市燈光帶 1:1 放大比較(最能看出量化差別)

⑤ 時間與畫質紀錄

模型 大小 45 影格總時間 ms/影格 張/秒 vs PyTorch PSNR ↑ SSIM ↑ 加速比 ↑
pytorch_fp32(原始) 67.0 MB 663.5 s 14745 0.07 基準 1.0000 1.0×
openvino_fp16 35.1 MB 211.5 s 4700 0.21 45.41 dB 0.9927 3.1×
openvino_int8_w8a8 19.5 MB 124.8 s 2773 0.36 36.90 dB 0.9284 5.3×
openvino_int4_mixed_w4a16 15.2 MB 214.6 s 4769 0.21 31.31 dB 0.9344 3.1×
openvino_int4_w4a16 11.0 MB 246.8 s 5485 0.18 25.40 dB 0.7894 2.7×

直接下載:samples/video_airglow/00_source_lr.mp4、10_pytorch_fp32_4x.mp4、11_openvino_fp16_4x.mp4、 12_openvino_int8_w8a8_4x.mp4、13_openvino_int4_mixed_4x.mp4、14_openvino_int4_4x.mp4、 20_side_by_side_1920x780.mp4(完整說明見 samples/video_airglow/README.md)

重點結論:INT8 (W8A8) 在 CPU 上比 PyTorch FP32 快 5.3× 且畫質僅 36.9 dB / SSIM 0.93; INT4 在 CPU 上不會更快(CPU plugin 執行時會把 4-bit 權重解壓回 FP32/FP16), 它的優勢是模型只有 11 MB;若要在 4-bit 尺寸下兼顧畫質,請用 int4-mixed。


2. 安裝

pip install openvino opencv-python numpy huggingface_hub
# 只有要「自己重新轉換模型」才需要:
pip install torch onnx onnxscript nncf

realesrgan_ov.py 會在第一次執行時自動從本 repo 下載 *.xml / *.bin, 並快取到 ~/.cache/realesrgan_openvino(可用環境變數 REALESRGAN_OV_CACHE 改位置)。 也可以手動下載後用 --model-path / -m /path/to/model.xml 指定本機檔案。


3. 快速開始:單一 py 檔

下載兩個檔案即可(realesrgan_ov.py + benchmark_models.py),模型會自動下載。

# quick_start.py
import cv2
from realesrgan_ov import RealESRGAN

upscaler = RealESRGAN(model='int4')              # int4 | int4-mixed | int8 | fp16 | 自訂 .xml 路徑
img = cv2.imread('input.jpg')                     # BGR uint8
sr, scale = upscaler.enhance(img, tile=128)       # 4 倍超解析(記憶體吃太多時用 tile 分塊)
cv2.imwrite('output.png', sr)
print('output:', sr.shape, 'scale:', scale)

inputs/ 內附 3 張測試圖(來自 Real-ESRGAN 官方 inputs),samples/ 內附 1 張公有領域照片與 1 段公有領域影片(都已用所有模型跑過),可直接試跑:

python realesrgan_ov.py -i inputs/00003.png -o results/ -m int4-mixed -t 128
# 命令列
python realesrgan_ov.py -i input.jpg                     # 4x 超解析,輸出 results/
python realesrgan_ov.py -i ./photos -o out/ -m int8      # 整個資料夾
python realesrgan_ov.py -i clip.mp4 -o clip_4x.mp4      # 影片逐影格
python realesrgan_ov.py -i big.png -t 64 --tile-pad 16   # 大圖分塊(256→1024 這類)
python realesrgan_ov.py -i input.jpg --outscale 2        # 只放大 2 倍(先 4x 再 downscale)
python realesrgan_ov.py -d GPU -m int8 -i input.jpg      # 用 GPU / NPU / AUTO
python realesrgan_ov.py --benchmark --images input.jpg   # 只測效能
python benchmark_models.py --images a.png b.png          # 比較所有版本的畫質/速度

支援 PNG/BMP/JPG/WEBP(含 alpha 通道,alpha 會用 bicubic 放大)、資料夾遞迴、影片(mp4/avi/mov/mkv)。


4. Python API

from realesrgan_ov import RealESRGAN

up = RealESRGAN(
    model='int4',        # int4 / int4-mixed / int4-w4a8 / int8 / fp16 / '/path/model.xml'
    device='CPU',        # CPU / GPU / NPU / AUTO
    precision='auto',    # auto(預設)/ f32 / f16 / bf16
    num_threads=8,       # 0 = 使用全部核心
)

sr, scale = up.enhance(img, tile=128, tile_pad=16, pre_pad=0, outscale=4)
sr_bgra   = up.enhance_alpha(bgra_img, tile=128)     # 自動處理 alpha

precision 說明(很重要):新版 OpenVINO CPU 對 FP32 graph 的預設 INFERENCE_PRECISION_HINT 是 bf16,會讓這個網路的 PSNR 掉約 20 dB。 realesrgan_ov.py 的 auto 會對 FP32/FP16/INT4(FP16 activation) 的模型自動釘成 f32, INT8 graph 不受影響。若你自己用 ov.Core().compile_model(),請記得傳 {'INFERENCE_PRECISION_HINT': 'f32'}。


5. 從 PyTorch 權重重新轉換成 OpenVINO INT4

python export_openvino_int4.py \
    --weights RealESRGAN_x4plus.pth \
    --data-dir inputs \        # 校準用的低解析度圖
    --out-dir models \
    --calib-tile 128 --calib-samples 6 \
    --weight-ratio 1.0 \       # 1.0 = 全部 4-bit;0.5 = 一半 4-bit、其餘 INT8
    --a8-bits 8 --a8-name int8_w8a8

流程:RRDBNet(重建) → strict=True 載入 checkpoint → ONNX(opset17, 動態 H/W) → OpenVINO FP32/FP16 → NNCF 量化 → IR

這個管線裡有兩個必須處理的地方(否則 INT4 會被靜默降級成 INT8):

  1. nncf.compress_weights(mode='int4_*') 預設只把 MatMul 壓成 4-bit,卷積一律落到 INT8 backup → enable_int4_for_convolutions() 擴充 _get_ratio_defining_params,讓 Conv 也吃 4-bit。
  2. NNCF 的 CPU hardware config 只允許 8/16-bit 權重 → enable_int4_weights_in_cpu_hw_config() 把 q4_w 加進 Convolution 的 qspace,產生 u4 常數(OpenVINO CPU plugin 可執行)。
  3. openvino.convert_model() 預設會把權重壓成 FP16(本模型會掉到 ~27 dB)→ convert_onnx() 以 compress_to_fp16=False 轉換,得到真正的 FP32 graph。

驗證方式(本 repo 產出的模型都跑過):

python - <<'PY'
from openvino import Core
m = Core().read_model('models/realesrgan_x4plus_int4_w4a16.xml')
print(sorted({str(n.get_element_type()) for n in m.get_ops() if n.get_type_name() == 'Constant'}))
PY
# 應該看到 '<Type: 'uint4_t'>',代表權重真的是 4-bit

export_openvino_int4.py 也能轉換其他 Real-ESRGAN 權重,只要架構相同(--stem 改檔名):

  • RealESRGAN_x4plus.pth(本 repo 使用,num_feat=64 / num_block=23 / num_grow_ch=32)
  • RealESRNet_x4plus.pth、RealESRGAN_x4plus_anime_6B.pth、RealESRGAN_x2plus.pth(需在 rrdbnet_arch.py 對應調整 scale / num_block / num_feat)

6. 疑難排解

症狀 原因 / 解決
輸出整張有雜訊、幾乎是亂圖 用了 int4:Real-ESRGAN 對 4-bit 非常敏感,請改用 int4-mixed 或 int8
輸出整體偏色 / 邊緣有雜訊 沒有指定 INFERENCE_PRECISION_HINT='f32',CPU 預設跑 bf16。用 precision='f32' 或 -m int8
大圖 OOM 用 -t 128(或 64)分塊,--tile-pad 16;不要 -t 0
分塊後有接縫 --tile-pad 調大到 20~32
下載權重失敗 export HF_TOKEN=hf_xxx(私有 repo)或先手動下載 models/*.xml *.bin 後用 -m 指定
匯出時 KeyError: 'input' LWC 傳入 dataset 的已知問題,script 預設不傳(--lwc-dataset 可開啟 AWQ / scale-estimation)
想改輸出的放大倍率 --outscale 2(>4 會再 cubic 放大)

7. 出處與授權

  • 程式碼/權重出處:xinntao/Real-ESRGAN、XPixelGroup/BasicSR(Apache-2.0)
  • rrdbnet_arch.py 為 BasicSR RRDBNet 的獨立重寫(不需安裝 BasicSR),參數名稱與官方 checkpoint 完全一致
  • 量化:OpenVINO NNCF
  • 本 repo 的推論程式、轉換腳本與本說明文檔為本專案新增。
  • 樣品照片:NASA Kennedy Space Center(公有領域),見 samples/eagle_4x/README.md 的來源與授權說明
  • 樣品影片:NASA/Goddard ICON 任務 ISS 氣輝畫面(公有領域),見 samples/video_airglow/README.md
  • License:Apache-2.0(見 LICENSE)。

Identity and Version

Repository
HelloSun/Real-ESRGAN-OpenVINO-INT4
Publisher
Max Sun
Task
Image to image
Modality
Image
Library
openvino
Parameters
Not stated by the source
Languages
Not stated by the source
Revision
29e3ef0ad3ef7e274c235ab3da1f5c1892e36f7c
First published
2026-10-02
Last updated
2026-10-02

Files and Weights

50 files, 105.5 MB in total. The weights are 5 files totalling 79.7 MB in bin.

Weights5 files · 79.7 MB
Configuration8 files · 53.5 KB
Documentation4 files · 31.9 KB
Other31 files · 25.7 MB
Repository2 files · 2.0 KB
Every file
FileTypeSizeSHA-256
models/realesrgan_x4plus_fp16.binWeights33.4 MB a2da277ec4d2
models/realesrgan_x4plus_int4_mixed_w4a16.binWeights12.6 MB c7240c0cd9e9
models/realesrgan_x4plus_int4_w4a16.binWeights8.4 MB 92fe92857c7e
models/realesrgan_x4plus_int4_w4a8.binWeights8.4 MB 1a319144d5cf
models/realesrgan_x4plus_int8_w8a8.binWeights16.8 MB 90113c7a8319
benchmark_models.pyConfiguration4.0 KB —
export_openvino_int4.pyConfiguration12.0 KB —
realesrgan_ov.pyConfiguration13.5 KB —
rrdbnet_arch.pyConfiguration3.3 KB —
samples/eagle_4x/metrics.jsonConfiguration1.2 KB —
samples/video_airglow/metrics.jsonConfiguration1.5 KB —
tools/make_sample.pyConfiguration7.3 KB —
tools/make_video_sample.pyConfiguration10.6 KB —
LICENSEDocumentation1.5 KB —
README.mdDocumentation17.5 KB —
samples/eagle_4x/README.mdDocumentation4.5 KB —
samples/video_airglow/README.mdDocumentation8.4 KB —
input/nasa_ICON_airglow_640x360_63s-66s.mp4Other933.6 KB 74cfe01f6dc0
input/nasa_bald_eagle_KSC-20230425-PH-JBS02_0003.jpgOther270.5 KB 9221f63bb6e5
inputs/00003.pngOther164.4 KB d37932ae7d31
inputs/0014.jpgOther9.6 KB —
inputs/0030.jpgOther12.0 KB —
models/realesrgan_x4plus_fp16.xmlOther1.7 MB a64a0e2410bd
models/realesrgan_x4plus_int4_mixed_w4a16.xmlOther2.5 MB 1f8aad8edc47
models/realesrgan_x4plus_int4_w4a16.xmlOther2.5 MB ff5cfc36ff0b
models/realesrgan_x4plus_int4_w4a8.xmlOther3.4 MB f625a7bbc86e
models/realesrgan_x4plus_int8_w8a8.xmlOther2.7 MB 07e3fa1898d1
requirements.txtOther198 B —
samples/eagle_4x/00_input_lr_x4.pngOther122.2 KB 55e6c84c5d08
samples/eagle_4x/01_bicubic_x4.jpgOther213.2 KB 44941d735a74
samples/eagle_4x/02_ground_truth.jpgOther366.4 KB 55b8da1db95a
samples/eagle_4x/03_pytorch_fp32_reference.jpgOther390.5 KB 297c42fb8a6d
samples/eagle_4x/10_detail_montage.jpgOther275.0 KB 4dd065de88ee
samples/eagle_4x/openvino_fp16.jpgOther304.0 KB b28805300360
samples/eagle_4x/openvino_int4_mixed_w4a16.jpgOther335.2 KB e2ef0b284c79
samples/eagle_4x/openvino_int4_w4a16.jpgOther338.3 KB dfe1b18b2e39
samples/eagle_4x/openvino_int4_w4a8.jpgOther355.2 KB 523ee1b4f2b7
samples/eagle_4x/openvino_int8_w8a8.jpgOther296.8 KB af6e1ab61426
samples/video_airglow/00_source_lr.mp4Other405.1 KB 0bf16c8a6d16
samples/video_airglow/10_pytorch_fp32_4x.mp4Other787.2 KB 5b171def478b
samples/video_airglow/11_openvino_fp16_4x.mp4Other779.2 KB 1a2b42cc7734
samples/video_airglow/12_openvino_int8_w8a8_4x.mp4Other927.3 KB 51423cc8a28d
samples/video_airglow/13_openvino_int4_mixed_4x.mp4Other847.1 KB cb7d64813c41
samples/video_airglow/14_openvino_int4_4x.mp4Other1.3 MB 97b5dee51569
samples/video_airglow/20_side_by_side_1920x780.mp4Other418.0 KB b11618fa15f5
samples/video_airglow/21_poster_frame.jpgOther365.3 KB 6f807a991524
samples/video_airglow/22_poster_detail_zoom.jpgOther205.7 KB d85ae2b872fc
samples/video_airglow/22_side_by_side_preview.gifOther2.3 MB 49302583a56f
.gitattributesRepository1.9 KB —
.gitignoreRepository47 B —

License and Download

License
apache-2.0
Access
Open weights, no gate
Download size
79.7 MB
Download from Max Sun

Released by Max Sun through its official repository on Hugging Face. Read the license.

Memory Requirements

PrecisionWeights in memory
As published79.7 MB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About Real-ESRGAN-OpenVINO-INT4

Can I use Real-ESRGAN-OpenVINO-INT4 commercially?

Yes. Real-ESRGAN-OpenVINO-INT4 is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

Similar Models

Model · Image to image

FLUX.2-klein-4B-GGUF

Unsloth AI

This is a GGUF quantized version of FLUX.2-klein-4B. unsloth/FLUX.2-klein-4B-GGUF uses Unsloth Dynamic 2.0 methodology for SOTA performance. - Important layers are upcasted to higher precision. - Uses tooling from ComfyUI-GGUF by city96. The FLUX.2 [klein] model family are our fastest image models to date. FLUX.2 [klein] unifies generation and editing in a single compact architecture, delivering state-of-the-art quality with end-to-end inference in as low as under a second. Built for applications that require real-time image generation without sacrificing quality, and runs on consumer hardware, with as little as 13GB VRAM. FLUX.2 [klein] 4B is a 4 billion parameter rectified flow…

Open weights apache-2.0 ggml

Model · Image to image

FLUX.2-klein-4b-fp8

Black Forest Labs

FLUX.2 [klein] 4B is a 4 billion parameter rectified flow transformer capable of generating images from text descriptions and supports multi-reference editing capabilities. For more information, please read our blog post. This repository holds an FP8 version of FLUX.2 [klein] 4B. The main repository of this model (full BF16 weights) can be found here. Limitations - This model is not intended or able to provide factual information. - While the model can output text, text rendered may be inaccurate or subject to distortion. - As a statistical model, this checkpoint may represent or amplify biases observed in the training data. - The model may fail to generate output that matches the prompts.…

Open weights apache-2.0 diffusion-single-file

A LoRA adapter for the Qwen-Image-Edit-2511 diffusion pipeline, enabling NSFW content generation and editing capabilities. The easiest way to use this LoRA is through the ScottzillaSystems Image Editor — select " MCNL-NSFW-v1" from the adapter dropdown. Use these concepts in your prompts to activate specific capabilities: nsfw nipples vagina penis missionary cowgirlout reversecowgirlpov blowjob cumonface creamp1e l1ck

Open weights openrail++ diffusers

FLUX.2 [klein] 4B Base is a 4 billion parameter rectified flow transformer capable of generating images from text descriptions and supports multi-reference editing capabilities. For more information, please read our blog post. This repository holds an FP8 version of FLUX.2 [klein] 4B Base. The main repository of this model (full BF16 weights) can be found here. Limitations - This model is not intended or able to provide factual information. - While the model can output text, text rendered may be inaccurate or subject to distortion. - As a statistical model, this checkpoint may represent or amplify biases observed in the training data. - The model may fail to generate output that matches the…

Open weights apache-2.0 diffusion-single-file

Model · Image to image

safetensored-upscalers

Ashtaka

compact is quick and decent i guess - dat is very slow but great quality (dark go to) - drct similar quality to dat but slightly faster - esrgan is old and mediocre - nafnet is an image restoration architecture - omnisr is quick and good quality (nub go to) - rcan is old but good - real-esrgan is esrgan but real (mediocre) - span is quick and good quality (my go to) I do not own any models in this repo. These are just converted pickles to safetensors. The script used is modified from https://github.com/OpenModelDB/model-hub/blob/main/tools/convert-pth.py with Gemini 2.5 Pro.

Open weights mit

Model · Image to image

stopmotion-consistency-v1-lora

SSH

Built with Qwen. An experimental next-frame image-editing adapter fine-tuned from the original BF16 Qwen-Image-2.1. The adapter was trained on 112 curated stop-motion transitions using an RTX PRO 6000. Quality is limited. Matched tests show improved static-area retention, but motion magnitude, direction and shape consistency remain unreliable. This is a small research pilot, not a proven solution for long, consistent animation. The adapter is stopmotion-consistency-v1.safetensors, 67,141,056 bytes. It contains only attention LoRA updates, not the base model. SHA-256: d68479ed2c246bc7ebea8a62019b10da903772d30854385302b2d66de99df51a. Use DiffSynth-Studio at commit…

Open weights other