Hugging Face's logo

nunchaku-ai
/
nunchaku-qwen-image-edit

Text-to-Image
Diffusers
English
image-editing
SVDQuant
Qwen-Image-Edit
Diffusion
Quantization
ICLR2025
Model card Files
xet
 Community
12
Lmxyy commited on
Commit
1e4b8c0
·
verified ·
1 Parent(s): 4739070

Upload ./README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +54 -13
README.md CHANGED
@@ -46,18 +46,59 @@ No recent news. Stay tuned for updates!
46
 
47
  ### Model Files
48
 
49
- - [`svdq-int4_r32-qwen-image-edit.safetensors`](./svdq-int4_r32-qwen-image-edit.safetensors): SVDQuant INT4 (rank 32) Qwen-Image-Edit model. For users with non-Blackwell GPUs (pre-50-series).
50
- - [`svdq-int4_r128-qwen-image-edit.safetensors`](./svdq-int4_r128-qwen-image-edit.safetensors): SVDQuant INT4 (rank 128) Qwen-Image-Edit model. For users with non-Blackwell GPUs (pre-50-series). It offers better quality than the rank 32 model, but it is slower.
51
- - [`svdq-int4_r32-qwen-image-edit-lightningv1.0-4steps.safetensors`](./svdq-int4_r32-qwen-image-edit-lightningv1.0-4steps.safetensors): SVDQuant INT4 (rank 32) 4-step Qwen-Image-Edit model by fusing [Qwen-Image-Edit-Lightning-4steps-V1.0-bf16.safetensors](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-4steps-V1.0-bf16.safetensors) using LoRA strength = 1.0. For users with non-Blackwell GPUs (pre-50-series).
52
- - [`svdq-int4_r128-qwen-image-edit-lightningv1.0-4steps.safetensors`](./svdq-int4_r128-qwen-image-edit-lightningv1.0-4steps.safetensors): SVDQuant INT4 (rank 128) 4-step Qwen-Image-Edit model by fusing [Qwen-Image-Edit-Lightning-4steps-V1.0-bf16.safetensors](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-4steps-V1.0-bf16.safetensors) using LoRA strength = 1.0. For users with non-Blackwell GPUs (pre-50-series).
53
- - [`svdq-int4_r32-qwen-image-edit-lightningv1.0-8steps.safetensors`](./svdq-int4_r32-qwen-image-edit-lightningv1.0-8steps.safetensors): SVDQuant INT4 (rank 32) 8-step Qwen-Image-Edit model by fusing [Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors) using LoRA strength = 1.0. For users with non-Blackwell GPUs (pre-50-series).
54
- - [`svdq-int4_r128-qwen-image-edit-lightningv1.0-8steps.safetensors`](./svdq-int4_r128-qwen-image-edit-lightningv1.0-8steps.safetensors): SVDQuant INT4 (rank 128) 8-step Qwen-Image-Edit model by fusing [Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors) using LoRA strength = 1.0. For users with non-Blackwell GPUs (pre-50-series).
55
- - [`svdq-fp4_r32-qwen-image-edit.safetensors`](./svdq-fp4_r32-qwen-image-edit.safetensors): SVDQuant NVFP4 (rank 32) Qwen-Image-Edit model. For users with Blackwell GPUs (50-series).
56
- - [`svdq-fp4_r128-qwen-image-edit.safetensors`](./svdq-fp4_r128-qwen-image-edit.safetensors): SVDQuant NVFP4 (rank 128) Qwen-Image-Edit model. For users with Blackwell GPUs (50-series). It offers better quality than the rank 32 model, but it is slower.
57
- - [`svdq-fp4_r32-qwen-image-edit-lightningv1.0-4steps.safetensors`](./svdq-fp4_r32-qwen-image-edit-lightningv1.0-4steps.safetensors): SVDQuant NVFP4 (rank 32) 4-step Qwen-Image-Edit model by fusing [Qwen-Image-Edit-Lightning-4steps-V1.0-bf16.safetensors](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-4steps-V1.0-bf16.safetensors) using LoRA strength = 1.0. For users with Blackwell GPUs (50-series).
58
- - [`svdq-fp4_r128-qwen-image-edit-lightningv1.0-4steps.safetensors`](./svdq-fp4_r128-qwen-image-edit-lightningv1.0-4steps.safetensors): SVDQuant NVFP4 (rank 128) 4-step Qwen-Image-Edit model by fusing [Qwen-Image-Edit-Lightning-4steps-V1.0-bf16.safetensors](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-4steps-V1.0-bf16.safetensors) using LoRA strength = 1.0. For users with Blackwell GPUs (50-series).
59
- - [`svdq-fp4_r32-qwen-image-edit-lightningv1.0-8steps.safetensors`](./svdq-fp4_r32-qwen-image-edit-lightningv1.0-8steps.safetensors): SVDQuant NVFP4 (rank 32) 8-step Qwen-Image-Edit model by fusing [Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors) using LoRA strength = 1.0. For users with Blackwell GPUs (50-series).
60
- - [`svdq-fp4_r128-qwen-image-edit-lightningv1.0-8steps.safetensors`](./svdq-fp4_r128-qwen-image-edit-lightningv1.0-8steps.safetensors): SVDQuant NVFP4 (rank 128) 8-step Qwen-Image-Edit model by fusing [Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors) using LoRA strength = 1.0. For users with Blackwell GPUs (50-series).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
61
 
62
 
63
  ### Model Sources
@@ -65,7 +106,7 @@ No recent news. Stay tuned for updates!
65
  - **Inference Engine:** [nunchaku](https://github.com/nunchaku-tech/nunchaku)
66
  - **Quantization Library:** [deepcompressor](https://github.com/nunchaku-tech/deepcompressor)
67
  - **Paper:** [SVDQuant: Absorbing Outliers by Low-Rank Components for 4-Bit Diffusion Models](http://arxiv.org/abs/2411.05007)
68
- - **Demo:** [svdquant.mit.edu](https://svdquant.mit.edu)
69
 
70
  ## Usage
71
 
 
46
 
47
  ### Model Files
48
 
49
+ **Data Type**: `INT4` for non-Blackwell GPUs (pre-50-series), `NVFP4` for Blackwell GPUs (50-series).
50
+ **Rank**: `r32` for faster inference, `r128` for better quality but slower inference.
51
+
52
+
53
+
54
+
55
+
56
+ ### Base Models
57
+
58
+ Standard inference speed models for general use
59
+
60
+
61
+ | Data Type | Rank | Model Name | Comment |
62
+ |-----------|------|----------|---------|
63
+ | INT4 | r32 | [`svdq-int4_r32-qwen-image-edit.safetensors`](./svdq-int4_r32-qwen-image-edit.safetensors) | |
64
+ | | r128 | [`svdq-int4_r128-qwen-image-edit.safetensors`](./svdq-int4_r128-qwen-image-edit.safetensors) | |
65
+ | NVFP4 | r32 | [`svdq-fp4_r32-qwen-image-edit.safetensors`](./svdq-fp4_r32-qwen-image-edit.safetensors) | |
66
+ | | r128 | [`svdq-fp4_r128-qwen-image-edit.safetensors`](./svdq-fp4_r128-qwen-image-edit.safetensors) | |
67
+
68
+
69
+
70
+
71
+
72
+ ### 4-Step Distilled Models
73
+
74
+ 4-step distilled models fused with [Qwen-Image-Edit-Lightning-4steps-V1.0 LoRA](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-4steps-V1.0-bf16.safetensors) using LoRA strength = 1.0
75
+
76
+
77
+ | Data Type | Rank | Model Name | Comment |
78
+ |-----------|------|----------|---------|
79
+ | INT4 | r32 | [`svdq-int4_r32-qwen-image-edit-lightningv1.0-4steps.safetensors`](./svdq-int4_r32-qwen-image-edit-lightningv1.0-4steps.safetensors) | Fused with [Qwen-Image-Edit-Lightning-4steps-V1.0 LoRA](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-4steps-V1.0-bf16.safetensors) |
80
+ | | r128 | [`svdq-int4_r128-qwen-image-edit-lightningv1.0-4steps.safetensors`](./svdq-int4_r128-qwen-image-edit-lightningv1.0-4steps.safetensors) | Fused with [Qwen-Image-Edit-Lightning-4steps-V1.0 LoRA](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-4steps-V1.0-bf16.safetensors). Better quality, slower inference |
81
+ | NVFP4 | r32 | [`svdq-fp4_r32-qwen-image-edit-lightningv1.0-4steps.safetensors`](./svdq-fp4_r32-qwen-image-edit-lightningv1.0-4steps.safetensors) | Fused with [Qwen-Image-Edit-Lightning-4steps-V1.0 LoRA](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-4steps-V1.0-bf16.safetensors) |
82
+ | | r128 | [`svdq-fp4_r128-qwen-image-edit-lightningv1.0-4steps.safetensors`](./svdq-fp4_r128-qwen-image-edit-lightningv1.0-4steps.safetensors) | Fused with [Qwen-Image-Edit-Lightning-4steps-V1.0 LoRA](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-4steps-V1.0-bf16.safetensors). Better quality, slower inference |
83
+
84
+
85
+
86
+
87
+
88
+ ### 8-Step Distilled Models
89
+
90
+ 8-step distilled models fused with [Qwen-Image-Edit-Lightning-8steps-V1.0 LoRA](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors) using LoRA strength = 1.0
91
+
92
+
93
+ | Data Type | Rank | Model Name | Comment |
94
+ |-----------|------|----------|---------|
95
+ | INT4 | r32 | [`svdq-int4_r32-qwen-image-edit-lightningv1.0-8steps.safetensors`](./svdq-int4_r32-qwen-image-edit-lightningv1.0-8steps.safetensors) | Fused with [Qwen-Image-Edit-Lightning-8steps-V1.0 LoRA](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors) |
96
+ | | r128 | [`svdq-int4_r128-qwen-image-edit-lightningv1.0-8steps.safetensors`](./svdq-int4_r128-qwen-image-edit-lightningv1.0-8steps.safetensors) | Fused with [Qwen-Image-Edit-Lightning-8steps-V1.0 LoRA](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors). Better quality, slower inference |
97
+ | NVFP4 | r32 | [`svdq-fp4_r32-qwen-image-edit-lightningv1.0-8steps.safetensors`](./svdq-fp4_r32-qwen-image-edit-lightningv1.0-8steps.safetensors) | Fused with [Qwen-Image-Edit-Lightning-8steps-V1.0 LoRA](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors) |
98
+ | | r128 | [`svdq-fp4_r128-qwen-image-edit-lightningv1.0-8steps.safetensors`](./svdq-fp4_r128-qwen-image-edit-lightningv1.0-8steps.safetensors) | Fused with [Qwen-Image-Edit-Lightning-8steps-V1.0 LoRA](https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors). Better quality, slower inference |
99
+
100
+
101
+
102
 
103
 
104
  ### Model Sources
 
106
  - **Inference Engine:** [nunchaku](https://github.com/nunchaku-tech/nunchaku)
107
  - **Quantization Library:** [deepcompressor](https://github.com/nunchaku-tech/deepcompressor)
108
  - **Paper:** [SVDQuant: Absorbing Outliers by Low-Rank Components for 4-Bit Diffusion Models](http://arxiv.org/abs/2411.05007)
109
+ - **Demo:** [demo.nunchaku.tech](https://demo.nunchaku.tech)
110
 
111
  ## Usage
112