Hugging Face's logo

InsecureErasure
/
Z-Image-Turbo-NVFP4

Text-to-Image
Diffusers
English
Chinese
comfyui
quantization
nvfp4
txt2img
Model card Files
xet
 Community
InsecureErasure commited on
Commit
513544a
·
verified ·
1 Parent(s): a3b00eb

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +15 -15
README.md CHANGED
@@ -103,10 +103,10 @@ convert_to_quant -i $1 \
103
 
104
  | File | Description |
105
  |---|---|
106
- | `z_image_turbo_nvfp4_mixed.safetensors` | Quantized weights |
107
- | `z_image_turbo_nvfp4_mixed_lora.safetensors` | Error-correction LoRA (rank 32) |
108
 
109
- Use the LoRA at **1.5–2.0** strength in ComfyUI for maximum fidelity.
110
 
111
  ## Requirements
112
 
@@ -115,18 +115,18 @@ Use the LoRA at **1.5–2.0** strength in ComfyUI for maximum fidelity.
115
 
116
  ## Comparison
117
 
118
- | | NVFP4 Mixed (this) | [MXFP8 Uniform](https://huggingface.co/InsecureErasure/Z-Image-Turbo-MXFP8) | [Official NVFP4](https://huggingface.co/Comfy-Org/z_image_turbo) |
119
- |---|---|---|---|---:|
120
- | **Size** | 4.84 GB | 6.23 GB | 4.51 GB |
121
- | **Base format** | NVFP4 (4-bit) | MXFP8 (8-bit) | NVFP4 (4-bit) |
122
- | **Custom layers** | ~100 tensors → MXFP8 | None | None |
123
- | **BF16 exclusions** | ~20 surgical | 8 patterns | Refiners fully BF16 |
124
- | **Learned rounding** | ✅ 6000 iter | ❌ `--simple` | ❌ |
125
- | **LoRA** | ✅ rank 32 | ❌ | ❌ |
126
- | **Refiner block 0** | MXFP8 | MXFP8 | BF16 |
127
- | **Late adaLN (22–29)** | BF16 | BF16 | NVFP4 ⚠️ |
128
- | **Last QKV (layer 29)** | BF16 | BF16 | NVFP4 ⚠️ |
129
- | **Quantization time¹** | ~60–90 min | ~5–10 min | N/A |
130
 
131
  ¹ Estimated on RTX 5060 (Blackwell) with `comfy-kitchen` CUDA kernels.
132
 
 
103
 
104
  | File | Description |
105
  |---|---|
106
+ | `z_image_turbo_nvfp4.safetensors` | Quantized weights |
107
+ | `z_image_turbo_nvfp4_lora.safetensors` | Error-correction LoRA (rank 32) |
108
 
109
+ Use the LoRA with variable strength in ComfyUI for improved fidelity.
110
 
111
  ## Requirements
112
 
 
115
 
116
  ## Comparison
117
 
118
+ | | NVFP4 Mixed (this) | MXFP8 Uniform | Official NVFP4 |
119
+ | --- | --- | --- | --- |
120
+ | Size | 4.84 GB | 6.23 GB | 4.51 GB |
121
+ | Base format | NVFP4 (4-bit) | MXFP8 (8-bit) | NVFP4 (4-bit) |
122
+ | Custom layers | ~100 tensors → MXFP8 | None | None |
123
+ | BF16 exclusions | ~20 tensors | 8 patterns | Refiners fully BF16 |
124
+ | Learned rounding | ✅ 6000 iter | ❌ --simple | ❌ |
125
+ | LoRA | ✅ rank 32 | ❌ | ❌ |
126
+ | Refiner block 0 | MXFP8 | MXFP8 | BF16 |
127
+ | Late adaLN (22–29) | BF16 | BF16 | NVFP4 ⚠️ |
128
+ | Last QKV (layer 29) | BF16 | BF16 | NVFP4 ⚠️ |
129
+ | Quantization time¹ | ~60–90 min | ~5–10 min | N/A |
130
 
131
  ¹ Estimated on RTX 5060 (Blackwell) with `comfy-kitchen` CUDA kernels.
132