DarthAffe/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2025-12-12 13:28:37 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat: add support for SD2.x with TINY U-Nets (#939)

Browse Source
This commit is contained in:
akleine 2025-11-09 15:47:37 +01:00 committed by GitHub
parent 0fa3e1a383
commit d2d3944f50
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
5 changed files with 68 additions and 47 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

79
docs/distilled_sd.md
View File

@ -1,40 +1,66 @@
# Running distilled models: SSD1B and SD1.x with tiny U-Nets
# Running distilled models: SSD1B and SDx.x with tiny U-Nets
## Preface
This kind of models have a reduced U-Net part.
Unlike other SDXL models the U-Net of SSD1B has only one middle block and lesser attention layers in up and down blocks, resulting in relatively smaller files. Running these models saves more than 33% of the time. For more details, refer to Segmind's paper on https://arxiv.org/abs/2401.02677v1 .
Unlike other SD 1.x models Tiny-UNet models consist of only 6 U-Net blocks, resulting in relatively smaller files (approximately 1 GB). Running these models saves almost 50% of the time. For more details, refer to the paper: https://arxiv.org/pdf/2305.15798.pdf .
These models feature a reduced U-Net architecture. Unlike standard SDXL models, the SSD-1B U-Net contains only one middle block and fewer attention layers in its up- and down-blocks, resulting in significantly smaller file sizes. Using these models can reduce inference time by more than 33%. For more details, refer to Segmind's paper: https://arxiv.org/abs/2401.02677v1.
Similarly, SD1.x- and SD2.x-style models with a tiny U-Net consist of only 6 U-Net blocks, leading to very small files and time savings of up to 50%. For more information, see the paper: https://arxiv.org/pdf/2305.15798.pdf.
## SSD1B
Unfortunately not all of this models follow the standard model parameter naming mapping.
Anyway there are some very useful SSD1B models available online, such as:
Note that not all of these models follow the standard parameter naming conventions. However, several useful SSD-1B models are available online, such as:
* https://huggingface.co/segmind/SSD-1B/resolve/main/SSD-1B-A1111.safetensors
* https://huggingface.co/hassenhamdi/SSD-1B-fp8_e4m3fn/resolve/main/SSD-1B_fp8_e4m3fn.safetensors
Also there are useful LORAs available:
Useful LoRAs are also available:
* https://huggingface.co/seungminh/lora-swarovski-SSD-1B/resolve/main/pytorch_lora_weights.safetensors
* https://huggingface.co/kylielee505/mylcmlorassd/resolve/main/pytorch_lora_weights.safetensors
You can use this files **out-of-the-box** - unlike models in next section.
These files can be used out-of-the-box, unlike the models described in the next section.
## SD1.x with tiny U-Nets
## SD1.x, SD2.x with tiny U-Nets
There are some Tiny SD 1.x models available online, such as:
These models require conversion before use. You will need a Python script provided by the diffusers team, available on GitHub:
* https://raw.githubusercontent.com/huggingface/diffusers/refs/heads/main/scripts/convert_diffusers_to_original_stable_diffusion.py
### SD2.x
NotaAI provides the following model online:
* https://huggingface.co/nota-ai/bk-sdm-v2-tiny
Creating a .safetensors file involves two steps. First, run this short Python script to download the model from Hugging Face:
```python
from diffusers import StableDiffusionPipeline
pipe = StableDiffusionPipeline.from_pretrained("nota-ai/bk-sdm-v2-tiny",cache_dir="./")
```
Second, create the .safetensors file by running:
```bash
python convert_diffusers_to_original_stable_diffusion.py \
--model_path models--nota-ai--bk-sdm-v2-tiny/snapshots/68277af553777858cd47e133f92e4db47321bc74 \
--checkpoint_path bk-sdm-v2-tiny.safetensors --half --use_safetensors
```
This will generate the **file bk-sdm-v2-tiny.safetensors**, which is now ready for use with sd.cpp.
### SD1.x
Several Tiny SD 1.x models are available online, such as:
* https://huggingface.co/segmind/tiny-sd
* https://huggingface.co/segmind/portrait-finetuned
* https://huggingface.co/nota-ai/bk-sdm-tiny
These models need some conversion, for example because partially tensors are **non contiguous** stored. To create a usable checkpoint file, follow these **easy** steps:
These models also require conversion, partly because some tensors are stored in a non-contiguous manner. To create a usable checkpoint file, follow these simple steps:
Download and prepare the model using Python:
### Download model from Hugging Face
Download the model using Python on your computer, for example this way:
##### Download the model using Python on your computer, for example this way:
```python
import torch
@ -46,35 +72,22 @@ for param in unet.parameters():
pipe.save_pretrained("segmindtiny-sd", safe_serialization=True)
```
### Convert that to a ckpt file
To convert the downloaded model to a checkpoint file, you need another Python script. Download the conversion script from here:
* https://raw.githubusercontent.com/huggingface/diffusers/refs/heads/main/scripts/convert_diffusers_to_original_stable_diffusion.py
### Run convert script
Now, run that conversion script:
##### Run the conversion script:
```bash
python convert_diffusers_to_original_stable_diffusion.py \
--model_path ./segmindtiny-sd \
--checkpoint_path ./segmind_tiny-sd.ckpt --half
--model_path ./segmindtiny-sd \
--checkpoint_path ./segmind_tiny-sd.ckpt --half
```
The file **segmind_tiny-sd.ckpt** will be generated and is now ready to use with sd.cpp
You can follow a similar process for other models mentioned above from Hugging Face.
The file segmind_tiny-sd.ckpt will be generated and is now ready for use with sd.cpp. You can follow a similar process for the other models mentioned above.
### Another ckpt file on the net
There is another model file available online:
### Another available .ckpt file:
* https://huggingface.co/ClashSAN/small-sd/resolve/main/tinySDdistilled.ckpt
If you want to use that, you have to adjust some **non-contiguous tensors** first:
To use this file, you must first adjust its non-contiguous tensors:
```python
import torch

3
model.cpp
View File

@ -1788,6 +1788,9 @@ SDVersion ModelLoader::get_sd_version() {
if (is_inpaint) {
return VERSION_SD2_INPAINT;
}
if (!has_middle_block_1) {
return VERSION_SD2_TINY_UNET;
}
return VERSION_SD2;
}
return VERSION_COUNT;

3
model.h
View File

@ -26,6 +26,7 @@ enum SDVersion {
VERSION_SD1_TINY_UNET,
VERSION_SD2,
VERSION_SD2_INPAINT,
VERSION_SD2_TINY_UNET,
VERSION_SDXL,
VERSION_SDXL_INPAINT,
VERSION_SDXL_PIX2PIX,
@ -52,7 +53,7 @@ static inline bool sd_version_is_sd1(SDVersion version) {
}
static inline bool sd_version_is_sd2(SDVersion version) {
if (version == VERSION_SD2 || version == VERSION_SD2_INPAINT) {
if (version == VERSION_SD2 || version == VERSION_SD2_INPAINT || version == VERSION_SD2_TINY_UNET) {
return true;
}
return false;

1
stable-diffusion.cpp
View File

@ -23,6 +23,7 @@ const char* model_version_to_str[] = {
"SD 1.x Tiny UNet",
"SD 2.x",
"SD 2.x Inpaint",
"SD 2.x Tiny UNet",
"SDXL",
"SDXL Inpaint",
"SDXL Instruct-Pix2Pix",

21
unet.hpp
View File

@ -180,6 +180,7 @@ protected:
int num_head_channels = -1; // channels // num_heads
int context_dim = 768; // 1024 for VERSION_SD2, 2048 for VERSION_SDXL
bool use_linear_projection = false;
bool tiny_unet = false;
public:
int model_channels = 320;
@ -208,15 +209,17 @@ public:
num_head_channels = 64;
num_heads = -1;
use_linear_projection = true;
} else if (version == VERSION_SD1_TINY_UNET) {
num_res_blocks = 1;
channel_mult = {1, 2, 4};
}
if (sd_version_is_inpaint(version)) {
in_channels = 9;
} else if (sd_version_is_unet_edit(version)) {
in_channels = 8;
}
if (version == VERSION_SD1_TINY_UNET || version == VERSION_SD2_TINY_UNET) {
num_res_blocks = 1;
channel_mult = {1, 2, 4};
tiny_unet = true;
}
// dims is always 2
// use_temporal_attention is always True for SVD
@ -290,7 +293,7 @@ public:
context_dim));
}
input_block_chans.push_back(ch);
if (version == VERSION_SD1_TINY_UNET) {
if (tiny_unet) {
input_block_idx++;
}
}
@ -311,7 +314,7 @@ public:
d_head = num_head_channels;
n_head = ch / d_head;
}
if (version != VERSION_SD1_TINY_UNET) {
if (!tiny_unet) {
blocks["middle_block.0"] = std::shared_ptr<GGMLBlock>(get_resblock(ch, time_embed_dim, ch));
if (version != VERSION_SDXL_SSD1B) {
blocks["middle_block.1"] = std::shared_ptr<GGMLBlock>(get_attention_layer(ch,
@ -358,7 +361,7 @@ public:
}
if (i > 0 && j == num_res_blocks) {
if (version == VERSION_SD1_TINY_UNET) {
if (tiny_unet) {
output_block_idx++;
if (output_block_idx == 2) {
up_sample_idx = 1;
@ -495,7 +498,7 @@ public:
}
hs.push_back(h);
}
if (version == VERSION_SD1_TINY_UNET) {
if (tiny_unet) {
input_block_idx++;
}
if (i != len_mults - 1) {
@ -512,7 +515,7 @@ public:
// [N, 4*model_channels, h/8, w/8]
// middle_block
if (version != VERSION_SD1_TINY_UNET) {
if (!tiny_unet) {
h = resblock_forward("middle_block.0", ctx, h, emb, num_video_frames); // [N, 4*model_channels, h/8, w/8]
if (version != VERSION_SDXL_SSD1B) {
h = attention_layer_forward("middle_block.1", ctx, h, context, num_video_frames); // [N, 4*model_channels, h/8, w/8]
@ -554,7 +557,7 @@ public:
}
if (i > 0 && j == num_res_blocks) {
if (version == VERSION_SD1_TINY_UNET) {
if (tiny_unet) {
output_block_idx++;
if (output_block_idx == 2) {
up_sample_idx = 1;