upstream-archive/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-07-23 19:30:54 -05:00
Code Issues Packages Projects Releases Wiki Activity

feat: add Qwen Image Edit support (#877)

Browse Source 
* add ref latent support for qwen image

* optimize clip_preprocess and fix get_first_stage_encoding

* add qwen2vl vit support

* add qwen image edit support

* fix qwen image edit pipeline

* add mmproj file support

* support dynamic number of Qwen image transformer blocks

* set prompt_template_encode_start_idx every time

* to_add_out precision fix

* to_out.0 precision fix

* update docs
This commit is contained in:
leejet
2025-10-13 23:17:18 +08:00
committed by GitHub
parent c64994dc1d
commit 2e9242e37f
18 changed files with 1339 additions and 365 deletions
Download Patch File Download Diff File Expand all files Collapse all files

68
util.cpp
View File

@@ -84,6 +84,7 @@ int round_up_to(int value, int base) {
}
#ifdef _WIN32 // code for windows
#define NOMINMAX
#include <windows.h>
bool file_exists(const std::string& filename) {
@@ -298,7 +299,7 @@ std::string trim(const std::string& s) {
static sd_log_cb_t sd_log_cb = NULL;
void* sd_log_cb_data = NULL;
#define LOG_BUFFER_SIZE 1024
#define LOG_BUFFER_SIZE 4096
void log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...) {
va_list args;
@@ -387,10 +388,10 @@ sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int
float original_x = (float)x * image.width / target_width;
float original_y = (float)y * image.height / target_height;
int x1 = (int)original_x;
int y1 = (int)original_y;
int x2 = x1 + 1;
int y2 = y1 + 1;
uint32_t x1 = (uint32_t)original_x;
uint32_t y1 = (uint32_t)original_y;
uint32_t x2 = std::min(x1 + 1, image.width - 1);
uint32_t y2 = std::min(y1 + 1, image.height - 1);
for (int k = 0; k < image.channel; k++) {
float v1 = *(image.data + y1 * image.width * image.channel + x1 * image.channel + k);
@@ -427,23 +428,26 @@ float means[3] = {0.48145466, 0.4578275, 0.40821073};
float stds[3] = {0.26862954, 0.26130258, 0.27577711};
// Function to clip and preprocess sd_image_f32_t
sd_image_f32_t clip_preprocess(sd_image_f32_t image, int size) {
float scale = (float)size / fmin(image.width, image.height);
sd_image_f32_t clip_preprocess(sd_image_f32_t image, int target_width, int target_height) {
float width_scale = (float)target_width / image.width;
float height_scale = (float)target_height / image.height;
float scale = std::fmax(width_scale, height_scale);
// Interpolation
int new_width = (int)(scale * image.width);
int new_height = (int)(scale * image.height);
float* resized_data = (float*)malloc(new_width * new_height * image.channel * sizeof(float));
int resized_width = (int)(scale * image.width);
int resized_height = (int)(scale * image.height);
float* resized_data = (float*)malloc(resized_width * resized_height * image.channel * sizeof(float));
for (int y = 0; y < new_height; y++) {
for (int x = 0; x < new_width; x++) {
float original_x = (float)x * image.width / new_width;
float original_y = (float)y * image.height / new_height;
for (int y = 0; y < resized_height; y++) {
for (int x = 0; x < resized_width; x++) {
float original_x = (float)x * image.width / resized_width;
float original_y = (float)y * image.height / resized_height;
int x1 = (int)original_x;
int y1 = (int)original_y;
int x2 = x1 + 1;
int y2 = y1 + 1;
uint32_t x1 = (uint32_t)original_x;
uint32_t y1 = (uint32_t)original_y;
uint32_t x2 = std::min(x1 + 1, image.width - 1);
uint32_t y2 = std::min(y1 + 1, image.height - 1);
for (int k = 0; k < image.channel; k++) {
float v1 = *(image.data + y1 * image.width * image.channel + x1 * image.channel + k);
@@ -456,26 +460,28 @@ sd_image_f32_t clip_preprocess(sd_image_f32_t image, int size) {
float value = interpolate(v1, v2, v3, v4, x_ratio, y_ratio);
*(resized_data + y * new_width * image.channel + x * image.channel + k) = value;
*(resized_data + y * resized_width * image.channel + x * image.channel + k) = value;
}
}
}
// Clip and preprocess
int h = (new_height - size) / 2;
int w = (new_width - size) / 2;
int h_offset = std::max((int)(resized_height - target_height) / 2, 0);
int w_offset = std::max((int)(resized_width - target_width) / 2, 0);
sd_image_f32_t result;
result.width = size;
result.height = size;
result.width = target_width;
result.height = target_height;
result.channel = image.channel;
result.data = (float*)malloc(size * size * image.channel * sizeof(float));
result.data = (float*)malloc(target_height * target_width * image.channel * sizeof(float));
for (int k = 0; k < image.channel; k++) {
for (int i = 0; i < size; i++) {
for (int j = 0; j < size; j++) {
*(result.data + i * size * image.channel + j * image.channel + k) =
fmin(fmax(*(resized_data + (i + h) * new_width * image.channel + (j + w) * image.channel + k), 0.0f), 255.0f) / 255.0f;
for (int i = 0; i < result.height; i++) {
for (int j = 0; j < result.width; j++) {
int src_y = std::min(i + h_offset, resized_height - 1);
int src_x = std::min(j + w_offset, resized_width - 1);
*(result.data + i * result.width * image.channel + j * image.channel + k) =
fmin(fmax(*(resized_data + src_y * resized_width * image.channel + src_x * image.channel + k), 0.0f), 255.0f) / 255.0f;
}
}
}
@@ -485,10 +491,10 @@ sd_image_f32_t clip_preprocess(sd_image_f32_t image, int size) {
// Normalize
for (int k = 0; k < image.channel; k++) {
for (int i = 0; i < size; i++) {
for (int j = 0; j < size; j++) {
for (int i = 0; i < result.height; i++) {
for (int j = 0; j < result.width; j++) {
// *(result.data + i * size * image.channel + j * image.channel + k) = 0.5f;
int offset = i * size * image.channel + j * image.channel + k;
int offset = i * result.width * image.channel + j * image.channel + k;
float value = *(result.data + offset);
value = (value - means[k]) / stds[k];
// value = 0.5f;