DarthAffe/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-02-04 10:53:34 +00:00
Code Issues Packages Projects Releases Wiki Activity

support longcat-image-edit

Browse Source 
Fix base rope offset for ref images
This commit is contained in:
Stéphane du Hamel 2025-12-08 13:43:37 +01:00
parent 9f225e4e63
commit c044a406c5
2 changed files with 130 additions and 60 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

75
conditioner.hpp
View File

@ -1690,7 +1690,7 @@ struct LLMEmbedder : public Conditioner {
std::string current_part;
for (char c : curr_text) {
if (c == '\'') {
if (c == '"') {
if (!current_part.empty()) {
parts.push_back(current_part);
current_part.clear();
@ -1711,7 +1711,7 @@ struct LLMEmbedder : public Conditioner {
for (const auto& part : parts) {
if (part.empty())
continue;
if (part[0] == '\'' && part.back() == '\'') {
if (part[0] == '"' && part.back() == '"') {
std::string quoted_content = part.substr(1, part.size() - 2);
for (char ch : quoted_content) {
std::string char_str(1, ch);
@ -1747,6 +1747,76 @@ struct LLMEmbedder : public Conditioner {
bool spell_quotes = false;
std::set<int> out_layers;
if (llm->enable_vision && conditioner_params.ref_images.size() > 0) {
if (sd_version_is_longcat(version)) {
LOG_INFO("LongCatEditPipeline");
prompt_template_encode_start_idx = 67;
// prompt_template_encode_end_idx = 5;
int image_embed_idx = 36 + 6;
int min_pixels = 384 * 384;
int max_pixels = 560 * 560;
std::string placeholder = "<|image_pad|>";
std::string img_prompt;
// Only one image is officicially supported by the model, not sure how it handles multiple images
for (int i = 0; i < conditioner_params.ref_images.size(); i++) {
sd_image_f32_t image = sd_image_t_to_sd_image_f32_t(*conditioner_params.ref_images[i]);
double factor = llm->params.vision.patch_size * llm->params.vision.spatial_merge_size;
int height = image.height;
int width = image.width;
int h_bar = static_cast<int>(std::round(height / factor)) * factor;
int w_bar = static_cast<int>(std::round(width / factor)) * factor;
if (static_cast<double>(h_bar) * w_bar > max_pixels) {
double beta = std::sqrt((height * width) / static_cast<double>(max_pixels));
h_bar = std::max(static_cast<int>(factor),
static_cast<int>(std::floor(height / beta / factor)) * static_cast<int>(factor));
w_bar = std::max(static_cast<int>(factor),
static_cast<int>(std::floor(width / beta / factor)) * static_cast<int>(factor));
} else if (static_cast<double>(h_bar) * w_bar < min_pixels) {
double beta = std::sqrt(static_cast<double>(min_pixels) / (height * width));
h_bar = static_cast<int>(std::ceil(height * beta / factor)) * static_cast<int>(factor);
w_bar = static_cast<int>(std::ceil(width * beta / factor)) * static_cast<int>(factor);
}
LOG_DEBUG("resize conditioner ref image %d from %dx%d to %dx%d", i, image.height, image.width, h_bar, w_bar);
sd_image_f32_t resized_image = clip_preprocess(image, w_bar, h_bar);
free(image.data);
image.data = nullptr;
ggml_tensor* image_tensor = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, resized_image.width, resized_image.height, 3, 1);
sd_image_f32_to_ggml_tensor(resized_image, image_tensor, false);
free(resized_image.data);
resized_image.data = nullptr;
ggml_tensor* image_embed = nullptr;
llm->encode_image(n_threads, image_tensor, &image_embed, work_ctx);
image_embeds.emplace_back(image_embed_idx, image_embed);
image_embed_idx += 1 + image_embed->ne[1] + 6;
img_prompt += "<|vision_start|>";
int64_t num_image_tokens = image_embed->ne[1];
img_prompt.reserve(num_image_tokens * placeholder.size());
for (int j = 0; j < num_image_tokens; j++) {
img_prompt += placeholder;
}
img_prompt += "<|vision_end|>";
}
max_length = 512;
spell_quotes = true;
prompt = "<|im_start|>system\nAs an image editing expert, first analyze the content and attributes of the input image(s). Then, based on the user's editing instructions, clearly and precisely determine how to modify the given image(s), ensuring that only the specified parts are altered and all other aspects remain consistent with the original(s).<|im_end|>\n<|im_start|>user\n";
prompt += img_prompt;
prompt_attn_range.first = static_cast<int>(prompt.size());
prompt += conditioner_params.text;
prompt_attn_range.second = static_cast<int>(prompt.size());
prompt += "<|im_end|>\n<|im_start|>assistant\n";
} else {
LOG_INFO("QwenImageEditPlusPipeline");
prompt_template_encode_start_idx = 64;
int image_embed_idx = 64 + 6;
@ -1809,6 +1879,7 @@ struct LLMEmbedder : public Conditioner {
prompt_attn_range.second = static_cast<int>(prompt.size());
prompt += "<|im_end|>\n<|im_start|>assistant\n";
}
} else if (sd_version_is_flux2(version)) {
prompt_template_encode_start_idx = 0;
out_layers = {10, 20, 30};

9
rope.hpp
View File

@ -107,7 +107,6 @@ namespace Rope {
std::vector<float> row_ids = linspace<float>(h_offset, h_len - 1 + h_offset, h_len);
std::vector<float> col_ids = linspace<float>(w_offset, w_len - 1 + w_offset, w_len);
for (int i = 0; i < h_len; ++i) {
for (int j = 0; j < w_len; ++j) {
img_ids[i * w_len + j][0] = index;
@ -183,8 +182,8 @@ namespace Rope {
float ref_index_scale,
int base_offset = 0) {
std::vector<std::vector<float>> ids;
uint64_t curr_h_offset = base_offset;
uint64_t curr_w_offset = base_offset;
uint64_t curr_h_offset = 0;
uint64_t curr_w_offset = 0;
int index = start_index;
for (ggml_tensor* ref : ref_latents) {
uint64_t h_offset = 0;
@ -203,8 +202,8 @@ namespace Rope {
bs,
axes_dim_num,
static_cast<int>(index * ref_index_scale),
h_offset,
w_offset);
h_offset + base_offset,
w_offset + base_offset);
ids = concat_ids(ids, ref_ids, bs);
if (increase_ref_index) {