From a1175cf34f163b0a6023ef3ece6e92ec69a93d18 Mon Sep 17 00:00:00 2001 From: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Sat, 28 Jun 2025 22:57:07 +0800 Subject: [PATCH] merged leejet changes --- otherarch/sdcpp/diffusion_model.hpp | 58 ++++----- otherarch/sdcpp/flux.hpp | 187 ++++++++++++++++----------- otherarch/sdcpp/main.cpp | 137 ++++++++++++-------- otherarch/sdcpp/sdtype_adapter.cpp | 54 +++++++- otherarch/sdcpp/stable-diffusion.cpp | 36 +++--- 5 files changed, 296 insertions(+), 176 deletions(-) diff --git a/otherarch/sdcpp/diffusion_model.hpp b/otherarch/sdcpp/diffusion_model.hpp index dc66d3105..48522b25f 100644 --- a/otherarch/sdcpp/diffusion_model.hpp +++ b/otherarch/sdcpp/diffusion_model.hpp @@ -13,13 +13,13 @@ struct DiffusionModel { struct ggml_tensor* c_concat, struct ggml_tensor* y, struct ggml_tensor* guidance, - int num_video_frames = -1, - std::vector controls = {}, - float control_strength = 0.f, - std::vector kontext_imgs = std::vector(), - struct ggml_tensor** output = NULL, - struct ggml_context* output_ctx = NULL, - std::vector skip_layers = std::vector()) = 0; + std::vector ref_latents = {}, + int num_video_frames = -1, + std::vector controls = {}, + float control_strength = 0.f, + struct ggml_tensor** output = NULL, + struct ggml_context* output_ctx = NULL, + std::vector skip_layers = std::vector()) = 0; virtual void alloc_params_buffer() = 0; virtual void free_params_buffer() = 0; virtual void free_compute_buffer() = 0; @@ -69,13 +69,13 @@ struct UNetModel : public DiffusionModel { struct ggml_tensor* c_concat, struct ggml_tensor* y, struct ggml_tensor* guidance, - int num_video_frames = -1, - std::vector controls = {}, - float control_strength = 0.f, - std::vector kontext_imgs = std::vector(), - struct ggml_tensor** output = NULL, - struct ggml_context* output_ctx = NULL, - std::vector skip_layers = std::vector()) { + std::vector ref_latents = {}, + int num_video_frames = -1, + std::vector controls = {}, + float control_strength = 0.f, + struct ggml_tensor** output = NULL, + struct ggml_context* output_ctx = NULL, + std::vector skip_layers = std::vector()) { (void)skip_layers; // SLG doesn't work with UNet models return unet.compute(n_threads, x, timesteps, context, c_concat, y, num_video_frames, controls, control_strength, output, output_ctx); } @@ -120,13 +120,13 @@ struct MMDiTModel : public DiffusionModel { struct ggml_tensor* c_concat, struct ggml_tensor* y, struct ggml_tensor* guidance, - int num_video_frames = -1, - std::vector controls = {}, - float control_strength = 0.f, - std::vector kontext_imgs = std::vector(), - struct ggml_tensor** output = NULL, - struct ggml_context* output_ctx = NULL, - std::vector skip_layers = std::vector()) { + std::vector ref_latents = {}, + int num_video_frames = -1, + std::vector controls = {}, + float control_strength = 0.f, + struct ggml_tensor** output = NULL, + struct ggml_context* output_ctx = NULL, + std::vector skip_layers = std::vector()) { return mmdit.compute(n_threads, x, timesteps, context, y, output, output_ctx, skip_layers); } }; @@ -172,14 +172,14 @@ struct FluxModel : public DiffusionModel { struct ggml_tensor* c_concat, struct ggml_tensor* y, struct ggml_tensor* guidance, - int num_video_frames = -1, - std::vector controls = {}, - float control_strength = 0.f, - std::vector kontext_imgs = std::vector(), - struct ggml_tensor** output = NULL, - struct ggml_context* output_ctx = NULL, - std::vector skip_layers = std::vector()) { - return flux.compute(n_threads, x, timesteps, context, c_concat, y, guidance, kontext_imgs, output, output_ctx, skip_layers); + std::vector ref_latents = {}, + int num_video_frames = -1, + std::vector controls = {}, + float control_strength = 0.f, + struct ggml_tensor** output = NULL, + struct ggml_context* output_ctx = NULL, + std::vector skip_layers = std::vector()) { + return flux.compute(n_threads, x, timesteps, context, c_concat, y, guidance, ref_latents, output, output_ctx, skip_layers); } }; diff --git a/otherarch/sdcpp/flux.hpp b/otherarch/sdcpp/flux.hpp index b3e8e4e8e..e6d941af9 100644 --- a/otherarch/sdcpp/flux.hpp +++ b/otherarch/sdcpp/flux.hpp @@ -672,51 +672,81 @@ namespace Flux { } // Generate IDs for image patches and text - std::vector> gen_ids(int h, int w, int patch_size, int index = 0) { + std::vector> gen_txt_ids(int bs, int context_len) { + return std::vector>(bs * context_len, std::vector(3, 0.0)); + } + + std::vector> gen_img_ids(int h, int w, int patch_size, int bs, int index = 0, int h_offset = 0, int w_offset = 0) { int h_len = (h + (patch_size / 2)) / patch_size; int w_len = (w + (patch_size / 2)) / patch_size; - std::vector> img_ids(h_len * w_len, std::vector(3, (float)index)); + std::vector> img_ids(h_len * w_len, std::vector(3, 0.0)); - std::vector row_ids = linspace(0, h_len - 1, h_len); - std::vector col_ids = linspace(0, w_len - 1, w_len); + std::vector row_ids = linspace(h_offset, h_len - 1 + h_offset, h_len); + std::vector col_ids = linspace(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; img_ids[i * w_len + j][1] = row_ids[i]; img_ids[i * w_len + j][2] = col_ids[j]; } } - return img_ids; + std::vector> img_ids_repeated(bs * img_ids.size(), std::vector(3)); + for (int i = 0; i < bs; ++i) { + for (int j = 0; j < img_ids.size(); ++j) { + img_ids_repeated[i * img_ids.size() + j] = img_ids[j]; + } + } + return img_ids_repeated; + } + + std::vector> concat_ids(const std::vector>& a, + const std::vector>& b, + int bs) { + size_t a_len = a.size() / bs; + size_t b_len = b.size() / bs; + std::vector> ids(a.size() + b.size(), std::vector(3)); + for (int i = 0; i < bs; ++i) { + for (int j = 0; j < a_len; ++j) { + ids[i * (a_len + b_len) + j] = a[i * a_len + j]; + } + for (int j = 0; j < b_len; ++j) { + ids[i * (a_len + b_len) + a_len + j] = b[i * b_len + j]; + } + } + return ids; + } + + std::vector> gen_ids(int h, int w, int patch_size, int bs, int context_len, std::vector ref_latents) { + auto txt_ids = gen_txt_ids(bs, context_len); + auto img_ids = gen_img_ids(h, w, patch_size, bs); + + auto ids = concat_ids(txt_ids, img_ids, bs); + uint64_t curr_h_offset = 0; + uint64_t curr_w_offset = 0; + for (ggml_tensor* ref : ref_latents) { + uint64_t h_offset = 0; + uint64_t w_offset = 0; + if (ref->ne[1] + curr_h_offset > ref->ne[0] + curr_w_offset) { + w_offset = curr_w_offset; + } else { + h_offset = curr_h_offset; + } + + auto ref_ids = gen_img_ids(ref->ne[1], ref->ne[0], patch_size, bs, 1, h_offset, w_offset); + ids = concat_ids(ids, ref_ids, bs); + + curr_h_offset = std::max(curr_h_offset, ref->ne[1] + h_offset); + curr_w_offset = std::max(curr_w_offset, ref->ne[0] + w_offset); + } + return ids; } // Generate positional embeddings - std::vector gen_pe(std::vector imgs, struct ggml_tensor* context, int patch_size, int theta, const std::vector& axes_dim) { - int context_len = context->ne[1]; - int bs = imgs[0]->ne[3]; - - std::vector> img_ids; - for (int i = 0; i < imgs.size(); i++) { - auto x = imgs[i]; - if (x) { - int h = x->ne[1]; - int w = x->ne[0]; - std::vector> img_ids_i = gen_ids(h, w, patch_size, i); - img_ids.insert(img_ids.end(), img_ids_i.begin(), img_ids_i.end()); - } - } - - std::vector> txt_ids(bs * context_len, std::vector(3, 0.0)); - std::vector> ids(bs * (context_len + img_ids.size()), std::vector(3)); - for (int i = 0; i < bs; ++i) { - for (int j = 0; j < context_len; ++j) { - ids[i * (context_len + img_ids.size()) + j] = txt_ids[j]; - } - for (int j = 0; j < img_ids.size(); ++j) { - ids[i * (context_len + img_ids.size()) + context_len + j] = img_ids[j]; - } - } + std::vector gen_pe(int h, int w, int patch_size, int bs, int context_len, std::vector ref_latents, int theta, const std::vector& axes_dim) { + std::vector> ids = gen_ids(h, w, patch_size, bs, context_len, ref_latents); std::vector> trans_ids = transpose(ids); size_t pos_len = ids.size(); @@ -843,7 +873,7 @@ namespace Flux { struct ggml_tensor* guidance, struct ggml_tensor* pe, struct ggml_tensor* arange = NULL, - std::vector skip_layers = std::vector()) { + std::vector skip_layers = {}) { auto img_in = std::dynamic_pointer_cast(blocks["img_in"]); auto txt_in = std::dynamic_pointer_cast(blocks["txt_in"]); auto final_layer = std::dynamic_pointer_cast(blocks["final_layer"]); @@ -929,8 +959,23 @@ namespace Flux { return img; } + struct ggml_tensor* process_img(struct ggml_context* ctx, + struct ggml_tensor* x) { + + int64_t W = x->ne[0]; + int64_t H = x->ne[1]; + int64_t patch_size = 2; + int pad_h = (patch_size - H % patch_size) % patch_size; + int pad_w = (patch_size - W % patch_size) % patch_size; + x = ggml_pad(ctx, x, pad_w, pad_h, 0, 0); // [N, C, H + pad_h, W + pad_w] + + // img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size) + auto img = patchify(ctx, x, patch_size); // [N, h*w, C * patch_size * patch_size] + return img; + } + struct ggml_tensor* forward(struct ggml_context* ctx, - std::vector imgs, + struct ggml_tensor* x, struct ggml_tensor* timestep, struct ggml_tensor* context, struct ggml_tensor* c_concat, @@ -938,8 +983,8 @@ namespace Flux { struct ggml_tensor* guidance, struct ggml_tensor* pe, struct ggml_tensor* arange = NULL, - std::vector skip_layers = std::vector(), - SDVersion version = VERSION_FLUX) { + std::vector ref_latents = {}, + std::vector skip_layers = {}) { // Forward pass of DiT. // x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images) // timestep: (N,) tensor of diffusion timesteps @@ -950,47 +995,41 @@ namespace Flux { // pe: (L, d_head/2, 2, 2) // return: (N, C, H, W) - auto x = imgs[0]; GGML_ASSERT(x->ne[3] == 1); int64_t W = x->ne[0]; int64_t H = x->ne[1]; int64_t C = x->ne[2]; int64_t patch_size = 2; - int pad_h = (patch_size - x->ne[0] % patch_size) % patch_size; - int pad_w = (patch_size - x->ne[1] % patch_size) % patch_size; + int pad_h = (patch_size - H % patch_size) % patch_size; + int pad_w = (patch_size - W % patch_size) % patch_size; - // img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size) - ggml_tensor* img = NULL; // [N, h*w, C * patch_size * patch_size] - int64_t patchified_img_size; - for (auto& x : imgs) { - int pad_h = (patch_size - x->ne[0] % patch_size) % patch_size; - int pad_w = (patch_size - x->ne[1] % patch_size) % patch_size; - ggml_tensor* pad_x = ggml_pad(ctx, x, pad_w, pad_h, 0, 0); - pad_x = patchify(ctx, pad_x, patch_size); - if (img) { - img = ggml_concat(ctx, img, pad_x, 1); - } else { - img = pad_x; - patchified_img_size = img->ne[1]; - } - } + auto img = process_img(ctx, x); + uint64_t img_tokens = img->ne[1]; if (c_concat != NULL) { ggml_tensor* masked = ggml_view_4d(ctx, c_concat, c_concat->ne[0], c_concat->ne[1], C, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], 0); ggml_tensor* mask = ggml_view_4d(ctx, c_concat, c_concat->ne[0], c_concat->ne[1], 8 * 8, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], c_concat->nb[2] * C); - masked = ggml_pad(ctx, masked, pad_w, pad_h, 0, 0); - mask = ggml_pad(ctx, mask, pad_w, pad_h, 0, 0); - - masked = patchify(ctx, masked, patch_size); - mask = patchify(ctx, mask, patch_size); + masked = process_img(ctx, masked); + mask = process_img(ctx, mask); img = ggml_concat(ctx, img, ggml_concat(ctx, masked, mask, 0), 0); } - auto out = forward_orig(ctx, img, context, timestep, y, guidance, pe, arange, skip_layers); // [N, h*w, C * patch_size * patch_size] - out = ggml_cont(ctx, ggml_view_2d(ctx, out, out->ne[0], patchified_img_size, out->nb[1], 0)); + if (ref_latents.size() > 0) { + for (ggml_tensor* ref : ref_latents) { + ref = process_img(ctx, ref); + img = ggml_concat(ctx, img, ref, 1); + } + } + + auto out = forward_orig(ctx, img, context, timestep, y, guidance, pe, arange, skip_layers); // [N, num_tokens, C * patch_size * patch_size] + if (out->ne[1] > img_tokens) { + out = ggml_cont(ctx, ggml_permute(ctx, out, 0, 2, 1, 3)); // [num_tokens, N, C * patch_size * patch_size] + out = ggml_view_3d(ctx, out, out->ne[0], out->ne[1], img_tokens, out->nb[1], out->nb[2], 0); + out = ggml_cont(ctx, ggml_permute(ctx, out, 0, 2, 1, 3)); // [N, h*w, C * patch_size * patch_size] + } // rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2) out = unpatchify(ctx, out, (H + pad_h) / patch_size, (W + pad_w) / patch_size, patch_size); // [N, C, H + pad_h, W + pad_w] @@ -1076,8 +1115,8 @@ namespace Flux { struct ggml_tensor* c_concat, struct ggml_tensor* y, struct ggml_tensor* guidance, - std::vector kontext_imgs = std::vector(), - std::vector skip_layers = std::vector()) { + std::vector ref_latents = {}, + std::vector skip_layers = std::vector()) { GGML_ASSERT(x->ne[3] == 1); struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, FLUX_GRAPH_SIZE, false); @@ -1088,9 +1127,7 @@ namespace Flux { if (c_concat != NULL) { c_concat = to_backend(c_concat); } - for (auto &img : kontext_imgs){ - img = to_backend(img); - } + if (flux_params.is_chroma) { const char* SD_CHROMA_ENABLE_GUIDANCE = getenv("SD_CHROMA_ENABLE_GUIDANCE"); bool disable_guidance = true; @@ -1131,10 +1168,11 @@ namespace Flux { if (flux_params.guidance_embed || flux_params.is_chroma) { guidance = to_backend(guidance); } - auto imgs = kontext_imgs; - imgs.insert(imgs.begin(), x); + for (int i = 0; i < ref_latents.size(); i++) { + ref_latents[i] = to_backend(ref_latents[i]); + } - pe_vec = flux.gen_pe(imgs, context, 2, flux_params.theta, flux_params.axes_dim); + pe_vec = flux.gen_pe(x->ne[1], x->ne[0], 2, x->ne[3], context->ne[1], ref_latents, flux_params.theta, flux_params.axes_dim); int pos_len = pe_vec.size() / flux_params.axes_dim_sum / 2; // LOG_DEBUG("pos_len %d", pos_len); auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, flux_params.axes_dim_sum / 2, pos_len); @@ -1144,7 +1182,7 @@ namespace Flux { set_backend_tensor_data(pe, pe_vec.data()); struct ggml_tensor* out = flux.forward(compute_ctx, - imgs, + x, timesteps, context, c_concat, @@ -1152,6 +1190,7 @@ namespace Flux { guidance, pe, precompute_arange, + ref_latents, skip_layers); ggml_build_forward_expand(gf, out); @@ -1166,17 +1205,17 @@ namespace Flux { struct ggml_tensor* c_concat, struct ggml_tensor* y, struct ggml_tensor* guidance, - std::vector kontext_imgs = std::vector(), - struct ggml_tensor** output = NULL, - struct ggml_context* output_ctx = NULL, - std::vector skip_layers = std::vector()) { + std::vector ref_latents = {}, + struct ggml_tensor** output = NULL, + struct ggml_context* output_ctx = NULL, + std::vector skip_layers = std::vector()) { // x: [N, in_channels, h, w] // timesteps: [N, ] // context: [N, max_position, hidden_size] // y: [N, adm_in_channels] or [1, adm_in_channels] // guidance: [N, ] auto get_graph = [&]() -> struct ggml_cgraph* { - return build_graph(x, timesteps, context, c_concat, y, guidance, kontext_imgs, skip_layers); + return build_graph(x, timesteps, context, c_concat, y, guidance, ref_latents, skip_layers); }; GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx); @@ -1216,7 +1255,7 @@ namespace Flux { struct ggml_tensor* out = NULL; int t0 = ggml_time_ms(); - compute(8, x, timesteps, context, NULL, y, guidance, std::vector(), &out, work_ctx); + compute(8, x, timesteps, context, NULL, y, guidance, {}, &out, work_ctx); int t1 = ggml_time_ms(); print_ggml_tensor(out); diff --git a/otherarch/sdcpp/main.cpp b/otherarch/sdcpp/main.cpp index db3365cb1..9ff38f60a 100644 --- a/otherarch/sdcpp/main.cpp +++ b/otherarch/sdcpp/main.cpp @@ -54,6 +54,7 @@ const char* modes_str[] = { "txt2img", "img2img", "img2vid", + "edit", "convert", }; @@ -61,6 +62,7 @@ enum SDMode { TXT2IMG, IMG2IMG, IMG2VID, + EDIT, CONVERT, MODE_COUNT }; @@ -86,8 +88,7 @@ struct SDParams { std::string input_path; std::string mask_path; std::string control_image_path; - - std::vector kontext_image_paths; + std::vector ref_image_paths; std::string prompt; std::string negative_prompt; @@ -153,6 +154,10 @@ void print_params(SDParams params) { printf(" init_img: %s\n", params.input_path.c_str()); printf(" mask_img: %s\n", params.mask_path.c_str()); printf(" control_image: %s\n", params.control_image_path.c_str()); + printf(" ref_images_paths:\n"); + for (auto& path : params.ref_image_paths) { + printf(" %s\n", path.c_str()); + }; printf(" clip on cpu: %s\n", params.clip_on_cpu ? "true" : "false"); printf(" controlnet cpu: %s\n", params.control_net_cpu ? "true" : "false"); printf(" vae decoder on cpu:%s\n", params.vae_on_cpu ? "true" : "false"); @@ -207,6 +212,7 @@ void print_usage(int argc, const char* argv[]) { printf(" -i, --init-img [IMAGE] path to the input image, required by img2img\n"); printf(" --mask [MASK] path to the mask image, required by img2img with mask\n"); printf(" --control-image [IMAGE] path to image condition, control net\n"); + printf(" -r, --ref_image [PATH] reference image for Flux Kontext models (can be used multiple times) \n"); printf(" -o, --output OUTPUT path to write result image to (default: ./output.png)\n"); printf(" -p, --prompt [PROMPT] the prompt to render\n"); printf(" -n, --negative-prompt PROMPT the negative prompt (default: \"\")\n"); @@ -242,9 +248,8 @@ void print_usage(int argc, const char* argv[]) { printf(" This might crash if it is not supported by the backend.\n"); printf(" --control-net-cpu keep controlnet in cpu (for low vram)\n"); printf(" --canny apply canny preprocessor (edge detection)\n"); - printf(" --color Colors the logging tags according to level\n"); + printf(" --color colors the logging tags according to level\n"); printf(" -v, --verbose print extra info\n"); - printf(" -ki, --kontext_img [PATH] Reference image for Flux Kontext models (can be used multiple times) \n"); } void parse_args(int argc, const char** argv, SDParams& params) { @@ -629,12 +634,12 @@ void parse_args(int argc, const char** argv, SDParams& params) { break; } params.skip_layer_end = std::stof(argv[i]); - } else if (arg == "-ki" || arg == "--kontext-img") { + } else if (arg == "-r" || arg == "--ref-image") { if (++i >= argc) { invalid_arg = true; break; } - params.kontext_image_paths.push_back(argv[i]); + params.ref_image_paths.push_back(argv[i]); } else { fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); print_usage(argc, argv); @@ -663,7 +668,13 @@ void parse_args(int argc, const char** argv, SDParams& params) { } if ((params.mode == IMG2IMG || params.mode == IMG2VID) && params.input_path.length() == 0) { - fprintf(stderr, "error: when using the img2img mode, the following arguments are required: init-img\n"); + fprintf(stderr, "error: when using the img2img/img2vid mode, the following arguments are required: init-img\n"); + print_usage(argc, argv); + exit(1); + } + + if (params.mode == EDIT && params.ref_image_paths.size() == 0) { + fprintf(stderr, "error: when using the edit mode, the following arguments are required: ref-image\n"); print_usage(argc, argv); exit(1); } @@ -827,43 +838,12 @@ int main(int argc, const char* argv[]) { fprintf(stderr, "SVD support is broken, do not use it!!!\n"); return 1; } - bool vae_decode_only = true; - - std::vector kontext_imgs; - for (auto& path : params.kontext_image_paths) { - vae_decode_only = false; - int c = 0; - int width = 0; - int height = 0; - uint8_t* image_buffer = stbi_load(path.c_str(), &width, &height, &c, 3); - if (image_buffer == NULL) { - fprintf(stderr, "load image from '%s' failed\n", path.c_str()); - return 1; - } - if (c < 3) { - fprintf(stderr, "the number of channels for the input image must be >= 3, but got %d channels\n", c); - free(image_buffer); - return 1; - } - if (width <= 0) { - fprintf(stderr, "error: the width of image must be greater than 0\n"); - free(image_buffer); - return 1; - } - if (height <= 0) { - fprintf(stderr, "error: the height of image must be greater than 0\n"); - free(image_buffer); - return 1; - } - kontext_imgs.push_back({(uint32_t)width, - (uint32_t)height, - 3, - image_buffer}); - } + bool vae_decode_only = true; uint8_t* input_image_buffer = NULL; uint8_t* control_image_buffer = NULL; uint8_t* mask_image_buffer = NULL; + std::vector ref_images; if (params.mode == IMG2IMG || params.mode == IMG2VID) { vae_decode_only = false; @@ -915,6 +895,37 @@ int main(int argc, const char* argv[]) { free(input_image_buffer); input_image_buffer = resized_image_buffer; } + } else if (params.mode == EDIT) { + vae_decode_only = false; + for (auto& path : params.ref_image_paths) { + int c = 0; + int width = 0; + int height = 0; + uint8_t* image_buffer = stbi_load(path.c_str(), &width, &height, &c, 3); + if (image_buffer == NULL) { + fprintf(stderr, "load image from '%s' failed\n", path.c_str()); + return 1; + } + if (c < 3) { + fprintf(stderr, "the number of channels for the input image must be >= 3, but got %d channels\n", c); + free(image_buffer); + return 1; + } + if (width <= 0) { + fprintf(stderr, "error: the width of image must be greater than 0\n"); + free(image_buffer); + return 1; + } + if (height <= 0) { + fprintf(stderr, "error: the height of image must be greater than 0\n"); + free(image_buffer); + return 1; + } + ref_images.push_back({(uint32_t)width, + (uint32_t)height, + 3, + image_buffer}); + } } sd_ctx_t* sd_ctx = new_sd_ctx(params.model_path.c_str(), @@ -1001,14 +1012,12 @@ int main(int argc, const char* argv[]) { params.style_ratio, params.normalize_input, params.input_id_images_path.c_str(), - kontext_imgs.data(), kontext_imgs.size(), params.skip_layers.data(), params.skip_layers.size(), params.slg_scale, params.skip_layer_start, - params.skip_layer_end, - nullptr); - } else { + params.skip_layer_end); + } else if (params.mode == IMG2IMG || params.mode == IMG2VID) { sd_image_t input_image = {(uint32_t)params.width, (uint32_t)params.height, 3, @@ -1072,14 +1081,38 @@ int main(int argc, const char* argv[]) { params.style_ratio, params.normalize_input, params.input_id_images_path.c_str(), - kontext_imgs.data(), kontext_imgs.size(), params.skip_layers.data(), params.skip_layers.size(), params.slg_scale, params.skip_layer_start, - params.skip_layer_end, - nullptr); + params.skip_layer_end); } + } else { // EDIT + results = edit(sd_ctx, + ref_images.data(), + ref_images.size(), + params.prompt.c_str(), + params.negative_prompt.c_str(), + params.clip_skip, + params.cfg_scale, + params.guidance, + params.eta, + params.width, + params.height, + params.sample_method, + params.sample_steps, + params.strength, + params.seed, + params.batch_count, + control_image, + params.control_strength, + params.style_ratio, + params.normalize_input, + params.skip_layers.data(), + params.skip_layers.size(), + params.slg_scale, + params.skip_layer_start, + params.skip_layer_end); } if (results == NULL) { @@ -1117,11 +1150,11 @@ int main(int argc, const char* argv[]) { std::string dummy_name, ext, lc_ext; bool is_jpg; - size_t last = params.output_path.find_last_of("."); + size_t last = params.output_path.find_last_of("."); size_t last_path = std::min(params.output_path.find_last_of("/"), params.output_path.find_last_of("\\")); - if (last != std::string::npos // filename has extension - && (last_path == std::string::npos || last > last_path)) { + if (last != std::string::npos // filename has extension + && (last_path == std::string::npos || last > last_path)) { dummy_name = params.output_path.substr(0, last); ext = lc_ext = params.output_path.substr(last); std::transform(ext.begin(), ext.end(), lc_ext.begin(), ::tolower); @@ -1129,7 +1162,7 @@ int main(int argc, const char* argv[]) { } else { dummy_name = params.output_path; ext = lc_ext = ""; - is_jpg = false; + is_jpg = false; } // appending ".png" to absent or unknown extension if (!is_jpg && lc_ext != ".png") { @@ -1141,7 +1174,7 @@ int main(int argc, const char* argv[]) { continue; } std::string final_image_path = i > 0 ? dummy_name + "_" + std::to_string(i + 1) + ext : dummy_name + ext; - if (is_jpg) { + if(is_jpg) { stbi_write_jpg(final_image_path.c_str(), results[i].width, results[i].height, results[i].channel, results[i].data, 90); printf("save result JPEG image to '%s'\n", final_image_path.c_str()); diff --git a/otherarch/sdcpp/sdtype_adapter.cpp b/otherarch/sdcpp/sdtype_adapter.cpp index 3991efe9a..4657d38e9 100644 --- a/otherarch/sdcpp/sdtype_adapter.cpp +++ b/otherarch/sdcpp/sdtype_adapter.cpp @@ -587,7 +587,47 @@ sd_generation_outputs sdtype_generate(const sd_generation_inputs inputs) extraimage_buffers.push_back(kcpp_base64_decode(extra_image_data[i])); input_extraimage_buffers.push_back(stbi_load_from_memory(extraimage_buffers[i].data(), extraimage_buffers[i].size(), &nx2, &ny2, &nc2, desiredchannels)); // Resize the image - int resok = stbir_resize_uint8(input_extraimage_buffers[i], nx2, ny2, 0, resized_extraimage_bufs[i].data(), img2imgW, img2imgH, 0, desiredchannels); + int desiredWidth = nx2; + int desiredHeight = ny2; + float aspect_ratio = static_cast(nx2) / ny2; + int maxsize = 1024; // no image can exceed this + int minsize = 256; + + if (desiredWidth > maxsize || desiredHeight > maxsize) { // Enforce maxsize first + if (aspect_ratio > 1.0f) { // wider than tall + desiredWidth = maxsize; + desiredHeight = static_cast(maxsize / aspect_ratio); + } else { // taller than wide or square + desiredHeight = maxsize; + desiredWidth = static_cast(maxsize * aspect_ratio); + } + } + + if (desiredWidth < minsize || desiredHeight < minsize) { // Enforce minsize only if it won't exceed maxsize + if (aspect_ratio > 1.0f) { // wider than tall + // Try to scale width up to max of (minsize, maxsize) + desiredWidth = std::min(maxsize, std::max(minsize, desiredWidth)); + desiredHeight = static_cast(desiredWidth / aspect_ratio); + if (desiredHeight > maxsize) { // If height now exceeds maxsize, clamp based on height instead + desiredHeight = maxsize; + desiredWidth = static_cast(maxsize * aspect_ratio); + } + } else { + // Taller than wide or square + desiredHeight = std::min(maxsize, std::max(minsize, desiredHeight)); + desiredWidth = static_cast(desiredHeight * aspect_ratio); + if (desiredWidth > maxsize) { + desiredWidth = maxsize; + desiredHeight = static_cast(maxsize / aspect_ratio); + } + } + } + + if(!sd_is_quiet && sddebugmode==1) + { + printf("Resize Extraimg: %dx%d to %dx%d\n",nx2,ny2,desiredWidth,desiredHeight); + } + int resok = stbir_resize_uint8(input_extraimage_buffers[i], nx2, ny2, 0, resized_extraimage_bufs[i].data(), desiredWidth, desiredHeight, 0, desiredchannels); if (!resok) { printf("\nKCPP SD: resize extra image failed!\n"); output.data = ""; @@ -595,8 +635,8 @@ sd_generation_outputs sdtype_generate(const sd_generation_inputs inputs) return output; } sd_image_t extraimage_reference; - extraimage_reference.width = img2imgW; - extraimage_reference.height = img2imgH; + extraimage_reference.width = desiredWidth; + extraimage_reference.height = desiredHeight; extraimage_reference.channel = desiredchannels; extraimage_reference.data = resized_extraimage_bufs[i].data(); extraimage_references.push_back(extraimage_reference); @@ -716,6 +756,10 @@ sd_generation_outputs sdtype_generate(const sd_generation_inputs inputs) } // Resize the image + if(!sd_is_quiet && sddebugmode==1) + { + printf("Resize Img2Img: %dx%d to %dx%d\n",nx,ny,img2imgW,img2imgH); + } int resok = stbir_resize_uint8(input_image_buffer, nx, ny, 0, resized_image_buf.data(), img2imgW, img2imgH, 0, img2imgC); if (!resok) { printf("\nKCPP SD: resize image failed!\n"); @@ -735,6 +779,10 @@ sd_generation_outputs sdtype_generate(const sd_generation_inputs inputs) image_mask_buffer = kcpp_base64_decode(img2img_mask); input_mask_buffer = stbi_load_from_memory(image_mask_buffer.data(), image_mask_buffer.size(), &nx2, &ny2, &nc2, 1); // Resize the image + if(!sd_is_quiet && sddebugmode==1) + { + printf("Resize Mask: %dx%d to %dx%d\n",nx2,ny2,img2imgW,img2imgH); + } int resok = stbir_resize_uint8(input_mask_buffer, nx2, ny2, 0, resized_mask_buf.data(), img2imgW, img2imgH, 0, 1); if (!resok) { printf("\nKCPP SD: resize image failed!\n"); diff --git a/otherarch/sdcpp/stable-diffusion.cpp b/otherarch/sdcpp/stable-diffusion.cpp index 6b9313dae..5e0a79d01 100644 --- a/otherarch/sdcpp/stable-diffusion.cpp +++ b/otherarch/sdcpp/stable-diffusion.cpp @@ -678,7 +678,7 @@ public: int64_t t0 = ggml_time_ms(); struct ggml_tensor* out = ggml_dup_tensor(work_ctx, x_t); - diffusion_model->compute(n_threads, x_t, timesteps, c, concat, NULL, NULL, -1, {}, 0.f, std::vector(), &out); + diffusion_model->compute(n_threads, x_t, timesteps, c, concat, NULL, NULL, {}, -1, {}, 0.f, &out); diffusion_model->free_compute_buffer(); double result = 0.f; @@ -892,12 +892,12 @@ public: const std::vector& sigmas, int start_merge_step, SDCondition id_cond, - std::vector skip_layers = {}, - float slg_scale = 0, - float skip_layer_start = 0.01, - float skip_layer_end = 0.2, - std::vector kontext_imgs = std::vector(), - ggml_tensor* noise_mask = NULL) { + std::vector ref_latents = {}, + std::vector skip_layers = {}, + float slg_scale = 0, + float skip_layer_start = 0.01, + float skip_layer_end = 0.2, + ggml_tensor* noise_mask = nullptr) { LOG_DEBUG("Sample"); struct ggml_init_params params; size_t data_size = ggml_row_size(init_latent->type, init_latent->ne[0]); @@ -980,10 +980,10 @@ public: cond.c_concat, cond.c_vector, guidance_tensor, + ref_latents, -1, controls, control_strength, - kontext_imgs, &out_cond); } else { diffusion_model->compute(n_threads, @@ -993,10 +993,10 @@ public: cond.c_concat, id_cond.c_vector, guidance_tensor, + ref_latents, -1, controls, control_strength, - kontext_imgs, &out_cond); } @@ -1014,10 +1014,10 @@ public: uncond.c_concat, uncond.c_vector, guidance_tensor, + ref_latents, -1, controls, control_strength, - kontext_imgs, &out_uncond); negative_data = (float*)out_uncond->data; } @@ -1035,10 +1035,10 @@ public: cond.c_concat, cond.c_vector, guidance_tensor, + ref_latents, -1, controls, control_strength, - kontext_imgs, &out_skip, NULL, skip_layers); @@ -1416,12 +1416,12 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, float style_ratio, bool normalize_input, std::string input_id_images_path, - std::vector kontext_imgs = std::vector(), - std::vector skip_layers = {}, - float slg_scale = 0, - float skip_layer_start = 0.01, - float skip_layer_end = 0.2, - ggml_tensor* masked_image = NULL, + std::vector ref_latents, + std::vector skip_layers = {}, + float slg_scale = 0, + float skip_layer_start = 0.01, + float skip_layer_end = 0.2, + ggml_tensor* masked_image = NULL, const std::vector photomaker_references = std::vector()) { if (seed < 0) { // Generally, when using the provided command line, the seed is always >0. @@ -1712,11 +1712,11 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, sigmas, start_merge_step, id_cond, + ref_latents, skip_layers, slg_scale, skip_layer_start, skip_layer_end, - kontext_imgs, noise_mask); // struct ggml_tensor* x_0 = load_tensor_from_file(ctx, "samples_ddim.bin");