Feat: Temporal tile custom size with overlap

examples/cli/README.md

@@ -107,6 +107,8 @@ Generation Options:
   --extra-sample-args <string>             extra sampler/scheduler args, key=value list. lcm supports noise_clip_std,
                                            noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift,
                                            stretch, terminal; euler_ge supports gamma
+  --extra-tiling-args <string>             extra VAE tiling args, key=value list. LTX video VAE supports
+                                           temporal_tile_frames (default: 4), temporal_tile_overlap (default: 1)
   -H, --height <int>                       image height, in pixel space (default: 512)
   -W, --width <int>                        image width, in pixel space (default: 512)
   --steps <int>                            number of sample steps (default: 20)

examples/common/common.cpp

@@ -835,6 +835,10 @@ ArgOptions SDGenerationParams::get_options() {
          "--extra-sample-args",
          "extra sampler/scheduler args, key=value list. lcm supports noise_clip_std, noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift, stretch, terminal; euler_ge supports gamma",
          &extra_sample_args},
+        {"",
+         "--extra-tiling-args",
+         "extra VAE tiling args, key=value list. LTX video VAE supports temporal_tile_frames (default: 4), temporal_tile_overlap (default: 1)",
+         &extra_tiling_args},
     };
 
     options.int_options = {
@@ -1780,6 +1784,9 @@ bool SDGenerationParams::from_json_str(
         if (tiling_json.contains("rel_size_y") && tiling_json["rel_size_y"].is_number()) {
             vae_tiling_params.rel_size_y = tiling_json["rel_size_y"];
         }
+        if (tiling_json.contains("extra_tiling_args") && tiling_json["extra_tiling_args"].is_string()) {
+            extra_tiling_args = tiling_json["extra_tiling_args"].get<std::string>();
+        }
     }
 
     if (!parse_lora_json_field(j, lora_path_resolver, lora_map, high_noise_lora_map)) {
@@ -2002,6 +2009,8 @@ bool SDGenerationParams::initialize_cache_params() {
 }
 
 bool SDGenerationParams::resolve(const std::string& lora_model_dir, const std::string& hires_upscalers_dir, bool strict) {
+    vae_tiling_params.extra_tiling_args = extra_tiling_args.empty() ? nullptr : extra_tiling_args.c_str();
+
     if (high_noise_sample_params.sample_steps <= 0) {
         high_noise_sample_params.sample_steps = -1;
     }
@@ -2188,6 +2197,7 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
     sample_params.custom_sigmas_count                 = static_cast<int>(custom_sigmas.size());
     sample_params.extra_sample_args                   = extra_sample_args.empty() ? nullptr : extra_sample_args.c_str();
     high_noise_sample_params.extra_sample_args        = high_noise_extra_sample_args.empty() ? nullptr : high_noise_extra_sample_args.c_str();
+    vae_tiling_params.extra_tiling_args               = extra_tiling_args.empty() ? nullptr : extra_tiling_args.c_str();
     cache_params.scm_mask                             = scm_mask.empty() ? nullptr : scm_mask.c_str();
 
     sd_pm_params_t pm_params = {
@@ -2261,6 +2271,7 @@ sd_vid_gen_params_t SDGenerationParams::to_sd_vid_gen_params_t() {
     sample_params.custom_sigmas_count                 = static_cast<int>(custom_sigmas.size());
     sample_params.extra_sample_args                   = extra_sample_args.empty() ? nullptr : extra_sample_args.c_str();
     high_noise_sample_params.extra_sample_args        = high_noise_extra_sample_args.empty() ? nullptr : high_noise_extra_sample_args.c_str();
+    vae_tiling_params.extra_tiling_args               = extra_tiling_args.empty() ? nullptr : extra_tiling_args.c_str();
     cache_params.scm_mask                             = scm_mask.empty() ? nullptr : scm_mask.c_str();
 
     params.loras                     = lora_vec.empty() ? nullptr : lora_vec.data();
@@ -2386,7 +2397,8 @@ std::string SDGenerationParams::to_string() const {
         << vae_tiling_params.tile_size_y << ", "
         << vae_tiling_params.target_overlap << ", "
         << vae_tiling_params.rel_size_x << ", "
-        << vae_tiling_params.rel_size_y << " },\n"
+        << vae_tiling_params.rel_size_y << ", "
+        << "\"" << extra_tiling_args << "\" },\n"
         << "}";
     return oss.str();
 }
@@ -2565,14 +2577,18 @@ std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
         };
     }
 
-    if (gen_params.vae_tiling_params.enabled) {
+    if (gen_params.vae_tiling_params.enabled ||
+        gen_params.vae_tiling_params.temporal_tiling ||
+        !gen_params.extra_tiling_args.empty()) {
         root["vae_tiling"] = {
             {"enabled", gen_params.vae_tiling_params.enabled},
+            {"temporal_tiling", gen_params.vae_tiling_params.temporal_tiling},
             {"tile_size_x", gen_params.vae_tiling_params.tile_size_x},
             {"tile_size_y", gen_params.vae_tiling_params.tile_size_y},
             {"target_overlap", gen_params.vae_tiling_params.target_overlap},
             {"rel_size_x", gen_params.vae_tiling_params.rel_size_x},
             {"rel_size_y", gen_params.vae_tiling_params.rel_size_y},
+            {"extra_tiling_args", gen_params.extra_tiling_args},
         };
     }
 

examples/common/common.h

@@ -189,7 +189,8 @@ struct SDGenerationParams {
     int video_frames                     = 1;
     int fps                              = 16;
     float vace_strength                  = 1.f;
-    sd_tiling_params_t vae_tiling_params = {false, false, 0, 0, 0.5f, 0.0f, 0.0f};
+    sd_tiling_params_t vae_tiling_params = {false, false, 0, 0, 0.5f, 0.0f, 0.0f, nullptr};
+    std::string extra_tiling_args;
 
     std::string pm_id_images_dir;
     std::string pm_id_embed_path;

examples/server/README.md

@@ -209,6 +209,8 @@ Default Generation Options:
   --extra-sample-args <string>             extra sampler/scheduler args, key=value list. lcm supports noise_clip_std,
                                            noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift,
                                            stretch, terminal; euler_ge supports gamma
+  --extra-tiling-args <string>             extra VAE tiling args, key=value list. LTX video VAE supports
+                                           temporal_tile_frames (default: 4), temporal_tile_overlap (default: 1)
   -H, --height <int>                       image height, in pixel space (default: 512)
   -W, --width <int>                        image width, in pixel space (default: 512)
   --steps <int>                            number of sample steps (default: 20)
@@ -264,6 +266,7 @@ Default Generation Options:
   --disable-auto-resize-ref-image          disable auto resize of ref images
   --disable-image-metadata                 do not embed generation metadata on image files
   --vae-tiling                             process vae in tiles to reduce memory usage
+  --temporal-tiling                        enable temporal tiling for LTX video VAE decode
   --hires                                  enable highres fix
   -s, --seed                               RNG seed (default: 42, use random seed for < 0)
   --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m,

examples/server/api.md

@@ -504,18 +504,22 @@ Shared default fields used by both `img_gen` and `vid_gen`:
 | `sample_params.guidance.slg.scale` | `number` |
 | `vae_tiling_params` | `object` |
 | `vae_tiling_params.enabled` | `boolean` |
+| `vae_tiling_params.temporal_tiling` | `boolean` |
 | `vae_tiling_params.tile_size_x` | `integer` |
 | `vae_tiling_params.tile_size_y` | `integer` |
 | `vae_tiling_params.target_overlap` | `number` |
 | `vae_tiling_params.rel_size_x` | `number` |
 | `vae_tiling_params.rel_size_y` | `number` |
+| `vae_tiling_params.extra_tiling_args` | `string` |
 | `cache_mode` | `string` |
 | `cache_option` | `string` |
 | `scm_mask` | `string` |
 | `scm_policy_dynamic` | `boolean` |
 | `output_format` | `string` |
 | `output_compression` | `integer` |
 
+`vae_tiling_params.extra_tiling_args` accepts a key=value list. For LTX video VAE temporal tiling, `temporal_tile_frames` defaults to `4` and `temporal_tile_overlap` defaults to `1`.
+
 `img_gen`-specific default fields:
 
 | Field | Type |
@@ -692,11 +696,13 @@ Example:
 
   "vae_tiling_params": {
     "enabled": false,
+    "temporal_tiling": false,
     "tile_size_x": 0,
     "tile_size_y": 0,
     "target_overlap": 0.5,
     "rel_size_x": 0.0,
-    "rel_size_y": 0.0
+    "rel_size_y": 0.0,
+    "extra_tiling_args": ""
   },
 
   "cache_mode": "disabled",
@@ -804,6 +810,14 @@ Other native fields:
 | `hires.custom_sigmas` | `array<number>` |
 | `hires.upscale_tile_size` | `integer` |
 | `vae_tiling_params` | `object` |
+| `vae_tiling_params.enabled` | `boolean` |
+| `vae_tiling_params.temporal_tiling` | `boolean` |
+| `vae_tiling_params.tile_size_x` | `integer` |
+| `vae_tiling_params.tile_size_y` | `integer` |
+| `vae_tiling_params.target_overlap` | `number` |
+| `vae_tiling_params.rel_size_x` | `number` |
+| `vae_tiling_params.rel_size_y` | `number` |
+| `vae_tiling_params.extra_tiling_args` | `string` |
 | `cache_mode` | `string` |
 | `cache_option` | `string` |
 | `scm_mask` | `string` |
@@ -1012,11 +1026,13 @@ Example:
 
   "vae_tiling_params": {
     "enabled": false,
+    "temporal_tiling": false,
     "tile_size_x": 0,
     "tile_size_y": 0,
     "target_overlap": 0.5,
     "rel_size_x": 0.0,
-    "rel_size_y": 0.0
+    "rel_size_y": 0.0,
+    "extra_tiling_args": ""
   },
 
   "cache_mode": "disabled",
@@ -1134,6 +1150,14 @@ Other native fields:
 | Field | Type |
 | --- | --- |
 | `vae_tiling_params` | `object` |
+| `vae_tiling_params.enabled` | `boolean` |
+| `vae_tiling_params.temporal_tiling` | `boolean` |
+| `vae_tiling_params.tile_size_x` | `integer` |
+| `vae_tiling_params.tile_size_y` | `integer` |
+| `vae_tiling_params.target_overlap` | `number` |
+| `vae_tiling_params.rel_size_x` | `number` |
+| `vae_tiling_params.rel_size_y` | `number` |
+| `vae_tiling_params.extra_tiling_args` | `string` |
 | `cache_mode` | `string` |
 | `cache_option` | `string` |
 | `scm_mask` | `string` |

examples/server/routes_sdcpp.cpp

@@ -56,11 +56,13 @@ static const char* capability_sample_method_name(enum sample_method_t sample_met
 static json make_vae_tiling_json(const sd_tiling_params_t& params) {
     return {
         {"enabled", params.enabled},
+        {"temporal_tiling", params.temporal_tiling},
         {"tile_size_x", params.tile_size_x},
         {"tile_size_y", params.tile_size_y},
         {"target_overlap", params.target_overlap},
         {"rel_size_x", params.rel_size_x},
         {"rel_size_y", params.rel_size_y},
+        {"extra_tiling_args", params.extra_tiling_args ? params.extra_tiling_args : ""},
     };
 }
 

include/stable-diffusion.h

@@ -160,6 +160,7 @@ typedef struct {
     float target_overlap;
     float rel_size_x;
     float rel_size_y;
+    const char* extra_tiling_args;
 } sd_tiling_params_t;
 
 typedef struct {

src/ggml_extend.hpp

@@ -3172,7 +3172,7 @@ class Conv2d_grouped : public UnaryBlock {
     void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map, const std::string prefix = "") override {
         this->prefix         = prefix;
         enum ggml_type wtype = GGML_TYPE_F16;
-        params["weight"] = ggml_new_tensor_4d(ctx, wtype, kernel_size.second, kernel_size.first, in_channels / groups, out_channels);
+        params["weight"]     = ggml_new_tensor_4d(ctx, wtype, kernel_size.second, kernel_size.first, in_channels / groups, out_channels);
         if (bias) {
             enum ggml_type wtype = GGML_TYPE_F32;
             params["bias"]       = ggml_new_tensor_1d(ctx, wtype, out_channels);