mtmd: fix longest_edge ignoring min/max pixels (#26638)

* mtmd: fix longest_edge ignoring min/max pixels

* nits

diff --git a/tools/mtmd/clip-model.h b/tools/mtmd/clip-model.h
index 101f49cd1849d1af1da705e7ef7377e3c2bfe763..7db01b576bd7659fb9c089492ac75d395b5197e4 100644 (file)
--- a/tools/mtmd/clip-model.h
+++ b/tools/mtmd/clip-model.h
@@ -170,6 +170,17 @@ struct clip_hparams {
         warmup_image_size = static_cast<int>(std::sqrt(image_max_pixels));
     }
 
+    // used by longest_edge preprocessor (no model-specific value for min/max tokens)
+    void set_limit_image_tokens() {
+        const int patch_area = patch_size * patch_size * n_merge * n_merge;
+        if (custom_image_min_tokens > 0) {
+            image_min_pixels = custom_image_min_tokens * patch_area;
+        }
+        if (custom_image_max_tokens > 0) {
+            image_max_pixels = custom_image_max_tokens * patch_area;
+        }
+    }
+
     void set_warmup_n_tokens(int n_tokens) {
         int n_tok_per_side = static_cast<int>(std::sqrt(n_tokens));
         GGML_ASSERT(n_tok_per_side * n_tok_per_side == n_tokens && "n_tokens must be n*n");

diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
index b1360fd7d30924dacac49d3572d4c8b9ff136ccd..3b610562987094a0752012c45f4acff16985d143 100644 (file)
--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -1434,6 +1434,7 @@ struct clip_model_loader {
                         // use default llava-uhd preprocessing params
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
                         get_u32(KEY_PREPROC_IMAGE_SIZE, hparams.image_longest_edge, false);
+                        hparams.set_limit_image_tokens();
                     } break;
                 case PROJECTOR_TYPE_LFM2:
                     {
@@ -1471,6 +1472,7 @@ struct clip_model_loader {
                         get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
                         hparams.image_longest_edge = hparams.image_size;
                         get_u32(KEY_PREPROC_IMAGE_SIZE, hparams.image_longest_edge, false);
+                        hparams.set_limit_image_tokens();
                         hparams.set_warmup_n_tokens(256); // avoid OOM on warmup
                     } break;
                 case PROJECTOR_TYPE_DOTS_OCR:
@@ -1595,6 +1597,7 @@ struct clip_model_loader {
                         if (hparams.image_longest_edge == 0) {
                             hparams.image_longest_edge = 3024;
                         }
+                        // note: the step3vl preprocessor slices based on a fixed window grid, so it does not support custom min/max image tokens
                         hparams.warmup_image_size = hparams.image_size;
                     } break;
                 case PROJECTOR_TYPE_YOUTUVL:

diff --git a/tools/mtmd/mtmd-image.cpp b/tools/mtmd/mtmd-image.cpp
index 10cfe52f56f773cb0aa441a820db4e732d3daa52..073d83d4534454dd6ebf4e70806f351af78d2e14 100644 (file)
--- a/tools/mtmd/mtmd-image.cpp
+++ b/tools/mtmd/mtmd-image.cpp
@@ -139,50 +139,46 @@ struct img_tool {
         }
     }
 
-    // calculate the size of the **resized** image, while preserving the aspect ratio
-    // the calculated size will be aligned to the nearest multiple of align_size
-    // if H or W size is larger than longest_edge, it will be resized to longest_edge
-    static clip_image_size calc_size_preserved_ratio(const clip_image_size & inp_size, const int align_size, const int longest_edge) {
-        GGML_ASSERT(align_size > 0);
-        if (inp_size.width <= 0 || inp_size.height <= 0 || longest_edge <= 0) {
-            return {0, 0};
-        }
-
-        float scale = std::min(static_cast<float>(longest_edge) / inp_size.width,
-                               static_cast<float>(longest_edge) / inp_size.height);
-
-        float target_width_f  = static_cast<float>(inp_size.width)  * scale;
-        float target_height_f = static_cast<float>(inp_size.height) * scale;
-
-        auto ceil_by_factor = [f = align_size](float x) { return static_cast<int>(std::ceil(x / static_cast<float>(f))) * f; };
-        int aligned_width  = ceil_by_factor(target_width_f);
-        int aligned_height = ceil_by_factor(target_height_f);
-
-        return {aligned_width, aligned_height};
-    }
+    struct calc_size_opt {
+        int align_size = 1;
+        int min_pixels = 0;   // 0 = disabled
+        int max_pixels = 0;   // 0 = disabled
+        // applied before min/max_pixels, so min_pixels can push an edge back above longest_edge
+        int longest_edge = 0; // 0 = disabled
+    };
 
-    // calculate the size of the **resized** image, while preserving the aspect ratio
-    // the calculated size will have min_pixels <= W*H <= max_pixels
-    // this is referred as "smart_resize" in transformers code
-    static clip_image_size calc_size_preserved_ratio(const clip_image_size & inp_size, const int align_size, const int min_pixels, const int max_pixels) {
-        GGML_ASSERT(align_size > 0);
+    // calculate the size of the **resized** image, while preserving the aspect ratio and
+    // aligning to the nearest multiple of align_size ("smart_resize" in transformers code)
+    static clip_image_size calc_size_preserved_ratio(const clip_image_size & inp_size, const calc_size_opt & opts) {
+        GGML_ASSERT(opts.align_size > 0);
         const int width  = inp_size.width;
         const int height = inp_size.height;
+        if (width <= 0 || height <= 0) {
+            return {0, 0};
+        }
 
-        auto round_by_factor = [f = align_size](float x) { return static_cast<int>(std::round(x / static_cast<float>(f))) * f; };
-        auto ceil_by_factor  = [f = align_size](float x) { return static_cast<int>(std::ceil(x / static_cast<float>(f))) * f; };
-        auto floor_by_factor = [f = align_size](float x) { return static_cast<int>(std::floor(x / static_cast<float>(f))) * f; };
+        auto round_by_factor = [f = opts.align_size](float x) { return static_cast<int>(std::round(x / static_cast<float>(f))) * f; };
+        auto ceil_by_factor  = [f = opts.align_size](float x) { return static_cast<int>(std::ceil(x / static_cast<float>(f))) * f; };
+        auto floor_by_factor = [f = opts.align_size](float x) { return static_cast<int>(std::floor(x / static_cast<float>(f))) * f; };
 
-        // always align up first
-        int h_bar = std::max(align_size, round_by_factor(height));
-        int w_bar = std::max(align_size, round_by_factor(width));
+        int w_bar, h_bar;
+        if (opts.longest_edge > 0) {
+            const float scale = std::min(static_cast<float>(opts.longest_edge) / width,
+                                          static_cast<float>(opts.longest_edge) / height);
+            w_bar = ceil_by_factor(width  * scale);
+            h_bar = ceil_by_factor(height * scale);
+        } else {
+            // always align up first
+            w_bar = std::max(opts.align_size, round_by_factor(width));
+            h_bar = std::max(opts.align_size, round_by_factor(height));
+        }
 
-        if (h_bar * w_bar > max_pixels) {
-            const auto beta = std::sqrt(static_cast<float>(height * width) / max_pixels);
-            h_bar = std::max(align_size, floor_by_factor(height / beta));
-            w_bar = std::max(align_size, floor_by_factor(width  / beta));
-        } else if (h_bar * w_bar < min_pixels) {
-            const auto beta = std::sqrt(static_cast<float>(min_pixels) / (height * width));
+        if (opts.max_pixels > 0 && h_bar * w_bar > opts.max_pixels) {
+            const auto beta = std::sqrt(static_cast<float>(height) * width / opts.max_pixels);
+            h_bar = std::max(opts.align_size, floor_by_factor(height / beta));
+            w_bar = std::max(opts.align_size, floor_by_factor(width  / beta));
+        } else if (opts.min_pixels > 0 && h_bar * w_bar < opts.min_pixels) {
+            const auto beta = std::sqrt(static_cast<float>(opts.min_pixels) / (static_cast<float>(height) * width));
             h_bar = ceil_by_factor(height * beta);
             w_bar = ceil_by_factor(width * beta);
         }
@@ -937,9 +933,12 @@ mtmd_image_preproc_out mtmd_image_preprocessor_dyn_size::preprocess(const clip_i
     const int cur_merge = hparams.n_merge;
     const clip_image_size target_size = img_tool::calc_size_preserved_ratio(
         original_size,
-        hparams.patch_size * cur_merge,
-        hparams.image_min_pixels,
-        hparams.image_max_pixels);
+        {
+            /* align_size   */ hparams.patch_size * cur_merge,
+            /* min_pixels   */ hparams.image_min_pixels,
+            /* max_pixels   */ hparams.image_max_pixels,
+            /* longest_edge */ 0,
+        });
     img_tool::resize(img, resized_image, target_size,
                         hparams.image_resize_algo,
                         hparams.image_resize_pad,
@@ -961,8 +960,12 @@ mtmd_image_preproc_out mtmd_image_preprocessor_longest_edge::preprocess(const cl
     const int cur_merge = hparams.n_merge == 0 ? 1 : hparams.n_merge;
     const clip_image_size target_size = img_tool::calc_size_preserved_ratio(
         original_size,
-        hparams.patch_size * cur_merge,
-        hparams.image_longest_edge);
+        {
+            /* align_size   */ hparams.patch_size * cur_merge,
+            /* min_pixels   */ std::max(0, hparams.image_min_pixels),
+            /* max_pixels   */ std::max(0, hparams.image_max_pixels),
+            /* longest_edge */ hparams.image_longest_edge,
+        });
     img_tool::resize(img, resized_image, target_size,
                         hparams.image_resize_algo,
                         hparams.image_resize_pad,
@@ -1000,8 +1003,8 @@ mtmd_image_preprocessor_llava_uhd::slice_instructions mtmd_image_preprocessor_lf
     mtmd_image_preprocessor_llava_uhd::slice_instructions inst;
     const int align_size = hparams.patch_size * hparams.n_merge;
     inst.overview_size = img_tool::calc_size_preserved_ratio(
-                            original_size, align_size,
-                            hparams.image_min_pixels, hparams.image_max_pixels);
+                            original_size,
+                            { align_size, hparams.image_min_pixels, hparams.image_max_pixels, 0 });
     // tile if either dimension exceeds tile_size with tolerance
     const bool needs_tiling = original_size.width > tile_size * max_pixels_tolerance || original_size.height > tile_size * max_pixels_tolerance;
 
@@ -1109,7 +1112,8 @@ mtmd_image_preproc_out mtmd_image_preprocessor_idefics3::preprocess(const clip_i
     // CITE: https://github.com/huggingface/transformers/blob/main/src/transformers/models/idefics3/image_processing_idefics3.py#L737
     const clip_image_size original_size = img.get_size();
     const clip_image_size refined_size = img_tool::calc_size_preserved_ratio(
-        original_size, hparams.image_size, hparams.image_longest_edge);
+        original_size,
+        { hparams.image_size, std::max(0, hparams.image_min_pixels), std::max(0, hparams.image_max_pixels), hparams.image_longest_edge });
     // LOG_INF("%s: original size: %d x %d, refined size: %d x %d\n",
     //         __func__, original_size.width, original_size.height,
     //         refined_size.width, refined_size.height);