feat: add Flux.2 VAE proj matrix for previews (#1017)

Author: stduhpf

latent-preview.h

@@ -91,6 +91,41 @@ const float flux_latent_rgb_proj[16][3] = {
     {-0.111849f, -0.055589f, -0.032361f}};
 float flux_latent_rgb_bias[3] = {0.024600f, -0.006937f, -0.008089f};
 
+const float flux2_latent_rgb_proj[32][3] = {
+    {0.000736f, -0.008385f, -0.019710f},
+    {-0.001352f, -0.016392f, 0.020693f},
+    {-0.006376f, 0.002428f, 0.036736f},
+    {0.039384f, 0.074167f, 0.119789f},
+    {0.007464f, -0.005705f, -0.004734f},
+    {-0.004086f, 0.005287f, -0.000409f},
+    {-0.032835f, 0.050802f, -0.028120f},
+    {-0.003158f, -0.000835f, 0.000406f},
+    {-0.112840f, -0.084337f, -0.023083f},
+    {0.001462f, -0.006656f, 0.000549f},
+    {-0.009980f, -0.007480f, 0.009702f},
+    {0.032540f, 0.000214f, -0.061388f},
+    {0.011023f, 0.000694f, 0.007143f},
+    {-0.001468f, -0.006723f, -0.001678f},
+    {-0.005921f, -0.010320f, -0.003907f},
+    {-0.028434f, 0.027584f, 0.018457f},
+    {0.014349f, 0.011523f, 0.000441f},
+    {0.009874f, 0.003081f, 0.001507f},
+    {0.002218f, 0.005712f, 0.001563f},
+    {0.053010f, -0.019844f, 0.008683f},
+    {-0.002507f, 0.005384f, 0.000938f},
+    {-0.002177f, -0.011366f, 0.003559f},
+    {-0.000261f, 0.015121f, -0.003240f},
+    {-0.003944f, -0.002083f, 0.005043f},
+    {-0.009138f, 0.011336f, 0.003781f},
+    {0.011429f, 0.003985f, -0.003855f},
+    {0.010518f, -0.005586f, 0.010131f},
+    {0.007883f, 0.002912f, -0.001473f},
+    {-0.003318f, -0.003160f, 0.003684f},
+    {-0.034560f, -0.008740f, 0.012996f},
+    {0.000166f, 0.001079f, -0.012153f},
+    {0.017772f, 0.000937f, -0.011953f}};
+float flux2_latent_rgb_bias[3] = {-0.028738f, -0.098463f, -0.107619f};
+
 // This one was taken straight from
 // https://github.com/Stability-AI/sd3.5/blob/8565799a3b41eb0c7ba976d18375f0f753f56402/sd3_impls.py#L288-L303
 // (MiT Licence)
@@ -128,16 +163,42 @@ const float sd_latent_rgb_proj[4][3] = {
     {-0.178022f, -0.200862f, -0.678514f}};
 float sd_latent_rgb_bias[3] = {-0.017478f, -0.055834f, -0.105825f};
 
-void preview_latent_video(uint8_t* buffer, struct ggml_tensor* latents, const float (*latent_rgb_proj)[3], const float latent_rgb_bias[3], int width, int height, int frames, int dim) {
+void preview_latent_video(uint8_t* buffer, struct ggml_tensor* latents, const float (*latent_rgb_proj)[3], const float latent_rgb_bias[3], int patch_size) {
     size_t buffer_head = 0;
+
+    uint32_t latent_width  = latents->ne[0];
+    uint32_t latent_height = latents->ne[1];
+    uint32_t dim           = latents->ne[ggml_n_dims(latents) - 1];
+    uint32_t frames        = 1;
+    if (ggml_n_dims(latents) == 4) {
+        frames = latents->ne[2];
+    }
+
+    uint32_t rgb_width  = latent_width * patch_size;
+    uint32_t rgb_height = latent_height * patch_size;
+
+    uint32_t unpatched_dim = dim / (patch_size * patch_size);
+
     for (int k = 0; k < frames; k++) {
-        for (int j = 0; j < height; j++) {
-            for (int i = 0; i < width; i++) {
-                size_t latent_id = (i * latents->nb[0] + j * latents->nb[1] + k * latents->nb[2]);
+        for (int rgb_x = 0; rgb_x < rgb_width; rgb_x++) {
+            for (int rgb_y = 0; rgb_y < rgb_height; rgb_y++) {
+                int latent_x = rgb_x / patch_size;
+                int latent_y = rgb_y / patch_size;
+
+                int channel_offset = 0;
+                if (patch_size > 1) {
+                    channel_offset = ((rgb_y % patch_size) * patch_size + (rgb_x % patch_size));
+                }
+
+                size_t latent_id = (latent_x * latents->nb[0] + latent_y * latents->nb[1] + k * latents->nb[2]);
+
+                // should be incremented by 1 for each pixel
+                size_t pixel_id = k * rgb_width * rgb_height + rgb_y * rgb_width + rgb_x;
+
                 float r = 0, g = 0, b = 0;
                 if (latent_rgb_proj != nullptr) {
-                    for (int d = 0; d < dim; d++) {
-                        float value = *(float*)((char*)latents->data + latent_id + d * latents->nb[ggml_n_dims(latents) - 1]);
+                    for (int d = 0; d < unpatched_dim; d++) {
+                        float value = *(float*)((char*)latents->data + latent_id + (d * patch_size * patch_size + channel_offset) * latents->nb[ggml_n_dims(latents) - 1]);
                         r += value * latent_rgb_proj[d][0];
                         g += value * latent_rgb_proj[d][1];
                         b += value * latent_rgb_proj[d][2];
@@ -164,9 +225,9 @@ void preview_latent_video(uint8_t* buffer, struct ggml_tensor* latents, const fl
                 g = g >= 0 ? g <= 1 ? g : 1 : 0;
                 b = b >= 0 ? b <= 1 ? b : 1 : 0;
 
-                buffer[buffer_head++] = (uint8_t)(r * 255);
-                buffer[buffer_head++] = (uint8_t)(g * 255);
-                buffer[buffer_head++] = (uint8_t)(b * 255);
+                buffer[pixel_id * 3 + 0] = (uint8_t)(r * 255);
+                buffer[pixel_id * 3 + 1] = (uint8_t)(g * 255);
+                buffer[pixel_id * 3 + 2] = (uint8_t)(b * 255);
             }
         }
     }

stable-diffusion.cpp

@@ -1326,10 +1326,17 @@ class StableDiffusionGGML {
         uint32_t dim           = latents->ne[ggml_n_dims(latents) - 1];
 
         if (preview_mode == PREVIEW_PROJ) {
+            int64_t patch_sz                       = 1;
             const float(*latent_rgb_proj)[channel] = nullptr;
             float* latent_rgb_bias                 = nullptr;
 
-            if (dim == 48) {
+            if (dim == 128) {
+                if (sd_version_is_flux2(version)) {
+                    latent_rgb_proj = flux2_latent_rgb_proj;
+                    latent_rgb_bias = flux2_latent_rgb_bias;
+                    patch_sz        = 2;
+                }
+            } else if (dim == 48) {
                 if (sd_version_is_wan(version)) {
                     latent_rgb_proj = wan_22_latent_rgb_proj;
                     latent_rgb_bias = wan_22_latent_rgb_bias;
@@ -1382,12 +1389,15 @@ class StableDiffusionGGML {
                 frames = latents->ne[2];
             }
 
-            uint8_t* data = (uint8_t*)malloc(frames * width * height * channel * sizeof(uint8_t));
+            uint32_t img_width  = width * patch_sz;
+            uint32_t img_height = height * patch_sz;
+
+            uint8_t* data = (uint8_t*)malloc(frames * img_width * img_height * channel * sizeof(uint8_t));
 
-            preview_latent_video(data, latents, latent_rgb_proj, latent_rgb_bias, width, height, frames, dim);
+            preview_latent_video(data, latents, latent_rgb_proj, latent_rgb_bias, patch_sz);
             sd_image_t* images = (sd_image_t*)malloc(frames * sizeof(sd_image_t));
             for (int i = 0; i < frames; i++) {
-                images[i] = {width, height, channel, data + i * width * height * channel};
+                images[i] = {img_width, img_height, channel, data + i * img_width * img_height * channel};
             }
             step_callback(step, frames, images, is_noisy, step_callback_data);
             free(data);