fix: resolve precision issues in SDXL VAE under fp16 (#888)

* fix: resolve precision issues in SDXL VAE under fp16

* add --force-sdxl-vae-conv-scale option

* update docs

README.md

@@ -17,7 +17,6 @@ API and command-line option may change frequently.***
   - Image Models
     - SD1.x, SD2.x, [SD-Turbo](https://huggingface.co/stabilityai/sd-turbo)
     - SDXL, [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo)
-      - !!!The VAE in SDXL encounters NaN issues under FP16, but unfortunately, the ggml_conv_2d only operates under FP16. Hence, a parameter is needed to specify the VAE that has fixed the FP16 NaN issue. You can find it here: [SDXL VAE FP16 Fix](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors).
     - [SD3/SD3.5](./docs/sd3.md)
     - [Flux-dev/Flux-schnell](./docs/flux.md)
     - [Chroma](./docs/chroma.md)
@@ -365,6 +364,7 @@ arguments:
   --vae-tile-size [X]x[Y]            tile size for vae tiling (default: 32x32)
   --vae-relative-tile-size [X]x[Y]   relative tile size for vae tiling, in fraction of image size if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)
   --vae-tile-overlap OVERLAP         tile overlap for vae tiling, in fraction of tile size (default: 0.5)
+  --force-sdxl-vae-conv-scale        force use of conv scale on sdxl vae
   --vae-on-cpu                       keep vae in cpu (for low vram)
   --clip-on-cpu                      keep clip in cpu (for low vram)
   --diffusion-fa                     use flash attention in the diffusion model (for low vram)

conditioner.hpp

@@ -1457,7 +1457,7 @@ struct Qwen2_5_VLCLIPEmbedder : public Conditioner {
                                       const ConditionerParams& conditioner_params) {
         std::string prompt;
         std::vector<std::pair<int, ggml_tensor*>> image_embeds;
-        size_t system_prompt_length = 0;
+        size_t system_prompt_length          = 0;
         int prompt_template_encode_start_idx = 34;
         if (qwenvl->enable_vision && conditioner_params.ref_images.size() > 0) {
             LOG_INFO("QwenImageEditPlusPipeline");

examples/cli/main.cpp

@@ -131,6 +131,7 @@ struct SDParams {
     prediction_t prediction = DEFAULT_PRED;
 
     sd_tiling_params_t vae_tiling_params = {false, 0, 0, 0.5f, 0.0f, 0.0f};
+    bool force_sdxl_vae_conv_scale       = false;
 
     SDParams() {
         sd_sample_params_init(&sample_params);
@@ -198,6 +199,7 @@ void print_params(SDParams params) {
     printf("    seed:                              %zd\n", params.seed);
     printf("    batch_count:                       %d\n", params.batch_count);
     printf("    vae_tiling:                        %s\n", params.vae_tiling_params.enabled ? "true" : "false");
+    printf("    force_sdxl_vae_conv_scale:         %s\n", params.force_sdxl_vae_conv_scale ? "true" : "false");
     printf("    upscale_repeats:                   %d\n", params.upscale_repeats);
     printf("    chroma_use_dit_mask:               %s\n", params.chroma_use_dit_mask ? "true" : "false");
     printf("    chroma_use_t5_mask:                %s\n", params.chroma_use_t5_mask ? "true" : "false");
@@ -292,6 +294,7 @@ void print_usage(int argc, const char* argv[]) {
     printf("  --vae-tile-size [X]x[Y]            tile size for vae tiling (default: 32x32)\n");
     printf("  --vae-relative-tile-size [X]x[Y]   relative tile size for vae tiling, in fraction of image size if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)\n");
     printf("  --vae-tile-overlap OVERLAP         tile overlap for vae tiling, in fraction of tile size (default: 0.5)\n");
+    printf("  --force-sdxl-vae-conv-scale        force use of conv scale on sdxl vae\n");
     printf("  --vae-on-cpu                       keep vae in cpu (for low vram)\n");
     printf("  --clip-on-cpu                      keep clip in cpu (for low vram)\n");
     printf("  --diffusion-fa                     use flash attention in the diffusion model (for low vram)\n");
@@ -562,6 +565,7 @@ void parse_args(int argc, const char** argv, SDParams& params) {
 
     options.bool_options = {
         {"", "--vae-tiling", "", true, &params.vae_tiling_params.enabled},
+        {"", "--force-sdxl-vae-conv-scale", "", true, &params.force_sdxl_vae_conv_scale},
         {"", "--offload-to-cpu", "", true, &params.offload_params_to_cpu},
         {"", "--control-net-cpu", "", true, &params.control_net_cpu},
         {"", "--clip-on-cpu", "", true, &params.clip_on_cpu},
@@ -1382,6 +1386,7 @@ int main(int argc, const char* argv[]) {
         params.diffusion_flash_attn,
         params.diffusion_conv_direct,
         params.vae_conv_direct,
+        params.force_sdxl_vae_conv_scale,
         params.chroma_use_dit_mask,
         params.chroma_use_t5_mask,
         params.chroma_t5_mask_pad,

ggml_extend.hpp

@@ -975,38 +975,28 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_conv_2d(struct ggml_context* ctx,
                                                       struct ggml_tensor* x,
                                                       struct ggml_tensor* w,
                                                       struct ggml_tensor* b,
-                                                      int s0 = 1,
+                                                      int s0      = 1,
-                                                      int s1 = 1,
+                                                      int s1      = 1,
-                                                      int p0 = 0,
+                                                      int p0      = 0,
-                                                      int p1 = 0,
+                                                      int p1      = 0,
-                                                      int d0 = 1,
+                                                      int d0      = 1,
-                                                      int d1 = 1) {
+                                                      int d1      = 1,
-    x = ggml_conv_2d(ctx, w, x, s0, s1, p0, p1, d0, d1);
+                                                      bool direct = false,
-    if (b != NULL) {
+                                                      float scale = 1.f) {
-        b = ggml_reshape_4d(ctx, b, 1, 1, b->ne[0], 1);
+    if (scale != 1.f) {
-        // b = ggml_repeat(ctx, b, x);
+        x = ggml_scale(ctx, x, scale);
-        x = ggml_add_inplace(ctx, x, b);
+    }
+    if (direct) {
+        x = ggml_conv_2d_direct(ctx, w, x, s0, s1, p0, p1, d0, d1);
+    } else {
+        x = ggml_conv_2d(ctx, w, x, s0, s1, p0, p1, d0, d1);
+    }
+    if (scale != 1.f) {
+        x = ggml_scale(ctx, x, 1.f / scale);
     }
-    return x;
-}
-
-// w: [OC*IC, KD, KH, KW]
-// x: [N*IC, ID, IH, IW]
-__STATIC_INLINE__ struct ggml_tensor* ggml_nn_conv_2d_direct(struct ggml_context* ctx,
-                                                             struct ggml_tensor* x,
-                                                             struct ggml_tensor* w,
-                                                             struct ggml_tensor* b,
-                                                             int s0 = 1,
-                                                             int s1 = 1,
-                                                             int p0 = 0,
-                                                             int p1 = 0,
-                                                             int d0 = 1,
-                                                             int d1 = 1) {
-    x = ggml_conv_2d_direct(ctx, w, x, s0, s1, p0, p1, d0, d1);
     if (b != NULL) {
         b = ggml_reshape_4d(ctx, b, 1, 1, b->ne[0], 1);
-        // b = ggml_repeat(ctx, b, x);
+        x = ggml_add_inplace(ctx, x, b);
-        x = ggml_add(ctx, x, b);
     }
     return x;
 }
@@ -2067,6 +2057,7 @@ protected:
     std::pair<int, int> dilation;
     bool bias;
     bool direct = false;
+    float scale = 1.f;
 
     void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types, const std::string prefix = "") {
         enum ggml_type wtype = GGML_TYPE_F16;
@@ -2097,6 +2088,10 @@ public:
         direct = true;
     }
 
+    void set_scale(float scale_value) {
+        scale = scale_value;
+    }
+
     std::string get_desc() {
         return "Conv2d";
     }
@@ -2107,11 +2102,18 @@ public:
         if (bias) {
             b = params["bias"];
         }
-        if (direct) {
+        return ggml_nn_conv_2d(ctx,
-            return ggml_nn_conv_2d_direct(ctx, x, w, b, stride.second, stride.first, padding.second, padding.first, dilation.second, dilation.first);
+                               x,
-        } else {
+                               w,
-            return ggml_nn_conv_2d(ctx, x, w, b, stride.second, stride.first, padding.second, padding.first, dilation.second, dilation.first);
+                               b,
-        }
+                               stride.second,
+                               stride.first,
+                               padding.second,
+                               padding.first,
+                               dilation.second,
+                               dilation.first,
+                               direct,
+                               scale);
     }
 };
 

qwen_image.hpp

@@ -535,7 +535,7 @@ namespace Qwen {
                 }
             }
             LOG_ERROR("qwen_image_params.num_layers: %ld", qwen_image_params.num_layers);
-            qwen_image                   = QwenImageModel(qwen_image_params);
+            qwen_image = QwenImageModel(qwen_image_params);
             qwen_image.init(params_ctx, tensor_types, prefix);
         }
 

stable-diffusion.cpp

@@ -330,13 +330,6 @@ public:
 
         if (sd_version_is_sdxl(version)) {
             scale_factor = 0.13025f;
-            if (strlen(SAFE_STR(sd_ctx_params->vae_path)) == 0 && strlen(SAFE_STR(sd_ctx_params->taesd_path)) == 0) {
-                LOG_WARN(
-                    "!!!It looks like you are using SDXL model. "
-                    "If you find that the generated images are completely black, "
-                    "try specifying SDXL VAE FP16 Fix with the --vae parameter. "
-                    "You can find it here: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors");
-            }
         } else if (sd_version_is_sd3(version)) {
             scale_factor = 1.5305f;
         } else if (sd_version_is_flux(version)) {
@@ -517,6 +510,15 @@ public:
                     LOG_INFO("Using Conv2d direct in the vae model");
                     first_stage_model->enable_conv2d_direct();
                 }
+                if (version == VERSION_SDXL &&
+                    (strlen(SAFE_STR(sd_ctx_params->vae_path)) == 0 || sd_ctx_params->force_sdxl_vae_conv_scale)) {
+                    float vae_conv_2d_scale = 1.f / 32.f;
+                    LOG_WARN(
+                        "No VAE specified with --vae or --force-sdxl-vae-conv-scale flag set, "
+                        "using Conv2D scale %.3f",
+                        vae_conv_2d_scale);
+                    first_stage_model->set_conv2d_scale(vae_conv_2d_scale);
+                }
                 first_stage_model->alloc_params_buffer();
                 first_stage_model->get_param_tensors(tensors, "first_stage_model");
             } else {

stable-diffusion.h

@@ -164,6 +164,7 @@ typedef struct {
     bool diffusion_flash_attn;
     bool diffusion_conv_direct;
     bool vae_conv_direct;
+    bool force_sdxl_vae_conv_scale;
     bool chroma_use_dit_mask;
     bool chroma_use_t5_mask;
     int chroma_t5_mask_pad;

vae.hpp

@@ -530,6 +530,7 @@ struct VAE : public GGMLRunner {
                          struct ggml_context* output_ctx)                                                         = 0;
     virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) = 0;
     virtual void enable_conv2d_direct(){};
+    virtual void set_conv2d_scale(float scale) { SD_UNUSED(scale); };
 };
 
 struct AutoEncoderKL : public VAE {
@@ -558,6 +559,17 @@ struct AutoEncoderKL : public VAE {
         }
     }
 
+    void set_conv2d_scale(float scale) {
+        std::vector<GGMLBlock*> blocks;
+        ae.get_all_blocks(blocks);
+        for (auto block : blocks) {
+            if (block->get_desc() == "Conv2d") {
+                auto conv_block = (Conv2d*)block;
+                conv_block->set_scale(scale);
+            }
+        }
+    }
+
     std::string get_desc() {
         return "vae";
     }