feat: add PixArt model family support (#2047)

README.md

@@ -64,6 +64,7 @@ API and command-line option may change frequently.***
     - [HiDream-O1-Image](./docs/hidream_o1_image.md)
     - [Ideogram4](./docs/ideogram4.md)
     - [LLaDA-Image](./docs/llada_image.md)
+    - [PixArt](./docs/pixart.md)
   - [Image Edit Models](./docs/edit.md)
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
     - [Qwen Image Edit series](./docs/qwen_image_edit.md)

docs/pixart.md

@@ -0,0 +1,46 @@
+# How to Use
+
+You can run PixArt-α / PixArt-Σ with stable-diffusion.cpp.
+
+PixArt is a DiT-based text-to-image model family conditioned by a T5-XXL text
+encoder and a 4-channel VAE: SDXL-style for PixArt-Σ and SD1.x-style for PixArt-α.
+
+## Download weights
+
+- Download the transformer (diffusion model)
+    - PixArt-Σ XL-2 1024-MS: https://huggingface.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS/tree/main/transformer
+    - PixArt-α XL-2 1024-MS: https://huggingface.co/PixArt-alpha/PixArt-XL-2-1024-MS/tree/main/transformer
+- Download the T5-XXL text encoder
+    - safetensors: https://huggingface.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS/tree/main/text_encoder
+- Download the VAE
+    - PixArt-Σ: https://huggingface.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS/tree/main/vae
+    - PixArt-α: https://huggingface.co/PixArt-alpha/PixArt-XL-2-1024-MS/tree/main/vae
+    - Use the VAE matching the checkpoint's latent space. For TAE decoding or
+      preview, use TAESDXL for PixArt-Σ and TAESD for PixArt-α.
+- Tokenizer: the T5 vocabulary is embedded; no extra tokenizer file is needed.
+
+## Examples
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\pixart_sigma_xl2_1024_ms.safetensors --t5xxl ..\models\text_encoders\t5xxl.safetensors --vae ..\models\vae\pixart_vae.safetensors -p "a lovely cat" --cfg-scale 4.5 -W 1024 -H 1024 --steps 20 -v
+```
+
+## Notes
+
+- The VAE scaling factor defaults to `0.13025` for PixArt-Σ. PixArt-α
+  checkpoints with resolution micro-condition weights use `0.18215`.
+  PixArt-α 512 has the same tensor layout as PixArt-Σ, so it requires an
+  explicit override: `--model-args "pixart_vae_scale_factor=0.18215"`.
+  This argument can also override the scale for other compatible checkpoints.
+- PixArt-Σ checkpoints compute 2D sincos positional embeddings at runtime;
+  the trained grid is 64x64 patches with an interpolation scale of 2.
+  For checkpoints trained at a different resolution, the positional embedding
+  parameters can be adjusted via model args:
+  `--model-args "pixart_pos_embed_base_size=<trained grid>,pixart_interpolation_scale=<scale>"`
+  (e.g. `pixart_pos_embed_base_size=32,pixart_interpolation_scale=1,pixart_vae_scale_factor=0.18215` for
+  PixArt-α XL-2 512).
+- Checkpoints carrying resolution/aspect-ratio micro-condition weights are
+  detected but those conditions are not applied yet; a warning is logged and
+  generation proceeds with the timestep embedding only.
+- The transformer predicts 8 channels (noise + learned variance); only the
+  noise half is used for sampling, matching the reference implementation.

src/model.h

@@ -62,6 +62,7 @@ enum SDVersion {
     VERSION_SENSENOVA_U1_5,
     VERSION_LLADA_IMAGE,
     VERSION_ESRGAN,
+    VERSION_PIXART,
     VERSION_COUNT,
 };
 
@@ -252,6 +253,10 @@ static inline bool sd_version_is_sensenova_u1(SDVersion version) {
     return version == VERSION_SENSENOVA_U1_5;
 }
 
+static inline bool sd_version_is_pixart(SDVersion version) {
+    return version == VERSION_PIXART;
+}
+
 static inline bool sd_version_supports_video_generation(SDVersion version) {
     return version == VERSION_SVD || sd_version_is_wan(version) || sd_version_is_hunyuan_video(version) || sd_version_is_lingbot_video(version) || sd_version_is_ltxav(version) || sd_version_is_minimax_h3(version);
 }
@@ -320,7 +325,8 @@ static inline bool sd_version_is_dit(SDVersion version) {
         sd_version_is_sefi_image(version) ||
         sd_version_is_krea2(version) ||
         sd_version_is_mage_flow(version) ||
-        sd_version_is_sensenova_u1(version)) {
+        sd_version_is_sensenova_u1(version) ||
+        sd_version_is_pixart(version)) {
         return true;
     }
     return false;

src/model/diffusion/pixart.hpp

@@ -0,0 +1,390 @@
+#ifndef __SD_MODEL_DIFFUSION_PIXART_HPP__
+#define __SD_MODEL_DIFFUSION_PIXART_HPP__
+
+#include <algorithm>
+#include <cinttypes>
+#include <cmath>
+#include <vector>
+
+#include "core/ggml_extend.h"
+#include "core/ggml_runner.h"
+#include "core/util.h"
+#include "model/common/ggml_block.hpp"
+#include "model/diffusion/dit.hpp"
+#include "model/diffusion/mmdit.hpp"
+#include "model/diffusion/model.hpp"
+#include "model_loader.h"
+
+// Ref: https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/pixart_transformer_2d.py
+// Ref: https://github.com/PixArt-alpha/PixArt-sigma
+
+namespace PixArt {
+    constexpr int PIXART_GRAPH_SIZE = 20480;
+    constexpr int ADALN_EMBED_DIM   = 256;
+
+    struct PixArtConfig {
+        int64_t in_channels         = 4;
+        int64_t out_channels        = 8;  // learn_sigma: noise prediction + learned variance
+        int64_t hidden_size         = 1152;
+        int64_t cross_attention_dim = 1152;
+        int64_t caption_channels    = 4096;
+        int64_t num_heads           = 16;
+        int64_t patch_size          = 2;
+        int64_t ffn_dim             = 4608;
+        int64_t pos_embed_base_size = 64;
+        float interpolation_scale   = 2.f;
+        int num_layers              = 28;
+
+        static PixArtConfig detect_from_weights(const String2TensorStorage& weights, const std::string& prefix) {
+            PixArtConfig config;
+            auto find = [&](const std::string& suffix) -> const TensorStorage* {
+                auto it = weights.find(prefix + "." + suffix);
+                return it == weights.end() ? nullptr : &it->second;
+            };
+            if (auto w = find("pos_embed.proj.weight")) {
+                config.hidden_size = w->ne[3];
+                config.in_channels = w->ne[2];
+                config.patch_size  = w->ne[0];
+            }
+            if (auto w = find("proj_out.weight")) {
+                config.out_channels = w->ne[1] / (config.patch_size * config.patch_size);
+            }
+            if (auto w = find("caption_projection.linear_1.weight")) {
+                config.caption_channels = w->ne[0];
+            }
+            if (auto w = find("transformer_blocks.0.attn2.to_k.weight")) {
+                config.cross_attention_dim = w->ne[0];
+            }
+            if (auto w = find("transformer_blocks.0.ff.net.0.proj.weight")) {
+                config.ffn_dim = w->ne[1];
+            }
+            if (find("adaln_single.emb.resolution_embedder.linear_1.weight") != nullptr) {
+                LOG_WARN("pixart: resolution/aspect-ratio micro conditions are not supported; output may differ from the reference");
+            }
+            int layers                     = 0;
+            const std::string block_prefix = prefix + ".transformer_blocks.";
+            for (const auto& [name, _] : weights) {
+                if (starts_with(name, block_prefix)) {
+                    layers = std::max(layers, atoi(name.substr(block_prefix.size()).c_str()) + 1);
+                }
+            }
+            if (layers > 0) {
+                config.num_layers = layers;
+                LOG_VERBOSE("pixart: layers = %d, hidden_size = %" PRId64,
+                            layers, config.hidden_size);
+            }
+            return config;
+        }
+    };
+
+    // Mirrors diffusers get_2d_sincos_pos_embed for a (gh, gw) patch grid.
+    static std::vector<float> gen_2d_sincos_pos_embed(int64_t dim,
+                                                      int64_t gh,
+                                                      int64_t gw,
+                                                      int64_t base_size,
+                                                      float interpolation_scale) {
+        // diffusers: meshgrid(grid_w, grid_h, indexing="xy") -> grid[0]=w, grid[1]=h,
+        // embedding = concat(sincos(w), sincos(h))
+        std::vector<float> out(static_cast<size_t>(gh) * gw * dim);
+        int64_t quarter = dim / 4;
+        for (int64_t h = 0; h < gh; ++h) {
+            float pos_h = static_cast<float>(h) / (static_cast<float>(gh) / base_size) / interpolation_scale;
+            for (int64_t w = 0; w < gw; ++w) {
+                float pos_w  = static_cast<float>(w) / (static_cast<float>(gw) / base_size) / interpolation_scale;
+                float* dst_w = out.data() + (h * gw + w) * dim;
+                float* dst_h = dst_w + dim / 2;
+                for (int64_t i = 0; i < quarter; ++i) {
+                    float omega        = 1.f / powf(10000.f, static_cast<float>(i) / quarter);
+                    dst_w[i]           = sinf(pos_w * omega);
+                    dst_w[i + quarter] = cosf(pos_w * omega);
+                    dst_h[i]           = sinf(pos_h * omega);
+                    dst_h[i + quarter] = cosf(pos_h * omega);
+                }
+            }
+        }
+        return out;
+    }
+
+    class PixArtTimestepEmbedding : public GGMLBlock {
+    public:
+        PixArtTimestepEmbedding(int64_t in_channels, int64_t out_dim) {
+            blocks["linear_1"] = std::make_shared<Linear>(in_channels, out_dim);
+            blocks["linear_2"] = std::make_shared<Linear>(out_dim, out_dim);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            x = std::dynamic_pointer_cast<Linear>(blocks["linear_1"])->forward(ctx, x);
+            x = ggml_silu(ctx->ggml_ctx, x);
+            return std::dynamic_pointer_cast<Linear>(blocks["linear_2"])->forward(ctx, x);
+        }
+    };
+
+    class PixArtAttention : public GGMLBlock {
+        int64_t num_heads;
+
+    public:
+        PixArtAttention(int64_t dim, int64_t num_heads, int64_t context_dim)
+            : num_heads(num_heads) {
+            blocks["to_q"]     = std::make_shared<Linear>(dim, dim);
+            blocks["to_k"]     = std::make_shared<Linear>(context_dim, dim);
+            blocks["to_v"]     = std::make_shared<Linear>(context_dim, dim);
+            blocks["to_out.0"] = std::make_shared<Linear>(dim, dim);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* context, ggml_tensor* mask = nullptr) {
+            // x: [N, n_token, dim], context: [N, n_context, context_dim]
+            auto q = std::dynamic_pointer_cast<Linear>(blocks["to_q"])->forward(ctx, x);
+            auto k = std::dynamic_pointer_cast<Linear>(blocks["to_k"])->forward(ctx, context);
+            auto v = std::dynamic_pointer_cast<Linear>(blocks["to_v"])->forward(ctx, context);
+
+            auto out = ggml_ext_attention_ext(ctx, q, k, v, num_heads, mask, false, ctx->flash_attn_enabled);
+            return std::dynamic_pointer_cast<Linear>(blocks["to_out.0"])->forward(ctx, out);
+        }
+    };
+
+    class PixArtBlock : public GGMLBlock {
+        int64_t dim;
+
+        void init_params(ggml_context* ctx,
+                         const String2TensorStorage& tensor_storage_map = {},
+                         const std::string prefix                       = "") override {
+            ggml_type wtype             = get_type(prefix + "scale_shift_table", tensor_storage_map, GGML_TYPE_F32);
+            params["scale_shift_table"] = ggml_new_tensor_2d(ctx, wtype, dim, 6);
+        }
+
+    public:
+        PixArtBlock(int64_t dim, int64_t num_heads, int64_t context_dim, int64_t ffn_dim)
+            : dim(dim) {
+            blocks["attn1"]         = std::make_shared<PixArtAttention>(dim, num_heads, dim);
+            blocks["attn2"]         = std::make_shared<PixArtAttention>(dim, num_heads, context_dim);
+            blocks["ff.net.0.proj"] = std::make_shared<Linear>(dim, ffn_dim);
+            blocks["ff.net.2"]      = std::make_shared<Linear>(ffn_dim, dim);
+        }
+
+        static ggml_tensor* norm(ggml_context* ctx, ggml_tensor* x) {
+            return ggml_norm(ctx, x, 1e-6f);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* mod, ggml_tensor* context, ggml_tensor* context_mask) {
+            // x: [N, n_token, dim]
+            // mod: [N, 6 * dim], shared adaLN-single output
+            int64_t N = x->ne[2];
+
+            auto table = params["scale_shift_table"];
+            if (table->type != GGML_TYPE_F32) {
+                table = ggml_cast(ctx->ggml_ctx, table, GGML_TYPE_F32);
+            }
+            table = ggml_reshape_3d(ctx->ggml_ctx, table, dim, 6, 1);
+            auto m     = ggml_add(ctx->ggml_ctx, ggml_reshape_3d(ctx->ggml_ctx, mod, dim, 6, N), table);
+            auto mv    = ggml_ext_chunk(ctx->ggml_ctx, ggml_reshape_2d(ctx->ggml_ctx, ggml_ext_cont(ctx->ggml_ctx, m), dim * 6, N), 6, 0);
+
+            auto attn1 = std::dynamic_pointer_cast<PixArtAttention>(blocks["attn1"]);
+            auto attn2 = std::dynamic_pointer_cast<PixArtAttention>(blocks["attn2"]);
+            auto proj  = std::dynamic_pointer_cast<Linear>(blocks["ff.net.0.proj"]);
+            auto fc2   = std::dynamic_pointer_cast<Linear>(blocks["ff.net.2"]);
+
+            auto gate = [&](ggml_tensor* y, ggml_tensor* g) {
+                g = ggml_reshape_3d(ctx->ggml_ctx, g, dim, 1, N);
+                return ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, y, g));
+            };
+
+            auto h = modulate(ctx->ggml_ctx, norm(ctx->ggml_ctx, x), mv[0], mv[1]);
+            x      = gate(attn1->forward(ctx, h, h), mv[2]);
+            // ada_norm_single: no norm before cross-attention (PixArtMS.py)
+            x = ggml_add(ctx->ggml_ctx, x, attn2->forward(ctx, x, context, context_mask));
+            h = modulate(ctx->ggml_ctx, norm(ctx->ggml_ctx, x), mv[3], mv[4]);
+            h = proj->forward(ctx, h);
+            h = ggml_ext_gelu(ctx->ggml_ctx, h, true);
+            h = fc2->forward(ctx, h);
+            return gate(h, mv[5]);
+        }
+    };
+
+    class PixArtModel : public GGMLBlock {
+        PixArtConfig config;
+
+        void init_params(ggml_context* ctx,
+                         const String2TensorStorage& tensor_storage_map = {},
+                         const std::string prefix                       = "") override {
+            ggml_type wtype             = get_type(prefix + "scale_shift_table", tensor_storage_map, GGML_TYPE_F32);
+            params["scale_shift_table"] = ggml_new_tensor_2d(ctx, wtype, config.hidden_size, 2);
+        }
+
+    public:
+        PixArtModel() = default;
+        PixArtModel(const PixArtConfig& config)
+            : config(config) {
+            blocks["pos_embed.proj"]                     = std::make_shared<Conv2d>(config.in_channels,
+                                                                                    config.hidden_size,
+                                                                                    std::pair<int, int>{static_cast<int>(config.patch_size), static_cast<int>(config.patch_size)},
+                                                                                    std::pair<int, int>{static_cast<int>(config.patch_size), static_cast<int>(config.patch_size)});
+            blocks["adaln_single.emb.timestep_embedder"] = std::make_shared<PixArtTimestepEmbedding>(ADALN_EMBED_DIM, config.hidden_size);
+            blocks["adaln_single.linear"]                = std::make_shared<Linear>(config.hidden_size, 6 * config.hidden_size);
+            blocks["caption_projection.linear_1"]        = std::make_shared<Linear>(config.caption_channels, config.hidden_size);
+            blocks["caption_projection.linear_2"]        = std::make_shared<Linear>(config.hidden_size, config.cross_attention_dim);
+            for (int i = 0; i < config.num_layers; ++i) {
+                blocks["transformer_blocks." + std::to_string(i)] =
+                    std::make_shared<PixArtBlock>(config.hidden_size, config.num_heads, config.cross_attention_dim, config.ffn_dim);
+            }
+            blocks["norm_out"] = std::make_shared<LayerNorm>(config.hidden_size, 1e-6f, false);
+            blocks["proj_out"] = std::make_shared<Linear>(config.hidden_size,
+                                                          config.patch_size * config.patch_size * config.out_channels);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* timesteps,
+                             ggml_tensor* context,
+                             ggml_tensor* pos_embed,
+                             ggml_tensor* context_mask) {
+            // x: [N, C, H, W] latent, context: [N, n_ctx, caption_channels]
+            int64_t W  = x->ne[0];
+            int64_t H  = x->ne[1];
+            int64_t N  = x->ne[3];
+            int64_t p  = config.patch_size;
+            int64_t wp = W / p;
+            int64_t hp = H / p;
+
+            auto h = std::dynamic_pointer_cast<Conv2d>(blocks["pos_embed.proj"])->forward(ctx, x);  // [N, hidden, hp, wp]
+            h      = ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, h, 1, 2, 0, 3));      // [N, hp, wp, hidden] -> [N, hp*wp, hidden]
+            h      = ggml_reshape_3d(ctx->ggml_ctx, h, config.hidden_size, wp * hp, N);             // [N, hp*wp, hidden]
+            h      = ggml_add(ctx->ggml_ctx, h, pos_embed);
+
+            auto t   = ggml_ext_timestep_embedding(ctx->ggml_ctx, timesteps, ADALN_EMBED_DIM, 10000);
+            auto emb = std::dynamic_pointer_cast<PixArtTimestepEmbedding>(blocks["adaln_single.emb.timestep_embedder"])->forward(ctx, t);
+
+            auto mod = std::dynamic_pointer_cast<Linear>(blocks["adaln_single.linear"])
+                           ->forward(ctx, ggml_silu(ctx->ggml_ctx, emb));  // [N, 6 * hidden]
+
+            auto ctx_emb = std::dynamic_pointer_cast<Linear>(blocks["caption_projection.linear_1"])->forward(ctx, context);
+            ctx_emb      = ggml_ext_gelu(ctx->ggml_ctx, ctx_emb, true);
+            ctx_emb      = std::dynamic_pointer_cast<Linear>(blocks["caption_projection.linear_2"])->forward(ctx, ctx_emb);
+
+            for (int i = 0; i < config.num_layers; ++i) {
+                auto block = std::dynamic_pointer_cast<PixArtBlock>(blocks["transformer_blocks." + std::to_string(i)]);
+                h          = block->forward(ctx, h, mod, ctx_emb, context_mask);
+                sd::ggml_graph_cut::mark_graph_cut(h, "pixart.transformer_blocks." + std::to_string(i), "h");
+            }
+
+            // scale_shift_table + emb -> (shift, scale) for the affine-free final norm
+            auto tail_table = params["scale_shift_table"];
+            if (tail_table->type != GGML_TYPE_F32) {
+                tail_table = ggml_cast(ctx->ggml_ctx, tail_table, GGML_TYPE_F32);
+            }
+            auto ss    = ggml_add(ctx->ggml_ctx,
+                                  ggml_reshape_3d(ctx->ggml_ctx, tail_table, config.hidden_size, 2, 1),
+                                  ggml_reshape_3d(ctx->ggml_ctx, emb, config.hidden_size, 1, N));  // [2, hidden, N]
+            auto parts = ggml_ext_chunk(ctx->ggml_ctx,
+                                        ggml_reshape_2d(ctx->ggml_ctx, ggml_ext_cont(ctx->ggml_ctx, ss), config.hidden_size * 2, N),
+                                        2, 0);
+            h          = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_out"])->forward(ctx, h);
+            h          = modulate(ctx->ggml_ctx, h, parts[0], parts[1]);
+            h          = std::dynamic_pointer_cast<Linear>(blocks["proj_out"])->forward(ctx, h);  // [N, hp*wp, p*p*out_ch]
+            h          = DiT::unpatchify(ctx->ggml_ctx, h, hp, wp, static_cast<int>(p), static_cast<int>(p), false);
+            return h;  // [N, out_channels, H, W]
+        }
+    };
+
+    struct PixArtRunner : public DiffusionModelRunner {
+        PixArtConfig config;
+        PixArtModel model;
+        std::vector<float> pos_vec;
+
+        PixArtRunner(ggml_backend_t backend,
+                     const String2TensorStorage& tensor_storage_map      = {},
+                     const std::string prefix                            = "",
+                     std::shared_ptr<RunnerWeightManager> weight_manager = nullptr,
+                     const char* model_args                              = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(PixArtConfig::detect_from_weights(tensor_storage_map, prefix)) {
+            for (const auto& [key, value] : parse_key_value_args(model_args, "model arg")) {
+                if (key == "pixart_pos_embed_base_size") {
+                    int parsed = 0;
+                    if (parse_strict_int(value, parsed)) {
+                        config.pos_embed_base_size = parsed;
+                    } else {
+                        LOG_WARN("ignoring invalid PixArt model arg '%s=%s'", key.c_str(), value.c_str());
+                    }
+                } else if (key == "pixart_interpolation_scale") {
+                    float parsed = 0.f;
+                    if (parse_strict_float(value, parsed)) {
+                        config.interpolation_scale = parsed;
+                    } else {
+                        LOG_WARN("ignoring invalid PixArt model arg '%s=%s'", key.c_str(), value.c_str());
+                    }
+                }
+            }
+            model = PixArtModel(config);
+            model.init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "pixart";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor,
+                                 const sd::Tensor<float>& mask_tensor) {
+            ggml_cgraph* gf        = new_graph_custom(PIXART_GRAPH_SIZE);
+            ggml_tensor* x         = make_input(x_tensor);
+            ggml_tensor* timesteps = make_input(timesteps_tensor);
+            GGML_ASSERT(x->ne[3] == 1);
+            GGML_ASSERT(!context_tensor.empty());
+            ggml_tensor* context = make_input(context_tensor);
+
+            ggml_tensor* context_mask = nullptr;
+            if (!mask_tensor.empty()) {
+                // additive attention bias over context tokens: 0 keep / -inf discard
+                context_mask = ggml_reshape_4d(compute_ctx, make_input(mask_tensor), mask_tensor.shape()[0], 1, 1, 1);
+            }
+
+            int64_t W  = x->ne[0];
+            int64_t H  = x->ne[1];
+            int64_t wp = W / config.patch_size;
+            int64_t hp = H / config.patch_size;
+
+            pos_vec  = gen_2d_sincos_pos_embed(config.hidden_size, hp, wp,
+                                               config.pos_embed_base_size, config.interpolation_scale);
+            auto pos = ggml_new_tensor_3d(compute_ctx, GGML_TYPE_F32, config.hidden_size, wp * hp, 1);
+            set_backend_tensor_data(pos, pos_vec.data());
+
+            auto runner_ctx  = get_context();
+            ggml_tensor* out = model.forward(&runner_ctx, x, timesteps, context, pos, context_mask);
+            // learn_sigma: keep the noise prediction half of the output channels
+            out = ggml_ext_slice(compute_ctx, out, 2, 0, config.in_channels);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context,
+                                  const sd::Tensor<float>& context_mask) {
+            auto get_graph = [&]() -> ggml_cgraph* {
+                return build_graph(x, timesteps, context, context_mask);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute(get_graph, n_threads, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            auto context     = tensor_or_empty(diffusion_params.context);
+            auto context_msk = tensor_or_empty(diffusion_params.y);
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           context,
+                           context_msk);
+        }
+    };
+}  // namespace PixArt
+
+#endif  // __SD_MODEL_DIFFUSION_PIXART_HPP__

src/model/vae/auto_encoder_kl.hpp

@@ -544,7 +544,7 @@ public:
                        const String2TensorStorage& tensor_storage_map = {},
                        const std::string& prefix                      = "")
         : version(version), decode_only(decode_only), use_video_decoder(use_video_decoder) {
-        if (sd_version_is_dit(version)) {
+        if (sd_version_is_dit(version) && version != VERSION_PIXART) {
             if (sd_version_uses_flux2_vae(version)) {
                 dd_config.z_channels = 32;
                 embed_dim            = 32;
@@ -678,7 +678,7 @@ struct AutoEncoderKL : public VAE {
         if (sd_version_is_sd1(version) || sd_version_is_sd2(version)) {
             scale_factor = 0.18215f;
             shift_factor = 0.f;
-        } else if (sd_version_is_sdxl(version)) {
+        } else if (sd_version_is_sdxl(version) || sd_version_is_pixart(version)) {
             scale_factor = 0.13025f;
             shift_factor = 0.f;
         } else if (sd_version_is_sd3(version)) {

src/model/vae/tae.hpp

@@ -701,7 +701,7 @@ public:
         bool use_midblock_gn = false;
         taef2                = sd_version_uses_flux2_vae(version);
 
-        if (sd_version_is_dit(version)) {
+        if (sd_version_is_dit(version) && !sd_version_is_pixart(version)) {
             z_channels = 16;
         }
         if (taef2) {

src/model_loader.cpp

@@ -533,6 +533,10 @@ SDVersion ModelLoader::get_sd_version() const {
             return VERSION_ERNIE_IMAGE;
         }
         if (tensor_storage.name.find("model.diffusion_model.adaln_single.emb.timestep_embedder.linear_1.bias") != std::string::npos) {
+            // PixArt shares this signature with LTX-AV; pos_embed.proj is PixArt-only.
+            if (tensor_storage_map.find("model.diffusion_model.pos_embed.proj.weight") != tensor_storage_map.end()) {
+                return VERSION_PIXART;
+            }
             return VERSION_LTXAV;
         }
         if (tensor_storage.name.find("model.diffusion_model.video_patch_proj.weight") != std::string::npos &&
@@ -1597,6 +1601,9 @@ bool ModelLoader::tensor_should_be_converted(const TensorStorage& tensor_storage
             // Pass, do not convert. For Unet
         } else if (contains(name, "embedding")) {
             // Pass, do not convert embedding
+        } else if (contains(name, "scale_shift_table")) {
+            // Pass, do not convert. adaLN modulation tables (PixArt, LTXV) are sliced
+            // element-wise, which is invalid on quantized block layouts.
         } else if (ends_with(name, "_pad_token")) {
             // Pass, do not convert. LLaDA-Image stores its pad tokens far outside the f16
             // range, so any format with an f16 scale or payload turns them into inf.

src/pipeline/diffusion_engine.cpp

@@ -105,6 +105,7 @@ const char* model_version_to_str[] = {
     "SenseNova U1.5",
     "LLaDA-Image",
     "ESRGAN",
+    "PixArt",
 };
 
 static_assert(VERSION_COUNT == sizeof(model_version_to_str) / sizeof(model_version_to_str[0]),
@@ -125,6 +126,18 @@ void calculate_alphas_cumprod(float* alphas_cumprod,
     }
 }
 
+void calculate_alphas_cumprod_linear_beta(float* alphas_cumprod,
+                                          float beta_start,
+                                          float beta_end,
+                                          int timesteps = TIMESTEPS) {
+    float product = 1.0f;
+    for (int i = 0; i < timesteps; i++) {
+        float beta = beta_start + (beta_end - beta_start) * ((float)i / (timesteps - 1));
+        product *= 1.0f - beta;
+        alphas_cumprod[i] = product;
+    }
+}
+
 template <typename T, typename = void>
 struct has_set_runtime_backends : std::false_type {};
 template <typename T>
@@ -666,6 +679,10 @@ void StableDiffusionGGML::refresh_compvis_denoiser_sigmas() {
     std::vector<float> alphas_cumprod(TIMESTEPS);
     if (file_alphas_cumprod.size() == TIMESTEPS) {
         alphas_cumprod = file_alphas_cumprod;
+    } else if (sd_version_is_pixart(version)) {
+        // PixArt checkpoints train with a linear beta schedule (0.0001 -> 0.02)
+        // instead of the scaled_linear schedule used by SD1.x/SDXL.
+        calculate_alphas_cumprod_linear_beta(alphas_cumprod.data(), 0.0001f, 0.02f);
     } else {
         calculate_alphas_cumprod(alphas_cumprod.data());
     }
@@ -2731,7 +2748,7 @@ int StableDiffusionGGML::get_diffusion_model_down_factor() {
     if (sd_version_is_dit(version)) {
         if (sd_version_is_sensenova_u1(version)) {
             down_factor = 32;
-        } else if (version == VERSION_QWEN_IMAGE_2_1 || sd_version_is_wan(version) || sd_version_is_lingbot_video(version) || sd_version_is_minimax_h3(version)) {
+        } else if (version == VERSION_QWEN_IMAGE_2_1 || sd_version_is_wan(version) || sd_version_is_lingbot_video(version) || sd_version_is_minimax_h3(version) || sd_version_is_pixart(version)) {
             down_factor = 2;
         } else {
             down_factor = 1;
@@ -2769,6 +2786,8 @@ int StableDiffusionGGML::get_latent_channel() {
             latent_channel = 128;
         } else if (sd_version_is_mage_flow(version)) {
             latent_channel = 128;
+        } else if (sd_version_is_pixart(version)) {
+            latent_channel = 4;
         } else {
             latent_channel = 16;
         }

src/pipeline/model_builders.cpp

@@ -1,5 +1,6 @@
 #include "model_builders.h"
 
+#include <cmath>
 #include <cstring>
 #include <utility>
 
@@ -28,6 +29,7 @@
 #include "model/diffusion/mmdit.hpp"
 #include "model/diffusion/model.hpp"
 #include "model/diffusion/pid.hpp"
+#include "model/diffusion/pixart.hpp"
 #include "model/diffusion/qwen_image.hpp"
 #include "model/diffusion/qwen_image_2_1.hpp"
 #include "model/diffusion/sensenova_u1.h"
@@ -301,6 +303,19 @@ namespace sd::model_builders {
                                                                            weight_manager,
                                                                            sd_ctx_params->model_args);
             }
+        } else if (version == VERSION_PIXART) {
+            result.conditioner = std::make_shared<T5CLIPEmbedder>(ctx.backends.runtime_backend(SDBackendModule::TE),
+                                                                  tensor_storage_map,
+                                                                  true,
+                                                                  0,
+                                                                  false,
+                                                                  weight_manager,
+                                                                  sd_ctx_params->model_args);
+            result.diffusion   = std::make_shared<PixArt::PixArtRunner>(ctx.backends.runtime_backend(SDBackendModule::DIFFUSION),
+                                                                      tensor_storage_map,
+                                                                      "model.diffusion_model",
+                                                                      weight_manager,
+                                                                      sd_ctx_params->model_args);
         } else if (sd_version_is_mage_flow(version)) {
             result.conditioner = std::make_shared<LLMEmbedder>(ctx.backends.runtime_backend(SDBackendModule::TE),
                                                                tensor_storage_map,
@@ -562,6 +577,23 @@ namespace sd::model_builders {
                                                              false,
                                                              vae_version,
                                                              weight_manager);
+                if (sd_version_is_pixart(version)) {
+                    // Alpha-512 and Sigma share tensor layouts; Alpha-512 needs an explicit scale override.
+                    if (tensor_storage_map.count("model.diffusion_model.adaln_single.emb.resolution_embedder.linear_1.weight") != 0) {
+                        model->scale_factor = 0.18215f;
+                    }
+                    for (const auto& [key, value] : parse_key_value_args(sd_ctx_params->model_args, "model arg")) {
+                        if (key == "pixart_vae_scale_factor") {
+                            float parsed = 0.f;
+                            if (parse_strict_float(value, parsed) && std::isfinite(parsed) && parsed > 0.f) {
+                                model->scale_factor = parsed;
+                            } else {
+                                LOG_WARN("ignoring invalid PixArt model arg '%s=%s'", key.c_str(), value.c_str());
+                            }
+                        }
+                    }
+                    LOG_VERBOSE("pixart: VAE scale factor = %.5f", model->scale_factor);
+                }
                 if (sd_version_is_sdxl(version) &&
                     (strlen(SAFE_STR(sd_ctx_params->vae_path)) == 0 || sd_ctx_params->force_sdxl_vae_conv_scale || options.external_vae_is_invalid)) {
                     float vae_conv_2d_scale = 1.f / 32.f;

src/tokenizers/t5_unigram_tokenizer.cpp

@@ -295,6 +295,12 @@ bool T5UniGramTokenizer::encode(const std::string& input, std::vector<int>& resu
     std::vector<int32_t> tokens;
     std::vector<std::string> token_strs;
     std::string normalized = normalize(input);
+    if (normalized.empty()) {
+        // HF reference tokenizers emit no pieces for empty input; pad_tokens
+        // still appends EOS so the sequence becomes [EOS] + padding.
+        result = std::move(tokens);
+        return true;
+    }
     auto splited_texts     = split_with_special_tokens(normalized, special_tokens);
     if (splited_texts.empty()) {
         splited_texts.push_back(normalized);  // for empty string