feat: add MiniMax-H3-Image support (#2324)

2026-11-10 19:06:33 +00:00 · 2026-11-11 01:24:25 +08:00
1 parent f8417d02e1
commit 32d516c019
@@ -15,6 +15,7 @@ API and command-line option may change frequently.***

 ## 🔥Important News

+* **2026/11/11** 🚀 stable-diffusion.cpp adds **Day-0 support for MiniMax-H3-Image**
 * **2026/10/16** 🚀 stable-diffusion.cpp adds **Day-0 support for MOSS-Image-Base / MOSS-Image-SFT**
 * **2026/10/16** 🚀 stable-diffusion.cpp adds **Day-0 support for FLUX.3 Image Klein / Klein Base**
 * **2026/09/20** 🚀 stable-diffusion.cpp adds **Day-0 support for Qwen-Image-2.1**
@@ -60,6 +61,7 @@ API and command-line option may change frequently.***
    - [Z-Image L2P](./docs/z_image_l2p.md)
    - [MiniT2I](./docs/minit2i.md)
    - [P-H3-Image](./docs/p_h3_image.md)
+    - [MiniMax-H3-Image](./docs/minimax_h3_image.md)
    - [SenseNova U1.5](./docs/sensenova_u1.md)
    - [Ovis-Image](./docs/ovis_image.md)
    - [Anima](./docs/anima.md)
@@ -0,0 +1,65 @@
+# MiniMax-H3-Image
+
+MiniMax-H3-Image supports text-to-image generation and reference-image editing
+with the same checkpoint. It reuses the MiniMax-H3 text encoder and uses a
+dedicated image VAE.
+
+## Model files
+
+Pass the three MiniMax-H3-Image components separately:
+
+- `--diffusion-model`: MiniMax-H3-Image diffusion transformer
+- `--vae`: MiniMax-H3-Image VAE
+- `--llm`: the MiniMax-H3 Qwen3-VL-32B text encoder checkpoint
+
+Use the MiniMax-H3 text encoder, including its Qwen3-VL vision tower and
+DeepStack mergers. An existing MiniMax-H3 text encoder can be reused. If the
+vision tower is stored separately, pass it with `--llm_vision`.
+
+Use a single-file diffusion checkpoint with the six-step PDD output heads
+included, such as `minimax_h3_image_pruned_bf16.safetensors` or its INT8 ConvRot
+version. Use the MiniMax-H3-Image VAE; the MiniMax-H3 video VAE and P-H3-Image
+VAE are incompatible. No audio VAE or P-H3 text connector is needed.
+
+### Download weights
+
+- Download minimax_h3_image_pruned
+    - safetensors: https://huggingface.co/Comfy-Org/MiniMax-H3-Image/tree/main/diffusion_models
+- Download qwen3vl_32b_minimax_h3
+    - safetensors: https://huggingface.co/Comfy-Org/MiniMax-H3-Image/tree/main/text_encoders
+    - gguf: https://huggingface.co/leejet/MiniMax-H3-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/Comfy-Org/MiniMax-H3-Image/tree/main/vae
+
+## Text-to-image
+
+```sh
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\minimax_h3_image_pruned_bf16.safetensors --vae ..\models\vae\minimax_h3_image_vae_fp16.safetensors --llm ..\models\text_encoders\qwen3vl_32b_minimax_h3-Q4_K_M.gguf -p "A silver tabby kitten holding a sign reading 'sd.cpp', natural daylight" --cfg-scale 1.0 --steps 6 --sampling-method euler --sigmas minimax-h3-image -v -W 1024 -H 1024 --offload-to-cpu --rng cpu --diffusion-fa --vae-tiling
+```
+
+Use Euler, CFG 1, and `--sigmas minimax-h3-image` for the model's six-step
+schedule. The default discrete scheduler also produces this schedule with
+`--steps 6`. Prompt expansion is not included.
+
+## Reference-image editing
+
+Add `--ref-image` (or `-r`) for reference-image editing:
+
+```sh
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\minimax_h3_image_pruned_bf16.safetensors --vae ..\models\vae\minimax_h3_image_vae_fp16.safetensors --llm ..\models\text_encoders\qwen3vl_32b_minimax_h3-Q4_K_M.gguf -p "Edit <Picture 1>. Put a red scarf on the cat, preserving its face and the background." -r ..\assets\ernie_image\turbo_example.png --cfg-scale 1.0 --steps 6 --sampling-method euler --sigmas minimax-h3-image -v -W 1024 -H 1024 --offload-to-cpu --rng cpu --diffusion-fa --vae-tiling
+```
+
+`--ref-image` can be repeated. Reference images are numbered in argument order
+as `<Picture 1>`, `<Picture 2>`, and so on, with up to 16 references supported
+by the model. Prompts are passed through without an automatic editing prefix.
+
+Reference images keep their aspect ratio. The text encoder sees a 1024-pixel
+short edge, while VAE inputs use a short edge between 1024 and 2048 pixels.
+Small references are enlarged and dimensions are rounded to multiples of 32.
+
+## Shape and runtime notes
+
+- Width and height are aligned upward to a multiple of 32.
+- Start at 1024 × 1024. Larger canvases and more references increase memory use.
+- Use quantized weights, `--offload-to-cpu`, and `--vae-tiling` to reduce GPU
+  memory requirements.
@@ -1540,6 +1540,8 @@ ArgOptions SDGenerationParams::get_options() {
            sigmas_str = "1.0,0.978453,0.95418,0.926626,0.89508,0.845148,0.704534,0.414568,0.0";
        } else if (sigmas_str == "p-h3-image") {
            sigmas_str = "1.0,0.984753,0.967905,0.951688,0.925372,0.881923,0.810083,0.528718,0.0";
+        } else if (sigmas_str == "minimax-h3-image") {
+            sigmas_str = "1.0,0.9675236940383911,0.8875485062599182,0.75,0.5327739119529724,0.23200887441635132,0.0";
        } else if (sigmas_str == "hyperflow-h3-flash-4") {
            sigmas_str = "1.0,0.962575853,0.755781889,0.312497854,0.0";
        } else if (sigmas_str == "hyperflow-h3-flash-6") {
@@ -1793,11 +1795,11 @@ ArgOptions SDGenerationParams::get_options() {
         on_scheduler_arg},
        {"",
         "--sigmas",
-         "custom sigma values for the sampler, comma-separated (e.g., \"14.61,7.8,3.5,0.0\") or preset: qwen-image-2.1-turbo, p-h3-image, hyperflow-h3-flash-4, hyperflow-h3-flash-6, hyperflow-h3-flash-8.",
+         "custom sigma values for the sampler, comma-separated (e.g., \"14.61,7.8,3.5,0.0\") or preset: qwen-image-2.1-turbo, p-h3-image, minimax-h3-image, hyperflow-h3-flash-4, hyperflow-h3-flash-6, hyperflow-h3-flash-8.",
         on_sigmas_arg},
        {"",
         "--hires-sigmas",
-         "custom sigma values for the highres fix second pass, comma-separated (e.g., \"0.85,0.725,0.421875,0.0\") or preset: qwen-image-2.1-turbo, p-h3-image, hyperflow-h3-flash-4, hyperflow-h3-flash-6, hyperflow-h3-flash-8.",
+         "custom sigma values for the highres fix second pass, comma-separated (e.g., \"0.85,0.725,0.421875,0.0\") or preset: qwen-image-2.1-turbo, p-h3-image, minimax-h3-image, hyperflow-h3-flash-4, hyperflow-h3-flash-6, hyperflow-h3-flash-8.",
         on_hires_sigmas_arg},
        {"",
         "--skip-layers",
@@ -70,6 +70,7 @@ For community edit adapters and other optional extensions, these are the base-mo
 | `ming-image` | 1 | 12 | 1024 × 1024 | [Source](https://huggingface.co/inclusionAI/Ming-Image-0.1-Design)  |
 | `minimax-h3-fl2va` | 1 | — | 1344 × 768 | [Source](https://huggingface.co/docs/diffusers/main/en/api/pipelines/minimax_h3) Guidance-distilled; official 768p canvas. No fixed step count is prescribed here; the upstream API counts sigma grid points rather than sd.cpp denoising steps. |
 | `minimax-h3-ref2va` | 1 | — | 1344 × 768 | [Source](https://huggingface.co/docs/diffusers/main/en/api/pipelines/minimax_h3) Guidance-distilled; official 768p canvas. No fixed step count is prescribed here; the upstream API counts sigma grid points rather than sd.cpp denoising steps. |
+| `minimax-h3-image` | 1 | 6 | 1024 × 1024 | [Runtime guide](../../../docs/minimax_h3_image.md) Dedicated image VAE and fixed PDD sigma schedule. |
 | `minit2i` | 2.5 | 20 | 512 × 512 | [Source](https://huggingface.co/MiniT2I/MiniT2I/raw/main/pipeline.py) B/16 example and native model size; L/16 uses CFG 6. |
 | `moss-image-base` | 3 | 20 | 1024 × 1024 | [Source](https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base)  |
 | `moss-image-sft` | 3 | 20 | 1024 × 1024 | [Source](https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT)  |
@@ -0,0 +1,48 @@
+{
+  "schema_version": 1,
+  "id": "minimax-h3-image",
+  "name": "MiniMax H3 Image",
+  "size_multiple": 32,
+  "description": "Text-to-image and reference-image editing with the six-step PDD checkpoint. Requires the MiniMax-H3-Image 2D VAE and MiniMax-H3 Qwen3-VL-32B text encoder with its vision tower.",
+  "model_url": "https://huggingface.co/MiniMaxAI/MiniMax-H3-Image",
+  "slots": [
+    {
+      "id": "diffusion_model",
+      "required": true
+    },
+    {
+      "id": "llm",
+      "required": true,
+      "label": "MiniMax-H3 Qwen 3 VL 32B Text Encoder"
+    },
+    {
+      "id": "llm_vision",
+      "required": false,
+      "label": "Vision tower (required when not included in the text encoder)"
+    },
+    {
+      "id": "vae",
+      "required": true,
+      "label": "MiniMax-H3-Image 2D VAE"
+    }
+  ],
+  "defaults": {
+    "sampling_method": "euler",
+    "cfg_scale": 1,
+    "steps": 6,
+    "width": 1024,
+    "height": 1024,
+    "custom_sigmas": [
+      1.0,
+      0.9675236940383911,
+      0.8875485062599182,
+      0.75,
+      0.5327739119529724,
+      0.23200887441635132,
+      0.0
+    ]
+  },
+  "features": [
+    "ref_images"
+  ]
+}
@@ -2589,7 +2589,9 @@ struct LLMEmbedder : public Conditioner {
                } else if (conditioner_params.ref_images != nullptr) {
                    for (size_t i = 0; i < conditioner_params.ref_images->size(); ++i) {
                        auto resized = resize_for_vision((*conditioner_params.ref_images)[i]);
-                        prompt += "<Picture " + std::to_string(i + 1) + ">: ";
+                        prompt += version == VERSION_MINIMAX_H3_IMAGE
+                                      ? "Picture " + std::to_string(i + 1) + ": "
+                                      : "<Picture " + std::to_string(i + 1) + ">: ";
                        if (!add_vision_outputs(llm->encode_image_outputs(n_threads,
                                                                          resized,
                                                                          false),
@@ -71,6 +71,7 @@ enum SDVersion {
    VERSION_FIBO2,
    VERSION_CHROMA,
    VERSION_P_H3_IMAGE,
+    VERSION_MINIMAX_H3_IMAGE,
    VERSION_COUNT,
 };

@@ -167,7 +168,11 @@ static inline bool sd_version_is_ltxav(SDVersion version) {
 }

 static inline bool sd_version_is_minimax_h3(SDVersion version) {
-    return version == VERSION_MINIMAX_H3 || version == VERSION_P_H3_IMAGE;
+    return version == VERSION_MINIMAX_H3 || version == VERSION_P_H3_IMAGE || version == VERSION_MINIMAX_H3_IMAGE;
+}
+
+static inline bool sd_version_is_minimax_h3_image(SDVersion version) {
+    return version == VERSION_P_H3_IMAGE || version == VERSION_MINIMAX_H3_IMAGE;
 }

 static inline bool sd_version_is_wan(SDVersion version) {
@@ -2,8 +2,10 @@
 #define __SD_MODEL_DIFFUSION_MINIMAX_H3_HPP__

 #include <algorithm>
+#include <array>
 #include <cinttypes>
 #include <cmath>
+#include <cstddef>
 #include <functional>
 #include <set>
 #include <string>
@@ -24,6 +26,10 @@ namespace MiniMaxH3 {
    constexpr float FRAME_RESCALE                 = 5.f / 3.f;
    constexpr float VISUAL_COND_TIMESTEP          = 0.999f;

+    inline constexpr std::array<float, 7> pdd_sigmas = {
+        1.f, 0.9675236940383911f, 0.8875485062599182f, 0.75f,
+        0.5327739119529724f, 0.23200887441635132f, 0.f};
+
    struct LowRankShape {
        int64_t q   = 0;
        int64_t k   = 0;
@@ -52,6 +58,7 @@ namespace MiniMaxH3 {
        int64_t time_embed_dim           = 2688;
        int64_t rope_inv_freq_len        = 16;
        int64_t adaln_curve_grid         = 0;
+        int64_t pdd_num_inference_steps  = 1;
        int patch_t                      = 1;
        int patch_h                      = 2;
        int patch_w                      = 2;
@@ -95,6 +102,21 @@ namespace MiniMaxH3 {
            }
            if (const auto* weight = find("audio_patch_proj.weight")) {
                config.audio_latent_channels = weight->logical_ne0();
+            } else {
+                config.audio_latent_channels = 0;
+            }
+            if (const auto* weight = find("final_layer.video_out.weight")) {
+                const int64_t patch_dim = config.video_latent_channels * config.patch_t * config.patch_h * config.patch_w;
+                if (patch_dim <= 0 || weight->ne[1] % patch_dim != 0) {
+                    throw std::runtime_error("invalid MiniMax-H3 output head shape");
+                }
+                config.pdd_num_inference_steps = weight->ne[1] / patch_dim;
+                if (config.pdd_num_inference_steps != 1 && (config.audio_latent_channels != 0 || config.pdd_num_inference_steps != 6)) {
+                    throw std::runtime_error("unsupported MiniMax-H3 output head bank");
+                }
+            }
+            if (config.audio_latent_channels == 0 && config.pdd_num_inference_steps != 6) {
+                throw std::runtime_error("MiniMax-H3-Image requires a checkpoint with its six-step PDD heads included");
            }
            config.num_layers               = count_blocks(tensors, prefix + ".blocks.");
            config.token_refiner_num_layers = count_blocks(tensors, prefix + ".token_refiner.blocks.");
@@ -151,14 +173,16 @@ namespace MiniMaxH3 {
            }

            LOG_VERBOSE("minimax_h3: layers=%" PRId64 ", hidden=%" PRId64 ", heads=%" PRId64
-                        ", head_dim=%" PRId64 ", ffn=%" PRId64 ", adaln_curve=%" PRId64 ", two_time=%d",
+                        ", head_dim=%" PRId64 ", ffn=%" PRId64 ", adaln_curve=%" PRId64 ", two_time=%d, audio_latent_channels=%" PRId64 ", pdd_num_inference_steps=%" PRId64,
                        config.num_layers,
                        config.hidden_size,
                        config.num_attention_heads,
                        config.attention_head_dim,
                        config.ffn_hidden_size,
                        config.adaln_curve_grid,
-                        config.two_time);
+                        config.two_time,
+                        config.audio_latent_channels,
+                        config.pdd_num_inference_steps);
            return config;
        }
    };
@@ -624,6 +648,27 @@ namespace MiniMaxH3 {
        }
    };

+    static std::array<float, 6> pdd_head_weights(float sigma, float sigma_next) {
+        auto nearest = [](float value) {
+            size_t index = 0;
+            for (size_t i = 1; i < pdd_sigmas.size(); ++i) {
+                if (std::abs(pdd_sigmas[i] - value) < std::abs(pdd_sigmas[index] - value)) {
+                    index = i;
+                }
+            }
+            return index;
+        };
+        const size_t start = std::min(nearest(sigma), pdd_sigmas.size() - 2);
+        const size_t stop  = std::max(nearest(sigma_next), start + 1);
+        std::array<float, 6> weights{};
+        // Comfy checkpoints keep head zero as the base and store offsets in the remaining heads.
+        weights[0] = 1.f;
+        for (size_t i = std::max<size_t>(start, 1); i < stop; ++i) {
+            weights[i] = (pdd_sigmas[i] - pdd_sigmas[i + 1]) / (pdd_sigmas[start] - pdd_sigmas[stop]);
+        }
+        return weights;
+    }
+
    struct FinalLayer : public GGMLBlock {
        Config config;

@@ -637,19 +682,23 @@ namespace MiniMaxH3 {
                                                                            1,
                                                                            !config.uses_adaln_curves(),
                                                                            config.uses_adaln_curves());
-            blocks["video_out"]  = std::make_shared<Linear>(config.hidden_size, video_dim, true, true);
-            blocks["audio_out"]  = std::make_shared<Linear>(config.hidden_size, config.audio_latent_channels, true, true);
+            blocks["video_out"] =
+                std::make_shared<Linear>(config.hidden_size, video_dim * config.pdd_num_inference_steps, true, true);
+            if (config.audio_latent_channels > 0) {
+                blocks["audio_out"] = std::make_shared<Linear>(config.hidden_size, config.audio_latent_channels, true, true);
+            }
        }

        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
                                                      ggml_tensor* x,
                                                      ggml_tensor* t_emb,
                                                      const TokenModulationSpan& video,
-                                                      const TokenModulationSpan& audio) {
+                                                      const TokenModulationSpan& audio,
+                                                      float sigma,
+                                                      float sigma_next) {
            auto norm      = std::dynamic_pointer_cast<RMSNorm>(blocks["norm"]);
            auto adaln     = std::dynamic_pointer_cast<AdaLayerNormModulation>(blocks["adaln_proj"]);
            auto video_out = std::dynamic_pointer_cast<Linear>(blocks["video_out"]);
-            auto audio_out = std::dynamic_pointer_cast<Linear>(blocks["audio_out"]);
            auto mods      = adaln->forward(ctx, t_emb);
            auto apply     = [&](const TokenModulationSpan& segment) {
                auto row   = modulation_row(ctx, mods, config.hidden_size, 2, 1, segment.modulation_row);
@@ -658,8 +707,25 @@ namespace MiniMaxH3 {
                                ggml_add(ctx->ggml_ctx, value, ggml_mul(ctx->ggml_ctx, value, row[1])),
                                row[0]);
            };
-            return {video_out->forward(ctx, apply(video)),
-                    audio_out->forward(ctx, apply(audio))};
+            auto prediction = video_out->forward(ctx, apply(video));
+            if (config.pdd_num_inference_steps > 1) {
+                GGML_ASSERT(std::isfinite(sigma_next));
+                auto weights      = pdd_head_weights(sigma, sigma_next);
+                const int64_t dim = prediction->ne[0] / config.pdd_num_inference_steps;
+                auto blended      = ggml_ext_scale(ctx, ggml_ext_slice(ctx, prediction, 0, 0, dim), weights[0]);
+                for (int64_t i = 1; i < config.pdd_num_inference_steps; ++i) {
+                    if (weights[i] != 0.f) {
+                        auto offset = ggml_ext_slice(ctx, prediction, 0, i * dim, (i + 1) * dim);
+                        blended     = ggml_add(ctx->ggml_ctx, blended, ggml_ext_scale(ctx, offset, weights[i]));
+                    }
+                }
+                prediction = blended;
+            }
+            ggml_tensor* audio_prediction = nullptr;
+            if (config.audio_latent_channels > 0) {
+                audio_prediction = std::dynamic_pointer_cast<Linear>(blocks["audio_out"])->forward(ctx, apply(audio));
+            }
+            return {prediction, audio_prediction};
        }
    };

@@ -670,8 +736,10 @@ namespace MiniMaxH3 {
            : config(config) {
            int64_t video_dim          = config.video_latent_channels * config.patch_t * config.patch_h * config.patch_w;
            blocks["video_patch_proj"] = std::make_shared<Linear>(video_dim, config.hidden_size, true, true);
-            blocks["audio_patch_proj"] = std::make_shared<Linear>(config.audio_latent_channels, config.hidden_size, true, true);
-            blocks["condition_proj"]   = std::make_shared<Linear>(config.text_dim, config.hidden_size, true);
+            if (config.audio_latent_channels > 0) {
+                blocks["audio_patch_proj"] = std::make_shared<Linear>(config.audio_latent_channels, config.hidden_size, true, true);
+            }
+            blocks["condition_proj"] = std::make_shared<Linear>(config.text_dim, config.hidden_size, true);
            if (!config.uses_adaln_curves()) {
                blocks["time_embedder"] = std::make_shared<TimeEmbedder>(config.timestep_input_dim,
                                                                         config.time_embed_hidden_size,
@@ -788,9 +856,10 @@ namespace MiniMaxH3 {
                                                      const std::vector<TokenModulationSpan>& segments,
                                                      const std::vector<SequenceSegment>& sequence_segments,
                                                      const TokenModulationSpan& video_segment,
-                                                      const TokenModulationSpan& audio_segment) {
+                                                      const TokenModulationSpan& audio_segment,
+                                                      float sigma,
+                                                      float sigma_next) {
            auto video_proj = std::dynamic_pointer_cast<Linear>(blocks["video_patch_proj"]);
-            auto audio_proj = std::dynamic_pointer_cast<Linear>(blocks["audio_patch_proj"]);

            std::vector<std::pair<int64_t, int64_t>> condition_video_ranges;
            ggml_tensor* video_rows = nullptr;
@@ -836,11 +905,15 @@ namespace MiniMaxH3 {
                audio_offset += rows->ne[1];
                audio_rows = audio_rows == nullptr ? rows : ggml_concat(ctx->ggml_ctx, audio_rows, rows, 1);
            }
-            audio                                          = pack_audio_rows(audio);
-            std::pair<int64_t, int64_t> target_audio_range = {audio_offset, audio_offset + audio->ne[1]};
-            audio_rows                                     = audio_rows == nullptr ? audio : ggml_concat(ctx->ggml_ctx, audio_rows, audio, 1);
-            auto audio_embeds                              = audio_proj->forward(ctx, audio_rows);
-            context                                        = refine_context(ctx, context);
+            std::pair<int64_t, int64_t> target_audio_range{};
+            ggml_tensor* audio_embeds = nullptr;
+            if (audio != nullptr) {
+                audio              = pack_audio_rows(audio);
+                target_audio_range = {audio_offset, audio_offset + audio->ne[1]};
+                audio_rows         = audio_rows == nullptr ? audio : ggml_concat(ctx->ggml_ctx, audio_rows, audio, 1);
+                audio_embeds       = std::dynamic_pointer_cast<Linear>(blocks["audio_patch_proj"])->forward(ctx, audio_rows);
+            }
+            context = refine_context(ctx, context);

            ggml_tensor* h = nullptr;
            auto append    = [&](ggml_tensor* value) {
@@ -891,7 +964,7 @@ namespace MiniMaxH3 {
            }

            auto final_layer = std::dynamic_pointer_cast<FinalLayer>(blocks["final_layer"]);
-            auto output      = final_layer->forward(ctx, h, t_emb, video_segment, audio_segment);
+            auto output      = final_layer->forward(ctx, h, t_emb, video_segment, audio_segment, sigma, sigma_next);
            auto video_out   = DiT::unpatchify_3d(ctx,
                                                  output.first,
                                                  video->ne[2] / config.patch_t,
@@ -901,15 +974,18 @@ namespace MiniMaxH3 {
                                                  config.patch_h,
                                                  config.patch_w,
                                                  true);
-            auto audio_out   = ggml_reshape_4d(ctx->ggml_ctx,
-                                               output.second,
-                                               config.audio_latent_channels,
-                                               audio->ne[1] / 2,
-                                               2,
-                                               audio->ne[2]);
-            audio_out        = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx, audio_out, 1, 2, 0, 3));
            video_out        = ggml_ext_scale(ctx, video_out, -1.f);
-            audio_out        = ggml_ext_scale(ctx, audio_out, -1.f);
+            if (audio == nullptr) {
+                return {video_out, nullptr};
+            }
+            auto audio_out = ggml_reshape_4d(ctx->ggml_ctx,
+                                             output.second,
+                                             config.audio_latent_channels,
+                                             audio->ne[1] / 2,
+                                             2,
+                                             audio->ne[2]);
+            audio_out      = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx, audio_out, 1, 2, 0, 3));
+            audio_out      = ggml_ext_scale(ctx, audio_out, -1.f);
            return {video_out, audio_out};
        }
    };
@@ -995,7 +1071,7 @@ namespace MiniMaxH3 {
        int64_t run_start = 0;
        int current_tag   = text_tags.empty() ? 1 : text_tags[0];
        for (int64_t i = 1; i <= text_len; ++i) {
-            int tag = i < text_len && !text_tags.empty() ? text_tags[i] : -1;
+            int tag = i < text_len ? (text_tags.empty() ? 1 : text_tags[i]) : -1;
            if (i == text_len || tag != current_tag) {
                layout.segments.push_back({run_start,
                                           i,
@@ -1121,16 +1197,18 @@ namespace MiniMaxH3 {
            }
        }

-        int64_t audio_start = row;
-        append_audio_positions(audio_t, cursor, w_axis.front(), w_axis.back());
-        layout.audio_segment = {audio_start, row + audio_t * 2, audio_time_row};
-        layout.sequence_segments.push_back({audio_start,
-                                            row + audio_t * 2,
-                                            SequenceKind::TARGET_AUDIO});
-        layout.segments.push_back({audio_start,
-                                   row + audio_t * 2,
-                                   audio_time_row * 3 + 2});
-        row += audio_t * 2;
+        if (audio_t > 0) {
+            int64_t audio_start = row;
+            append_audio_positions(audio_t, cursor, w_axis.front(), w_axis.back());
+            layout.audio_segment = {audio_start, row + audio_t * 2, audio_time_row};
+            layout.sequence_segments.push_back({audio_start,
+                                                row + audio_t * 2,
+                                                SequenceKind::TARGET_AUDIO});
+            layout.segments.push_back({audio_start,
+                                       row + audio_t * 2,
+                                       audio_time_row * 3 + 2});
+            row += audio_t * 2;
+        }

        int64_t video_start = row;
        for (int64_t t = 0; t < latent_t; ++t) {
@@ -1232,26 +1310,29 @@ namespace MiniMaxH3 {
                                 float audio_shift,
                                 float video_sigma_next) {
            GGML_ASSERT(packed.dim() == 4 || packed.dim() == 5);
-            GGML_ASSERT(audio_length > 0);
+            GGML_ASSERT(audio_length > 0 || config.audio_latent_channels == 0);
            GGML_ASSERT(!context_tensor.empty());
            const int64_t spatial      = packed.shape()[0] * packed.shape()[1] * packed.shape()[2];
            const int64_t video_values = spatial * config.video_latent_channels;
            const int64_t audio_values = int64_t(audio_length) * 2 * config.audio_latent_channels;
            GGML_ASSERT(packed.numel() >= video_values + audio_values);
-            auto packed_input  = make_input(packed);
-            auto video         = ggml_view_4d(compute_ctx, packed_input,
-                                              packed.shape()[0], packed.shape()[1], packed.shape()[2],
-                                              config.video_latent_channels,
-                                              packed.shape()[0] * sizeof(float),
-                                              packed.shape()[0] * packed.shape()[1] * sizeof(float),
-                                              spatial * sizeof(float), 0);
-            auto audio_carrier = ggml_view_4d(compute_ctx, packed_input,
-                                              audio_length, 2, config.audio_latent_channels, 1,
-                                              audio_length * sizeof(float),
-                                              int64_t(audio_length) * 2 * sizeof(float),
-                                              audio_values * sizeof(float),
-                                              video_values * sizeof(float));
-            auto context       = make_input(context_tensor);
+            auto packed_input          = make_input(packed);
+            auto video                 = ggml_view_4d(compute_ctx, packed_input,
+                                                      packed.shape()[0], packed.shape()[1], packed.shape()[2],
+                                                      config.video_latent_channels,
+                                                      packed.shape()[0] * sizeof(float),
+                                                      packed.shape()[0] * packed.shape()[1] * sizeof(float),
+                                                      spatial * sizeof(float), 0);
+            ggml_tensor* audio_carrier = nullptr;
+            if (config.audio_latent_channels > 0) {
+                audio_carrier = ggml_view_4d(compute_ctx, packed_input,
+                                             audio_length, 2, config.audio_latent_channels, 1,
+                                             audio_length * sizeof(float),
+                                             int64_t(audio_length) * 2 * sizeof(float),
+                                             audio_values * sizeof(float),
+                                             video_values * sizeof(float));
+            }
+            auto context = make_input(context_tensor);
            std::vector<ggml_tensor*> condition_inputs;
            condition_inputs.reserve(condition_videos.size());
            for (const auto& condition : condition_videos) {
@@ -1282,7 +1363,7 @@ namespace MiniMaxH3 {
            // The sampler carries c_a = (sigma_v / sigma_a) * x_a so the packed
            // latent follows one sigma schedule. Restore x_a for the H3 network.
            auto audio_ctx = get_context();
-            auto audio     = ggml_ext_scale(&audio_ctx, audio_carrier, sigma_a / sigma_v);
+            auto audio     = audio_carrier == nullptr ? nullptr : ggml_ext_scale(&audio_ctx, audio_carrier, sigma_a / sigma_v);
            auto layout    = build_layout(context_tensor.shape()[1],
                                          packed.shape()[2],
                                          packed.shape()[1],
@@ -1370,14 +1451,18 @@ namespace MiniMaxH3 {
                                            layout.segments,
                                            layout.sequence_segments,
                                            layout.video_segment,
-                                            layout.audio_segment);
+                                            layout.audio_segment,
+                                            sigma_v,
+                                            video_sigma_next);
            // Convert the model's audio velocity to d(c_a) / d(sigma_v).
-            output.second = ggml_add(compute_ctx,
-                                     ggml_ext_scale(&runner_ctx, audio, 1.f - audio_scale),
-                                     ggml_ext_scale(&runner_ctx,
-                                                    output.second,
-                                                    1.f + (audio_scale - 1.f) * sigma_a));
-            auto merged   = merge_av_latents(&runner_ctx, output.first, output.second);
+            if (output.second != nullptr) {
+                output.second = ggml_add(compute_ctx,
+                                         ggml_ext_scale(&runner_ctx, audio, 1.f - audio_scale),
+                                         ggml_ext_scale(&runner_ctx,
+                                                        output.second,
+                                                        1.f + (audio_scale - 1.f) * sigma_a));
+            }
+            auto merged = output.second == nullptr ? output.first : merge_av_latents(&runner_ctx, output.first, output.second);
            ggml_build_forward_expand(graph, merged);
            return graph;
        }
@@ -41,6 +41,15 @@ const std::vector<ModelDefinition>& minimax_h3_definitions() {
            definition.build                            = build_minimax_h3;
            result.push_back(std::move(definition));
        }
+        {
+            ModelDefinition definition                  = result.front();
+            definition.version                          = VERSION_MINIMAX_H3_IMAGE;
+            definition.name                             = "MiniMax-H3-Image";
+            definition.traits.supports_video_generation = false;
+            definition.sampling.default_flow_shift      = 1.f;
+            definition.ref_image_preset                 = "qwen";
+            result.push_back(std::move(definition));
+        }
        {
            ModelDefinition definition                  = result.front();
            definition.version                          = VERSION_P_H3_IMAGE;
@@ -26,14 +26,19 @@ namespace MiniMaxH3VAE {
    }

    struct MiniMaxH3VAEConfig {
-        int num_layers   = 36;
-        int patch_size_t = 4;
-        bool sol         = false;
+        int num_layers      = 36;
+        int patch_size_t    = 4;
+        int hidden_size     = 2048;
+        bool sol            = false;
+        bool vit_decoder_2d = false;

        static MiniMaxH3VAEConfig detect_from_weights(const String2TensorStorage& tensors, const std::string& prefix) {
            MiniMaxH3VAEConfig config;
            config.sol = tensors.count(prefix + ".decoder.blocks.0.scale1") != 0;
            auto proj  = tensors.find(prefix + ".decoder.proj_out.weight");
+            if (const auto embedder = tensors.find(prefix + ".decoder.x_embedder.weight"); embedder != tensors.end()) {
+                config.hidden_size = static_cast<int>(embedder->second.ne[1]);
+            }
            if (proj != tensors.end()) {
                config.patch_size_t = static_cast<int>(proj->second.ne[1] / (3 * 16 * 16));
                if (config.patch_size_t != 1 && config.patch_size_t != 4) {
@@ -50,11 +55,13 @@ namespace MiniMaxH3VAE {
                    config.num_layers = layers;
                }
            }
+            config.vit_decoder_2d = config.patch_size_t == 1 && config.hidden_size == 1024;
            if (config.sol || proj != tensors.end()) {
-                LOG_VERBOSE("minimax_h3_vae: decoder=%s, layers=%d, patch_size_t=%d",
+                LOG_VERBOSE("minimax_h3_vae: decoder=%s, layers=%d, patch_size_t=%d, hidden_size=%d",
                            config.sol ? "sol" : "original",
                            config.num_layers,
-                            config.patch_size_t);
+                            config.patch_size_t,
+                            config.hidden_size);
            }
            return config;
        }
@@ -287,11 +294,12 @@ namespace MiniMaxH3VAE {
    }

    struct DecoderAttention : public GGMLBlock {
-        static constexpr int num_head = 32;
        static constexpr int head_dim = 64;
-        static constexpr int dim      = num_head * head_dim;
+        int dim;
+        int num_head;

-        DecoderAttention() {
+        explicit DecoderAttention(int dim = 2048)
+            : dim(dim), num_head(dim / head_dim) {
            blocks["to_qkv"] = std::make_shared<Linear>(dim, dim * 3, true);
            blocks["to_out"] = std::make_shared<Linear>(dim, dim, true);
        }
@@ -366,26 +374,23 @@ namespace MiniMaxH3VAE {
                };
                return to_out->forward(ctx, sol->forward(ctx, x, pe, num_head, project_rows));
            }
-            auto [q, k, v]      = project(ctx, norm.forward(ctx, x), pe);
-            auto out            = ggml_ext_attention_ext(ctx,
-                                                         q,
-                                                         k,
-                                                         v,
-                                                         num_head,
-                                                         nullptr,
-                                                         true,
-                                                         ctx->flash_attn_enabled);
+            auto [q, k, v] = project(ctx, norm.forward(ctx, x), pe);
+            auto out       = ggml_ext_attention_ext(ctx,
+                                                    q,
+                                                    k,
+                                                    v,
+                                                    num_head,
+                                                    nullptr,
+                                                    true,
+                                                    ctx->flash_attn_enabled);
            return to_out->forward(ctx, out);
        }
    };

    struct DecoderFeedForward : public GGMLBlock {
-        static constexpr int dim       = 2048;
-        static constexpr int kInnerDim = dim * 4;
-
-        DecoderFeedForward() {
-            blocks["w1"] = std::make_shared<Linear>(dim, kInnerDim * 2, true);
-            blocks["w2"] = std::make_shared<Linear>(kInnerDim, dim, true);
+        explicit DecoderFeedForward(int dim = 2048) {
+            blocks["w1"] = std::make_shared<Linear>(dim, dim * 8, true);
+            blocks["w2"] = std::make_shared<Linear>(dim * 4, dim, true);
        }

        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
@@ -397,13 +402,14 @@ namespace MiniMaxH3VAE {
    };

    struct DecoderBlock : public GGMLBlock {
-        static constexpr int dim = 2048;
+        int dim;

-        DecoderBlock() {
+        explicit DecoderBlock(int dim = 2048)
+            : dim(dim) {
            blocks["norm1"] = std::make_shared<RMSNorm>(dim, 1e-5f);
-            blocks["attn"]  = std::make_shared<DecoderAttention>();
+            blocks["attn"]  = std::make_shared<DecoderAttention>(dim);
            blocks["norm2"] = std::make_shared<RMSNorm>(dim, 1e-5f);
-            blocks["ff"]    = std::make_shared<DecoderFeedForward>();
+            blocks["ff"]    = std::make_shared<DecoderFeedForward>(dim);
        }

        void init_params(ggml_context* ctx,
@@ -420,10 +426,10 @@ namespace MiniMaxH3VAE {
                             ggml_tensor* pe,
                             const SolAttentionPlan::Inputs* sol,
                             int64_t rows_per_chunk) {
-            auto norm1 = std::dynamic_pointer_cast<RMSNorm>(blocks["norm1"]);
-            auto attn  = std::dynamic_pointer_cast<DecoderAttention>(blocks["attn"]);
-            auto norm2 = std::dynamic_pointer_cast<RMSNorm>(blocks["norm2"]);
-            auto ff    = std::dynamic_pointer_cast<DecoderFeedForward>(blocks["ff"]);
+            auto norm1  = std::dynamic_pointer_cast<RMSNorm>(blocks["norm1"]);
+            auto attn   = std::dynamic_pointer_cast<DecoderAttention>(blocks["attn"]);
+            auto norm2  = std::dynamic_pointer_cast<RMSNorm>(blocks["norm2"]);
+            auto ff     = std::dynamic_pointer_cast<DecoderFeedForward>(blocks["ff"]);
            x           = ggml_add(ctx->ggml_ctx,
                                   x,
                                   ggml_mul(ctx->ggml_ctx,
@@ -449,7 +455,7 @@ namespace MiniMaxH3VAE {
    };

    struct Decoder : public GGMLBlock {
-        static constexpr int dim = 2048;
+        int dim;
        int num_layers;
        static constexpr int num_register_tokens = 4;
        static constexpr int patch_size          = 16;
@@ -457,11 +463,11 @@ namespace MiniMaxH3VAE {
        bool sol;

        explicit Decoder(const MiniMaxH3VAEConfig& config)
-            : num_layers(config.num_layers), patch_size_t(config.patch_size_t), sol(config.sol) {
+            : dim(config.hidden_size), num_layers(config.num_layers), patch_size_t(config.patch_size_t), sol(config.sol) {
            blocks["x_embedder"] = std::make_shared<Linear>(24, dim, true);
            for (int i = 0; i < num_layers; ++i) {
                blocks[(sol ? "blocks." : "transformer_blocks.") + std::to_string(i)] =
-                    std::make_shared<DecoderBlock>();
+                    std::make_shared<DecoderBlock>(dim);
            }
            blocks["norm_out"] = std::make_shared<LayerNorm>(dim, 1e-5f, true, true);
            blocks["proj_out"] = std::make_shared<Linear>(dim,
@@ -574,15 +580,19 @@ namespace MiniMaxH3VAE {

    struct MiniMaxH3VideoVAE : public GGMLBlock {
        bool single_frame;
+        bool vit_decoder_2d;

        explicit MiniMaxH3VideoVAE(const MiniMaxH3VAEConfig& config)
-            : single_frame(config.patch_size_t == 1) {
+            : single_frame(config.patch_size_t == 1), vit_decoder_2d(config.vit_decoder_2d) {
            blocks["encoder"]         = std::make_shared<Encoder>();
            blocks["quant_conv"]      = std::make_shared<Conv3d>(48, 48, std::tuple{1, 1, 1});
            blocks["post_quant_conv"] = std::make_shared<Conv3d>(24,
                                                                 24,
                                                                 std::tuple{1, 1, 1});
-            blocks["decoder"]         = std::make_shared<Decoder>(config);
+            if (vit_decoder_2d) {
+                blocks["post_quant_conv"] = std::make_shared<Conv2d>(24, 24, std::pair{1, 1});
+            }
+            blocks["decoder"] = std::make_shared<Decoder>(config);
        }

        ggml_tensor* encode(GGMLRunnerContext* ctx,
@@ -605,9 +615,17 @@ namespace MiniMaxH3VAE {
                            const std::array<SolAttentionPlan::Inputs, 2>& sol_inputs,
                            ggml_tensor* pixel_mean,
                            ggml_tensor* pixel_std) {
-            auto post_quant = std::dynamic_pointer_cast<Conv3d>(blocks["post_quant_conv"]);
+            auto post_quant = std::dynamic_pointer_cast<UnaryBlock>(blocks["post_quant_conv"]);
            auto decoder    = std::dynamic_pointer_cast<Decoder>(blocks["decoder"]);
-            return decoder->forward(ctx, post_quant->forward(ctx, latent), pe, shifted_pe, sol_inputs, pixel_mean, pixel_std);
+            if (vit_decoder_2d) {
+                GGML_ASSERT(latent->ne[2] == 1);
+                latent = ggml_reshape_4d(ctx->ggml_ctx, latent, latent->ne[0], latent->ne[1], 24, latent->ne[3] / 24);
+            }
+            latent = post_quant->forward(ctx, latent);
+            if (vit_decoder_2d) {
+                latent = ggml_reshape_4d(ctx->ggml_ctx, latent, latent->ne[0], latent->ne[1], 1, latent->ne[2] * latent->ne[3]);
+            }
+            return decoder->forward(ctx, latent, pe, shifted_pe, sol_inputs, pixel_mean, pixel_std);
        }
    };

@@ -651,6 +669,9 @@ namespace MiniMaxH3VAE {
            if (config.sol) {
                return "sol_h3_vae";
            }
+            if (config.vit_decoder_2d) {
+                return "minimax_h3_image_vae";
+            }
            return config.patch_size_t == 1 ? "p_h3_image_vae" : "minimax_h3_video_vae";
        }

@@ -759,7 +780,7 @@ namespace MiniMaxH3VAE {
            auto input  = ensure_video_shape(x);
            auto tiling = resolve_original_tiling_params(tiling_params);
            if (config.patch_size_t == 1 && input.shape()[2] != 1) {
-                LOG_ERROR("P-H3-Image VAE encodes single-frame images only");
+                LOG_ERROR("H3 image VAE encodes single-frame images only");
                return {};
            }
            if (input.shape()[2] == 1) {
@@ -865,7 +886,7 @@ namespace MiniMaxH3VAE {
            }
            auto tiling = resolve_tiling_params(tiling_params);
            if (config.patch_size_t == 1 && input.shape()[2] != 1) {
-                LOG_ERROR("P-H3-Image VAE decodes single-frame latents only");
+                LOG_ERROR("H3 image VAE decodes single-frame latents only");
                return {};
            }
            if (input.shape()[2] == 1) {
@@ -966,7 +987,7 @@ namespace MiniMaxH3VAE {
            std::vector<std::vector<float>> ids;
            const int64_t global_rows = config.sol ? ((num_frames + sol_block_size - 1) / sol_block_size) * sol_global_rows : 5;
            ids.reserve(static_cast<size_t>(width * height * num_frames + global_rows));
-            constexpr float two_pi = 6.28318530717958647692f;
+            constexpr float two_pi      = 6.28318530717958647692f;
            const int64_t spatial_shift = shifted ? sol_spatial_shift : 0;
            const int64_t time_shift    = shifted ? sol_temporal_shift : 0;
            for (int64_t t = 0; t < num_frames; ++t) {
@@ -979,20 +1000,20 @@ namespace MiniMaxH3VAE {
                    for (int64_t w = 0; w < width; ++w) {
                        const float x = config.sol ? ((w + spatial_shift) % sol_window_size + 0.5f) / sol_window_size : (w + 0.5f) / width;
                        float pw      = (2.f * x - 1.f) * two_pi;
-                        ids.push_back({pt, ph, pw});
+                        ids.push_back(config.vit_decoder_2d ? std::vector<float>{ph, pw} : std::vector<float>{pt, ph, pw});
                    }
                }
            }
            for (int64_t i = 0; i < global_rows; ++i) {
-                ids.push_back({0.f, 0.f, 0.f});
+                ids.push_back(std::vector<float>(config.vit_decoder_2d ? 2 : 3, 0.f));
            }
            auto values = Rope::embed_nd(ids,
                                         1,
                                         100.f,
-                                         std::vector<int>{16, 16, 16});
+                                         config.vit_decoder_2d ? std::vector<int>{16, 16} : std::vector<int>{16, 16, 16});
            return sd::Tensor<float>({2,
                                      2,
-                                      24,
+                                      config.vit_decoder_2d ? 16 : 24,
                                      static_cast<int64_t>(ids.size())},
                                     std::move(values));
        }
@@ -577,7 +577,10 @@ SDVersion ModelLoader::get_sd_version() const {
            return VERSION_LTXAV;
        }
        if (tensor_storage.name.find("model.diffusion_model.video_patch_proj.weight") != std::string::npos &&
-            tensor_storage_map.find("model.diffusion_model.audio_patch_proj.weight") != tensor_storage_map.end()) {
+            tensor_storage_map.find("model.diffusion_model.condition_proj.weight") != tensor_storage_map.end()) {
+            if (tensor_storage_map.count("model.diffusion_model.audio_patch_proj.weight") == 0) {
+                return VERSION_MINIMAX_H3_IMAGE;
+            }
            // P-H3 uses the same DiT architecture as H3; its VAE or connector identifies the image pipeline.
            auto decoder = tensor_storage_map.find("first_stage_model.decoder.proj_out.weight");
            if (decoder == tensor_storage_map.end()) {
@@ -41,6 +41,7 @@
 #include "model/audio/wav2vec2.hpp"
 #include "model/diffusion/animatediff.hpp"
 #include "model/diffusion/control.hpp"
+#include "model/diffusion/minimax_h3.hpp"
 #include "model/diffusion/model.hpp"
 #include "model/diffusion/z_image.hpp"
 #include "model/vae/audio_vae.hpp"
@@ -1492,6 +1493,11 @@ bool StableDiffusionGGML::build_denoiser() {
            break;
        }
        case H3_AV_FLOW_PRED: {
+            if (version == VERSION_MINIMAX_H3_IMAGE) {
+                LOG_INFO("running in MiniMax H3 Image FLOW mode");
+                denoiser = std::make_shared<MiniMaxH3ImageDenoiser>(MiniMaxH3::pdd_sigmas);
+                break;
+            }
            LOG_INFO("running in MiniMax H3 AV FLOW mode");
            denoiser = std::make_shared<H3AVFlowDenoiser>(default_flow_shift, version == VERSION_P_H3_IMAGE ? 1.f : 3.f, get_latent_channel());
            break;
@@ -2615,7 +2621,7 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
                    condition.c_position_ids.empty() ? nullptr : &condition.c_position_ids,
                    condition.c_ref_audios.empty() ? nullptr : &condition.c_ref_audios,
                    condition.c_reference_blocks.empty() ? nullptr : &condition.c_reference_blocks,
-                    version == VERSION_P_H3_IMAGE ? 2 : audio_length,
+                    version == VERSION_MINIMAX_H3_IMAGE ? 0 : (version == VERSION_P_H3_IMAGE ? 2 : audio_length),
                    std::isfinite(active_flow_shift) ? active_flow_shift : 12.f,
                    version == VERSION_P_H3_IMAGE ? 1.f : 3.f,
                    target_index > 0 && target_index < sigmas.size() ? sigmas[target_index] : NAN};
@@ -3054,7 +3060,7 @@ sd::Tensor<float> StableDiffusionGGML::decode_first_stage(const sd::Tensor<float
            return {};
        }
        auto decoded = first_stage_model->decode(n_threads, latents, tiling_params, decode_video, circular_x, circular_y);
-        if (version == VERSION_P_H3_IMAGE && !decoded.empty() && decoded.dim() == 5) {
+        if (sd_version_is_minimax_h3_image(version) && !decoded.empty() && decoded.dim() == 5) {
            decoded = decoded.squeeze(2);
        }
        if (!decoded.empty() ||
@@ -172,6 +172,17 @@ namespace sd::pipeline {
        return packed;
    }

+    static sd::Tensor<float> resize_h3_image_reference(const sd::Tensor<float>& image, int64_t max_short_edge) {
+        const int64_t short_edge = std::min(image.shape()[0], image.shape()[1]);
+        const double scale       = static_cast<double>(std::clamp<int64_t>(short_edge, 1024, max_short_edge)) / short_edge;
+        const int64_t width      = std::max<int64_t>(32, static_cast<int64_t>(std::nearbyint(image.shape()[0] * scale / 32)) * 32);
+        const int64_t height     = std::max<int64_t>(32, static_cast<int64_t>(std::nearbyint(image.shape()[1] * scale / 32)) * 32);
+        if (width == image.shape()[0] && height == image.shape()[1]) {
+            return image;
+        }
+        return sd::ops::interpolate(image, {width, height, image.shape()[2], image.shape()[3]}, sd::ops::InterpolateMode::Lanczos);
+    }
+
    static std::optional<ImageGenerationLatents> prepare_image_generation_latents(StableDiffusionGGML* sd,
                                                                                  const sd_img_gen_params_t* sd_img_gen_params,
                                                                                  GenerationRequest* request,
@@ -354,7 +365,20 @@ namespace sd::pipeline {
                continue;
            }
            sd::Tensor<float> ref_latent;
-            if (sd->version == VERSION_P_H3_IMAGE) {
+            if (sd->version == VERSION_MINIMAX_H3_IMAGE) {
+                ref_latent = sd->encode_first_stage(resize_h3_image_reference(ref_images[i], 2048));
+                if (!ref_latent.empty()) {
+                    const int64_t width    = ref_latent.shape()[0];
+                    const int64_t height   = ref_latent.shape()[1];
+                    const int64_t channels = ref_latent.shape()[3];
+                    auto condition_rng     = std::make_shared<MT19937RNG>(static_cast<uint64_t>(request->seed));
+                    // Upstream augments patch rows, so the RNG follows token/channel/patch order.
+                    auto noise = sd::randn<float>({2, 2, channels, width / 2, height / 2}, condition_rng)
+                                     .permute({0, 3, 1, 4, 2})
+                                     .reshape(ref_latent.shape());
+                    ref_latent = ref_latent * 0.999f + noise * (1.f - 0.999f);
+                }
+            } else if (sd->version == VERSION_P_H3_IMAGE) {
                auto resized = resize_ph3_reference(ref_images[i], int64_t(request->width) * request->height);
                ref_latent   = sd->encode_first_stage(resized);
            } else if (ref_image_params.resize_before_vae && !sd_version_is_pid(sd->version)) {
@@ -484,14 +508,18 @@ namespace sd::pipeline {
        }

        ImageGenerationLatents latents;
-        if (sd->version == VERSION_P_H3_IMAGE) {
-            init_latent = pack_ph3_image_latent(init_latent);
+        if (sd_version_is_minimax_h3_image(sd->version)) {
+            if (sd->version == VERSION_P_H3_IMAGE) {
+                init_latent = pack_ph3_image_latent(init_latent);
+            } else if (init_latent.dim() == 4) {
+                init_latent = init_latent.unsqueeze(2);
+            }
            latent_mask = latent_mask.unsqueeze(2);
            for (size_t i = 0; i < ref_latents.size(); ++i) {
                latents.minimax_reference_blocks.push_back({MiniMaxH3ReferenceKind::IMAGE, static_cast<int32_t>(i), -1});
            }
            for (auto& image : ref_images) {
-                image = resize_ph3_reference(image, 1000000);
+                image = sd->version == VERSION_P_H3_IMAGE ? resize_ph3_reference(image, 1000000) : resize_h3_image_reference(image, 1024);
            }
        }
        latents.init_latent              = std::move(init_latent);
@@ -548,7 +576,7 @@ namespace sd::pipeline {
        if (cond.c_concat.empty() && ref_image_params.pass_to_dit) {
            cond.c_concat = latents->concat_latent;  // TODO: optimize
        }
-        if (sd->version == VERSION_P_H3_IMAGE && ref_image_params.pass_to_dit) {
+        if (sd_version_is_minimax_h3_image(sd->version) && ref_image_params.pass_to_dit) {
            cond.c_reference_blocks = latents->minimax_reference_blocks;
        }

@@ -596,7 +624,7 @@ namespace sd::pipeline {
            if (uncond.c_concat.empty() && ref_image_params.pass_to_dit) {
                uncond.c_concat = latents->concat_latent;  // TODO: optimize
            }
-            if (sd->version == VERSION_P_H3_IMAGE && ref_image_params.pass_to_dit) {
+            if (sd_version_is_minimax_h3_image(sd->version) && ref_image_params.pass_to_dit) {
                uncond.c_reference_blocks = latents->minimax_reference_blocks;
            }
        }
@@ -159,6 +159,14 @@ namespace sd::model_builders {
                                                     version,
                                                     weight_manager);
            } else if (find_model_traits(vae_version)->vae_kind == ModelVAEKind::MINIMAX_H3) {
+                if (version == VERSION_MINIMAX_H3_IMAGE) {
+                    auto proj     = tensor_storage_map.find("first_stage_model.decoder.proj_out.weight");
+                    auto embedder = tensor_storage_map.find("first_stage_model.decoder.x_embedder.weight");
+                    if (proj == tensor_storage_map.end() || proj->second.ne[1] != 768 ||
+                        embedder == tensor_storage_map.end() || embedder->second.ne[1] != 1024) {
+                        throw std::runtime_error("MiniMax-H3-Image requires its 2D image VAE; pass --vae minimax_h3_image_vae_fp16.safetensors");
+                    }
+                }
                if (version == VERSION_P_H3_IMAGE) {
                    // The video VAE also loads successfully, but uses different temporal decoding and posterior sampling.
                    auto proj = tensor_storage_map.find("first_stage_model.decoder.proj_out.weight");
@@ -2,6 +2,7 @@
 #define __SD_RUNTIME_DENOISER_HPP__

 #include <algorithm>
+#include <array>
 #include <cctype>
 #include <cmath>
 #include <cstring>
@@ -1405,6 +1406,38 @@ struct MingImageFlowDenoiser : DiscreteFlowDenoiser {
    }
 };

+struct MiniMaxH3ImageDenoiser : public DiscreteFlowDenoiser {
+    std::array<float, 7> pdd_sigmas;
+
+    explicit MiniMaxH3ImageDenoiser(const std::array<float, 7>& pdd_sigmas)
+        : DiscreteFlowDenoiser(1.f), pdd_sigmas(pdd_sigmas) {}
+
+    float sigma_min() override { return 0.f; }
+    float sigma_max() override { return 1.f; }
+
+    float t_to_sigma(float t) override {
+        const float position = (1.f - std::clamp(t / (TIMESTEPS - 1.f), 0.f, 1.f)) * 6.f;
+        const size_t index   = std::min<size_t>(static_cast<size_t>(position), 5);
+        const auto& grid     = pdd_sigmas;
+        return grid[index] + (position - index) * (grid[index + 1] - grid[index]);
+    }
+
+    std::vector<float> get_sigmas(uint32_t n, int image_seq_len, scheduler_t scheduler, SDVersion version, const char* extra_sample_args = nullptr) override {
+        if (scheduler != DISCRETE_SCHEDULER) {
+            return DiscreteFlowDenoiser::get_sigmas(n, image_seq_len, scheduler, version, extra_sample_args);
+        }
+        if (n == 0) {
+            return {};
+        }
+        std::vector<float> sigmas;
+        sigmas.reserve(n + 1);
+        for (uint32_t i = 0; i <= n; ++i) {
+            sigmas.push_back(t_to_sigma((1.f - static_cast<float>(i) / n) * (TIMESTEPS - 1.f)));
+        }
+        return sigmas;
+    }
+};
+
 struct H3AVFlowDenoiser : public DiscreteFlowDenoiser {
    int64_t video_channels;
    float audio_shift;