Support Hunyuan3d (#150)

* add memory predict model by lr

* 3d

* fix hunyuan bug

* ruff format

* Update diffsynth_engine/models/hunyuan3d/dino_image_encoder.py

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* Update diffsynth_engine/utils/memory/memory_predcit_model.py

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* bug fix

* Update diffsynth_engine/pipelines/hunyuan3d_shape.py

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

---------

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

Author: tenderness-git

diffsynth_engine/__init__.py

@@ -17,6 +17,7 @@
     SDImagePipeline,
     WanVideoPipeline,
     QwenImagePipeline,
+    Hunyuan3DShapePipeline,
 )
 from .models.flux import FluxControlNet, FluxIPAdapter, FluxRedux
 from .models.sd import SDControlNet
@@ -57,6 +58,7 @@
     "SDImagePipeline",
     "WanVideoPipeline",
     "QwenImagePipeline",
+    "Hunyuan3DShapePipeline",
     "FluxInpaintingTool",
     "FluxOutpaintingTool",
     "FluxIPAdapterRefTool",

diffsynth_engine/algorithm/noise_scheduler/flow_match/recifited_flow.py

@@ -1,7 +1,11 @@
 import torch
 import math
 
-from diffsynth_engine.algorithm.noise_scheduler.base_scheduler import append_zero, BaseScheduler
+from diffsynth_engine.algorithm.noise_scheduler.base_scheduler import BaseScheduler
+
+
+def append(x, value):
+    return torch.cat([x, x.new_ones([1]) * value])
 
 
 class RecifitedFlowScheduler(BaseScheduler):
@@ -37,6 +41,7 @@ def schedule(
         mu: float | None = None,
         sigma_min: float | None = None,
         sigma_max: float | None = None,
+        append_value: float = 0,
     ):
         sigma_min = self.sigma_min if sigma_min is None else sigma_min
         sigma_max = self.sigma_max if sigma_max is None else sigma_max
@@ -46,5 +51,5 @@ def schedule(
         else:
             sigmas = self._shift_sigma(sigmas, self.shift)
         timesteps = sigmas * self.num_train_timesteps
-        sigmas = append_zero(sigmas)
+        sigmas = append(sigmas, append_value)
         return sigmas, timesteps

diffsynth_engine/algorithm/sampler/flow_match/flow_match_euler.py

@@ -2,19 +2,14 @@
 
 
 class FlowMatchEulerSampler:
-    def initialize(self, init_latents, timesteps, sigmas, mask=None):
-        self.init_latents = init_latents
-        self.timesteps = timesteps
+    def initialize(self, sigmas):
         self.sigmas = sigmas
-        self.mask = mask
 
     def step(self, latents, model_outputs, i):
         dt = self.sigmas[i + 1] - self.sigmas[i]
         latents = latents.to(dtype=torch.float32)
         latents = latents + model_outputs * dt
         latents = latents.to(dtype=model_outputs.dtype)
-        if self.mask is not None:
-            latents = latents * self.mask + self.init_latents * (1 - self.mask)
         return latents
 
     def add_noise(self, latents, noise, sigma):

diffsynth_engine/configs/pipeline.py

@@ -184,6 +184,16 @@ def __post_init__(self):
         init_parallel_config(self)
 
 
+@dataclass
+class HunyuanPipelineConfig(BaseConfig):
+    model_path: str | os.PathLike | List[str | os.PathLike]
+    model_dtype: torch.dtype = torch.float16
+    vae_path: Optional[str | os.PathLike | List[str | os.PathLike]] = None
+    vae_dtype: torch.dtype = torch.float16
+    image_encoder_path: Optional[str | os.PathLike | List[str | os.PathLike]] = None
+    image_encoder_dtype: torch.dtype = torch.float16
+
+
 @dataclass
 class QwenImagePipelineConfig(AttentionConfig, OptimizationConfig, ParallelConfig, BaseConfig):
     model_path: str | os.PathLike | List[str | os.PathLike]

diffsynth_engine/models/basic/attention.py

@@ -111,7 +111,7 @@ def attention(
     q,
     k,
     v,
-    attn_impl: Optional[str] = None,
+    attn_impl: Optional[str] = "auto",
     attn_mask: Optional[torch.Tensor] = None,
     scale: Optional[float] = None,
     **kwargs,

diffsynth_engine/models/hunyuan3d/__init__.py

@@ -0,0 +1,9 @@
+from .hunyuan3d_dit import HunYuan3DDiT
+from .hunyuan3d_vae import ShapeVAEDecoder
+from .dino_image_encoder import ImageEncoder
+
+__all__ = [
+    "HunYuan3DDiT",
+    "ShapeVAEDecoder",
+    "ImageEncoder",
+]