VRAM requirements and quantization

[0xPD33]

Hey! First of all, great release. I'm able to generate models just fine on my mobile 5080 but unfortunately lack the VRAM for texture painting.

Tried to quantize using bitsandbytes and using xformers but no luck so far. CPU offloading is obviously way too slow.

Any pointers to getting this to work would be greatly appreciated.

Thanks!

[DenisKochetov]

Hey, did you try mmgp for offloading?
https://github.com/deepbeepmeep/mmgp
https://github.com/deepbeepmeep/Hunyuan3D-2GP

[0xPD33]

I just tried it and it works! Pretty fast too. Thank you so much!

For anyone interested here is the adapted demo.py which I used to get this to work (after running pip install mmgp). You may have to adapt the profile based on your system configuration:

import sys
sys.path.insert(0, './hy3dshape')
sys.path.insert(0, './hy3dpaint')

from mmgp import offload, profile_type

from PIL import Image
from hy3dshape.rembg import BackgroundRemover
from hy3dshape.pipelines import Hunyuan3DDiTFlowMatchingPipeline

from textureGenPipeline import Hunyuan3DPaintPipeline, Hunyuan3DPaintConfig

try:
    from torchvision_fix import apply_fix
    apply_fix()
except ImportError:
    print("Warning: torchvision_fix module not found, proceeding without compatibility fix")      
except Exception as e:
    print(f"Warning: Failed to apply torchvision fix: {e}")

model_path = 'tencent/Hunyuan3D-2.1'
print("Loading shape-generation pipeline... (GPU) ")
pipeline_shapegen = Hunyuan3DDiTFlowMatchingPipeline.from_pretrained(model_path)

image_path = 'assets/demo.png'
image = Image.open(image_path).convert("RGBA")
if image.mode == 'RGB':
    rembg = BackgroundRemover()
    image = rembg(image)

mesh = pipeline_shapegen(image=image)[0]
mesh.export('demo.glb')

max_num_view = 6  # can be 6 to 9
resolution = 512  # can be 768 or 512
conf = Hunyuan3DPaintConfig(max_num_view, resolution)

# Make sure the heavy diffusion pipeline is kept on CPU
conf.device = "cpu"

conf.realesrgan_ckpt_path = "hy3dpaint/ckpt/RealESRGAN_x4plus.pth"
conf.multiview_cfg_path = "hy3dpaint/cfgs/hunyuan-paint-pbr.yaml"
conf.custom_pipeline = "hy3dpaint/hunyuanpaintpbr/pipeline.py"

print("Instantiating paint pipeline...")
paint_pipeline = Hunyuan3DPaintPipeline(conf)

# Apply mmgp memory-management + quantisation
print("Enabling mmgp off-loading / quantisation...")
try:
    core_pipe = paint_pipeline.models["multiview_model"].pipeline
    offload.profile(core_pipe, profile_type.LowRAM_LowVRAM)
except Exception as e:
    print("[mmgp] Failed to apply off-loading profile ➜ continuing without it.\n", e)

output_mesh_path = 'demo_textured.glb'
print("Generating textured mesh... (this might take a while)")
output_mesh_path = paint_pipeline(
    mesh_path = "demo.glb", 
    image_path = 'assets/demo.png',
    output_mesh_path = output_mesh_path
)

print(f"Textured mesh saved to {output_mesh_path}") 

[tanh609]

Can confirm this work great with the demo. @0xPD33 will this work with the gradio_app?

[bc129697]

I just tried it and it works! Pretty fast too. Thank you so much!

For anyone interested here is the adapted demo.py which I used to get this to work (after running pip install mmgp). You may have to adapt the profile based on your system configuration:

import sys
sys.path.insert(0, './hy3dshape')
sys.path.insert(0, './hy3dpaint')

from mmgp import offload, profile_type

from PIL import Image
from hy3dshape.rembg import BackgroundRemover
from hy3dshape.pipelines import Hunyuan3DDiTFlowMatchingPipeline

from textureGenPipeline import Hunyuan3DPaintPipeline, Hunyuan3DPaintConfig

try:
from torchvision_fix import apply_fix
apply_fix()
except ImportError:
print("Warning: torchvision_fix module not found, proceeding without compatibility fix")
except Exception as e:
print(f"Warning: Failed to apply torchvision fix: {e}")

model_path = 'tencent/Hunyuan3D-2.1'
print("Loading shape-generation pipeline... (GPU) ")
pipeline_shapegen = Hunyuan3DDiTFlowMatchingPipeline.from_pretrained(model_path)

image_path = 'assets/demo.png'
image = Image.open(image_path).convert("RGBA")
if image.mode == 'RGB':
rembg = BackgroundRemover()
image = rembg(image)

mesh = pipeline_shapegen(image=image)[0]
mesh.export('demo.glb')

max_num_view = 6 # can be 6 to 9
resolution = 512 # can be 768 or 512
conf = Hunyuan3DPaintConfig(max_num_view, resolution)

Make sure the heavy diffusion pipeline is kept on CPU

conf.device = "cpu"

conf.realesrgan_ckpt_path = "hy3dpaint/ckpt/RealESRGAN_x4plus.pth"
conf.multiview_cfg_path = "hy3dpaint/cfgs/hunyuan-paint-pbr.yaml"
conf.custom_pipeline = "hy3dpaint/hunyuanpaintpbr/pipeline.py"

print("Instantiating paint pipeline...")
paint_pipeline = Hunyuan3DPaintPipeline(conf)

Apply mmgp memory-management + quantisation

print("Enabling mmgp off-loading / quantisation...")
try:
core_pipe = paint_pipeline.models["multiview_model"].pipeline
offload.profile(core_pipe, profile_type.LowRAM_LowVRAM)
except Exception as e:
print("[mmgp] Failed to apply off-loading profile ➜ continuing without it.\n", e)

output_mesh_path = 'demo_textured.glb'
print("Generating textured mesh... (this might take a while)")
output_mesh_path = paint_pipeline(
mesh_path = "demo.glb",
image_path = 'assets/demo.png',
output_mesh_path = output_mesh_path
)

print(f"Textured mesh saved to {output_mesh_path}")

Hi，Why is there no .pth file corresponding to realesrgan_ckpt_path?

[0xPD33]

Can confirm this work great with the demo. @0xPD33 will this work with the gradio_app?

Yes, I managed to do it this way in the gradio app:

    if not args.disable_tex:
        try:
            # Apply torchvision fix before importing basicsr/RealESRGAN
            print("Applying torchvision compatibility fix for texture generation...")
            try:
                from torchvision_fix import apply_fix
                fix_result = apply_fix()
                if not fix_result:
                    print("Warning: Torchvision fix may not have been applied successfully")
            except Exception as fix_error:
                print(f"Warning: Failed to apply torchvision fix: {fix_error}")
            
            # from hy3dgen.texgen import Hunyuan3DPaintPipeline
            # texgen_worker = Hunyuan3DPaintPipeline.from_pretrained(args.texgen_model_path)
            # if args.low_vram_mode:
            #     texgen_worker.enable_model_cpu_offload()

            from hy3dpaint.textureGenPipeline import Hunyuan3DPaintPipeline, Hunyuan3DPaintConfig
            conf = Hunyuan3DPaintConfig(max_num_view=6, resolution=512)
            # keep heavy diffusion part on CPU so mmgp can manage VRAM
            conf.device = "cpu"
            conf.realesrgan_ckpt_path = "hy3dpaint/ckpt/RealESRGAN_x4plus.pth"
            conf.multiview_cfg_path = "hy3dpaint/cfgs/hunyuan-paint-pbr.yaml"
            conf.custom_pipeline = "hy3dpaint/hunyuanpaintpbr/pipeline.py"
            tex_pipeline = Hunyuan3DPaintPipeline(conf)
        
            # Apply MMGP off-loading & 8-bit quantisation to the heavy
            # multiview diffusion pipeline inside the texture generator.
            try:
                core_pipe = tex_pipeline.models["multiview_model"].pipeline
                offload.profile(core_pipe, profile_type.LowRAM_LowVRAM)
                print("[mmgp] LowRAM_LowVRAM profile enabled for texture pipeline")
            except Exception as e:
                print(f"[mmgp] Failed to enable MMGP off-loading: {e}")
            
            # Not help much, ignore for now.
            # if args.compile:
            #     texgen_worker.models['delight_model'].pipeline.unet.compile()
            #     texgen_worker.models['delight_model'].pipeline.vae.compile()
            #     texgen_worker.models['multiview_model'].pipeline.unet.compile()
            #     texgen_worker.models['multiview_model'].pipeline.vae.compile()
            
            HAS_TEXTUREGEN = True
            
        except Exception as e:
            import traceback
            traceback.print_exc()
            print(f"Error loading texture generator: {e}")
            print("Failed to load texture generator.")
            print('Please try to install requirements by following README.md')
            HAS_TEXTUREGEN = False

Hi，Why is there no .pth file corresponding to realesrgan_ckpt_path?

You need to download the model, check the README in the hy3dpaint subfolder

[thruCD]

Can confirm this work great with the demo. @0xPD33 will this work with the gradio_app?

Yes, I managed to do it this way in the gradio app:

if not args.disable_tex:
    try:
        # Apply torchvision fix before importing basicsr/RealESRGAN
        print("Applying torchvision compatibility fix for texture generation...")
        try:
            from torchvision_fix import apply_fix
            fix_result = apply_fix()
            if not fix_result:
                print("Warning: Torchvision fix may not have been applied successfully")
        except Exception as fix_error:
            print(f"Warning: Failed to apply torchvision fix: {fix_error}")
        
        # from hy3dgen.texgen import Hunyuan3DPaintPipeline
        # texgen_worker = Hunyuan3DPaintPipeline.from_pretrained(args.texgen_model_path)
        # if args.low_vram_mode:
        #     texgen_worker.enable_model_cpu_offload()

        from hy3dpaint.textureGenPipeline import Hunyuan3DPaintPipeline, Hunyuan3DPaintConfig
        conf = Hunyuan3DPaintConfig(max_num_view=6, resolution=512)
        # keep heavy diffusion part on CPU so mmgp can manage VRAM
        conf.device = "cpu"
        conf.realesrgan_ckpt_path = "hy3dpaint/ckpt/RealESRGAN_x4plus.pth"
        conf.multiview_cfg_path = "hy3dpaint/cfgs/hunyuan-paint-pbr.yaml"
        conf.custom_pipeline = "hy3dpaint/hunyuanpaintpbr/pipeline.py"
        tex_pipeline = Hunyuan3DPaintPipeline(conf)
    
        # Apply MMGP off-loading & 8-bit quantisation to the heavy
        # multiview diffusion pipeline inside the texture generator.
        try:
            core_pipe = tex_pipeline.models["multiview_model"].pipeline
            offload.profile(core_pipe, profile_type.LowRAM_LowVRAM)
            print("[mmgp] LowRAM_LowVRAM profile enabled for texture pipeline")
        except Exception as e:
            print(f"[mmgp] Failed to enable MMGP off-loading: {e}")
        
        # Not help much, ignore for now.
        # if args.compile:
        #     texgen_worker.models['delight_model'].pipeline.unet.compile()
        #     texgen_worker.models['delight_model'].pipeline.vae.compile()
        #     texgen_worker.models['multiview_model'].pipeline.unet.compile()
        #     texgen_worker.models['multiview_model'].pipeline.vae.compile()
        
        HAS_TEXTUREGEN = True
        
    except Exception as e:
        import traceback
        traceback.print_exc()
        print(f"Error loading texture generator: {e}")
        print("Failed to load texture generator.")
        print('Please try to install requirements by following README.md')
        HAS_TEXTUREGEN = False

Hi，Why is there no .pth file corresponding to realesrgan_ckpt_path?

You need to download the model, check the README in the hy3dpaint subfolder

the shape generation works but when i use texutre generation it generates the shape and the does the face reduction and when it comes down to texture generation it just stops using gpu and doesnt show anything after face reduction in cmd

[baptisterajaut]

Hi
To make the demo work on 2080ti i had to be inspired from https://github.com/deepbeepmeep/Hunyuan3D-2GP, heres the git diff

diff --git a/gradio_app.py b/gradio_app.py
index 47ec574..80a6211 100644
--- a/gradio_app.py
+++ b/gradio_app.py
@@ -17,6 +17,7 @@
 import sys
 sys.path.insert(0, './hy3dshape')
 sys.path.insert(0, './hy3dpaint')
+from mmgp import offload, profile_type
 
 
 try:
@@ -84,6 +85,23 @@ else:
             def __call__(self, func):
                 return func 
 
+def replace_property_getter(instance, property_name, new_getter):
+    # Get the original class and property
+    original_class = type(instance)
+    original_property = getattr(original_class, property_name)
+    
+    # Create a custom subclass for this instance
+    custom_class = type(f'Custom{original_class.__name__}', (original_class,), {})
+    
+    # Create a new property with the new getter but same setter
+    new_property = property(new_getter, original_property.fset)
+    setattr(custom_class, property_name, new_property)
+    
+    # Change the instance's class
+    instance.__class__ = custom_class
+    
+    return instance
+
 def get_example_img_list():
     """
     Load and return a sorted list of example image file paths.
@@ -737,7 +755,7 @@ if __name__ == '__main__':
     parser.add_argument("--model_path", type=str, default='tencent/Hunyuan3D-2.1')
     parser.add_argument("--subfolder", type=str, default='hunyuan3d-dit-v2-1')
     parser.add_argument("--texgen_model_path", type=str, default='tencent/Hunyuan3D-2.1')
-    parser.add_argument('--port', type=int, default=8080)
+    parser.add_argument('--port', type=int, default=7860)
     parser.add_argument('--host', type=str, default='0.0.0.0')
     parser.add_argument('--device', type=str, default='cuda')
     parser.add_argument('--mc_algo', type=str, default='mc')
@@ -798,6 +816,8 @@ if __name__ == '__main__':
 
             from hy3dpaint.textureGenPipeline import Hunyuan3DPaintPipeline, Hunyuan3DPaintConfig
             conf = Hunyuan3DPaintConfig(max_num_view=8, resolution=768)
+           # keep heavy diffusion part on CPU so mmgp can manage VRAM
+            conf.device = "cpu"
             conf.realesrgan_ckpt_path = "hy3dpaint/ckpt/RealESRGAN_x4plus.pth"
             conf.multiview_cfg_path = "hy3dpaint/cfgs/hunyuan-paint-pbr.yaml"
             conf.custom_pipeline = "hy3dpaint/hunyuanpaintpbr"
@@ -820,7 +845,7 @@ if __name__ == '__main__':
             print('Please try to install requirements by following README.md')
             HAS_TEXTUREGEN = False
 
-    HAS_T2I = True
+    HAS_T2I = False
     if args.enable_t23d:
         from hy3dgen.text2image import HunyuanDiTPipeline
 
@@ -848,6 +873,22 @@ if __name__ == '__main__':
     floater_remove_worker = FloaterRemover()
     degenerate_face_remove_worker = DegenerateFaceRemover()
     face_reduce_worker = FaceReducer()
+    profile = 2
+    kwargs = {}
+    replace_property_getter(i23d_worker, "_execution_device", lambda self : "cuda")
+    pipe = offload.extract_models("i23d_worker", i23d_worker)
+    if HAS_TEXTUREGEN:
+        pipe.update(offload.extract_models( "tex_pipeline", tex_pipeline))
+        tex_pipeline.models["multiview_model"].pipeline.vae.use_slicing = True
+    if HAS_T2I:
+        pipe.update(offload.extract_models( "t2i_worker", t2i_worker))
+    #if profile < 5:
+        #kwargs["pinnedMemory"] = "i23d_worker/model"
+    if profile !=1 and profile !=3:
+        kwargs["budgets"] = { "*" : 2200 }
+    offload.default_verboseLevel = verboseLevel = 1
+    offload.profile(pipe, profile_no = profile, verboseLevel = 1, **kwargs)
 
     # https://discuss.huggingface.co/t/how-to-serve-an-html-file/33921/2
     # create a FastAPI app
diff --git a/requirements.txt b/requirements.txt
index d9f1e4d..db5349d 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -53,7 +53,7 @@ psutil==6.0.0
 cupy-cuda12x==13.4.1
 
 # Blender
-bpy==4.0
+bpy==4.2
 
 # ONNX Runtime
 onnxruntime==1.16.3

Offloading i23d_worker is not required for my 2080ti (gets the shape gen down to 3g but slows down the process by 1/6)

[bangboomtata]

does anyone knows if it works with intel arc B580 gpu?

[RiasadHuq10]

I just tried it and it works! Pretty fast too. Thank you so much!

For anyone interested here is the adapted demo.py which I used to get this to work (after running pip install mmgp). You may have to adapt the profile based on your system configuration:

import sys
sys.path.insert(0, './hy3dshape')
sys.path.insert(0, './hy3dpaint')

from mmgp import offload, profile_type

from PIL import Image
from hy3dshape.rembg import BackgroundRemover
from hy3dshape.pipelines import Hunyuan3DDiTFlowMatchingPipeline

from textureGenPipeline import Hunyuan3DPaintPipeline, Hunyuan3DPaintConfig

try:
from torchvision_fix import apply_fix
apply_fix()
except ImportError:
print("Warning: torchvision_fix module not found, proceeding without compatibility fix")
except Exception as e:
print(f"Warning: Failed to apply torchvision fix: {e}")

model_path = 'tencent/Hunyuan3D-2.1'
print("Loading shape-generation pipeline... (GPU) ")
pipeline_shapegen = Hunyuan3DDiTFlowMatchingPipeline.from_pretrained(model_path)

image_path = 'assets/demo.png'
image = Image.open(image_path).convert("RGBA")
if image.mode == 'RGB':
rembg = BackgroundRemover()
image = rembg(image)

mesh = pipeline_shapegen(image=image)[0]
mesh.export('demo.glb')

max_num_view = 6 # can be 6 to 9
resolution = 512 # can be 768 or 512
conf = Hunyuan3DPaintConfig(max_num_view, resolution)

Make sure the heavy diffusion pipeline is kept on CPU

conf.device = "cpu"

conf.realesrgan_ckpt_path = "hy3dpaint/ckpt/RealESRGAN_x4plus.pth"
conf.multiview_cfg_path = "hy3dpaint/cfgs/hunyuan-paint-pbr.yaml"
conf.custom_pipeline = "hy3dpaint/hunyuanpaintpbr/pipeline.py"

print("Instantiating paint pipeline...")
paint_pipeline = Hunyuan3DPaintPipeline(conf)

Apply mmgp memory-management + quantisation

print("Enabling mmgp off-loading / quantisation...")
try:
core_pipe = paint_pipeline.models["multiview_model"].pipeline
offload.profile(core_pipe, profile_type.LowRAM_LowVRAM)
except Exception as e:
print("[mmgp] Failed to apply off-loading profile ➜ continuing without it.\n", e)

output_mesh_path = 'demo_textured.glb'
print("Generating textured mesh... (this might take a while)")
output_mesh_path = paint_pipeline(
mesh_path = "demo.glb",
image_path = 'assets/demo.png',
output_mesh_path = output_mesh_path
)

print(f"Textured mesh saved to {output_mesh_path}")

Hi so im trying to run this on my 4070 super. I'm very new to this so kinda lost. Is there anything I should change cuz I have only 12GB of VRAM