Limitations

Due to differences in framework, some APIs & models will not be identical to huggingface/diffusers in the foreseeable future.

APIs

xxx.from_pretrained

• torch_dtype is renamed to mindspore_dtype
• device_map, max_memory, offload_folder, offload_state_dict, low_cpu_mem_usage will not be supported.

BaseOutput

• Default value of return_dict is changed to False, for GRAPH_MODE does not allow to construct an instance of it.

Output of AutoencoderKL.encode

Unlike the output posterior = DiagonalGaussianDistribution(latent), which can do sampling by posterior.sample(). We can only output the latent and then do sampling through AutoencoderKL.diag_gauss_dist.sample(latent).

self.config in construct()

For many models, parameters used in initialization will be registered in self.config. They are often accessed during the construct like using if self.config.xxx == xxx to determine execution paths in origin 🤗diffusers. However getting attributes like this is not supported by static graph syntax of MindSpore. Two feasible replacement options are

• set new attributes in initialization for self like self.xxx = self.config.xxx, then use self.xxx in construct instead.
• use self.config["xxx"] as self.config is an OrderedDict and getting items like this is supported in static graph mode.

When self.config.xxx changed, we change self.xxx and self.config["xxx"] both.

Models

The table below represents the current support in mindone/diffusers for each of those modules, whether they have support in Pynative fp16 mode, Graph fp16 mode, Pynative fp32 mode or Graph fp32 mode.

Names	Pynative FP16	Pynative FP32	Graph FP16	Graph FP32	Description
StableCascadeUNet	❌	✅	❌	✅	huggingface/diffusers output NaN when using float16.
nn.Conv3d	✅	❌	✅	❌	FP32 is not supported on Ascend
TemporalConvLayer	✅	❌	✅	❌	contains nn.Conv3d
TemporalResnetBlock	✅	❌	✅	❌	contains nn.Conv3d
SpatioTemporalResBlock	✅	❌	✅	❌	contains TemporalResnetBlock
UNetMidBlock3DCrossAttn	✅	❌	✅	❌	contains TemporalConvLayer
CrossAttnDownBlock3D	✅	❌	✅	❌	contains TemporalConvLayer
DownBlock3D	✅	❌	✅	❌	contains TemporalConvLayer
CrossAttnUpBlock3D	✅	❌	✅	❌	contains TemporalConvLayer
UpBlock3D	✅	❌	✅	❌	contains TemporalConvLayer
MidBlockTemporalDecoder	✅	❌	✅	❌	contains SpatioTemporalResBlock
UpBlockTemporalDecoder	✅	❌	✅	❌	contains SpatioTemporalResBlock
UNetMidBlockSpatioTemporal	✅	❌	✅	❌	contains SpatioTemporalResBlock
DownBlockSpatioTemporal	✅	❌	✅	❌	contains SpatioTemporalResBlock
CrossAttnDownBlockSpatioTemporal	✅	❌	✅	❌	contains SpatioTemporalResBlock
UpBlockSpatioTemporal	✅	❌	✅	❌	contains SpatioTemporalResBlock
CrossAttnUpBlockSpatioTemporal	✅	❌	✅	❌	contains SpatioTemporalResBlock
TemporalDecoder	✅	❌	✅	❌	contains nn.Conv3d, MidBlockTemporalDecoder etc.
UNet3DConditionModel	✅	❌	✅	❌	contains UNetMidBlock3DCrossAttn etc.
I2VGenXLUNet	✅	❌	✅	❌	contains UNetMidBlock3DCrossAttn etc.
AutoencoderKLTemporalDecoder	✅	❌	✅	❌	contains MidBlockTemporalDecoder etc.
UNetSpatioTemporalConditionModel	✅	❌	✅	❌	contains UNetMidBlockSpatioTemporal etc.
FirUpsample2D	❌	✅	✅	✅	ops.Conv2D has poor precision in fp16 and PyNative mode
FirDownsample2D	❌	✅	✅	✅	ops.Conv2D has poor precision in fp16 and PyNative mode
AttnSkipUpBlock2D	❌	✅	✅	✅	contains FirUpsample2D
SkipUpBlock2D	❌	✅	✅	✅	contains FirUpsample2D
AttnSkipDownBlock2D	❌	✅	✅	✅	contains FirDownsample2D
SkipDownBlock2D	❌	✅	✅	✅	contains FirDownsample2D
ResnetBlock2D (kernel='fir')	❌	✅	✅	✅	ops.Conv2D has poor precision in fp16 and PyNative mode
CogVideoXTransformer3DModel	✅	❌	✅	❌	Using FlashAttention which only supports FP16/BF16
AutoencoderKLCogVideoX	✅	❌	❌	❌	Only FP16/BF16: contains nn.Conv3d as sub cells & Only PyNative: Static Graph Syntax doesn't support cache operations inside
CogVideoXEncoder3D	✅	❌	❌	❌	Only FP16/BF16: contains nn.Conv3d as sub cells & Only PyNative: Static Graph Syntax doesn't support cache operations inside
CogVideoXDecoder3D	✅	❌	❌	❌	Only FP16/BF16: contains nn.Conv3d as sub cells & Only PyNative: Static Graph Syntax doesn't support cache operations inside

Pipelines

The table below represents the current support in mindone/diffusers for each of those pipelines in MindSpore 2.3.1, whether they have support in Pynative fp16 mode, Graph fp16 mode, Pynative fp32 mode or Graph fp32 mode.

precision issues of pipelines, the experiments in the table below default to upcasting GroupNorm to FP32 to avoid this issue.

Pipelines	Pynative FP16	Pynative FP32	Graph FP16	Graph FP32	Description
AnimateDiffPipeline	✅	✅	✅	✅
AnimateDiffVideoToVideoPipeline	✅	❌	✅	✅	In FP32 and Pynative mode, this pipeline will run out of memory
BlipDiffusionPipeline	✅	✅	✅	✅
ConsistencyModelPipeline	✅	✅	✅	✅
CogVideoXPipeline	✅	❌	❌	❌	Flash Attention used in transformer and nn.Conv3d used in VAE only support FP16/BF16; Cache opeartions in VAE is not supported by MindSpore Graph Mode
CogVideoXImageToVideoPipeline	✅	❌	❌	❌	Flash Attention used in transformer and nn.Conv3d used in VAE only support FP16/BF16; Cache opeartions in VAE is not supported by MindSpore Graph Mode
CogVideoXVideoToVideoPipeline	✅	❌	❌	❌	Flash Attention used in transformer and nn.Conv3d used in VAE only support FP16/BF16; Cache opeartions in VAE is not supported by MindSpore Graph Mode
DDIMPipeline	✅	✅	✅	✅
DDPMPipeline	✅	✅	✅	✅
DiTPipeline	✅	✅	✅	✅
I2VGenXLPipeline	✅	❌	✅	❌	ops.bmm and ops.softmax have precision issues under FP16, so we need to upcast them to FP32 to get a good result; FP32 is not supported since I2VGenXLPipeline contains nn.Conv3d
IFImg2ImgPipeline	✅	✅	✅	✅
IFImg2ImgSuperResolutionPipeline	✅	✅	✅	✅
IFInpaintingPipeline	✅	✅	✅	✅
IFInpaintingSuperResolutionPipeline	✅	✅	✅	✅
IFPipeline	✅	✅	✅	✅
IFSuperResolutionPipeline	✅	✅	✅	✅
Kandinsky3Img2ImgPipeline	❌	❌	❌	❌	Kandinsky3 only provides FP16 weights; additionally, T5 has precision issues, so to achieve the desired results, you need to directly input prompt_embeds and attention_mask.
Kandinsky3Pipeline	❌	❌	❌	❌	Kandinsky3 only provides FP16 weights; additionally, T5 has precision issues, so to achieve the desired results, you need to directly input prompt_embeds and attention_mask.
KandinskyImg2ImgPipeline	✅	✅	✅	✅
KandinskyInpaintPipeline	✅	✅	✅	✅
KandinskyPipeline	✅	✅	✅	✅
KandinskyV22ControlnetImg2ImgPipeline	✅	✅	✅	✅
KandinskyV22ControlnetPipeline	✅	✅	✅	✅
KandinskyV22Img2ImgPipeline	✅	✅	✅	✅
KandinskyV22InpaintPipeline	✅	✅	✅	✅
KandinskyV22Pipeline	✅	✅	✅	✅
LatentConsistencyModelImg2ImgPipeline	✅	✅	✅	✅
LatentConsistencyModelPipeline	✅	✅	✅	✅
LDMSuperResolutionPipeline	✅	✅	✅	✅
LDMTextToImagePipeline	✅	✅	✅	✅
PixArtAlphaPipeline	✅	✅	✅	✅
ShapEImg2ImgPipeline	✅	✅	❌	❌	The syntax in Render only supports Pynative mode
ShapEPipeline	✅	✅	❌	❌	The syntax in Render only supports Pynative mode
StableCascadePipeline	❌	✅	❌	✅	This pipeline does not support FP16 due to precision issues
StableDiffusion3Pipeline	✅	✅	✅	✅
StableDiffusionAdapterPipeline	✅	✅	✅	✅
StableDiffusionControlNetImg2ImgPipeline	✅	✅	✅	✅
StableDiffusionControlNetInpaintPipeline	✅	✅	✅	✅
StableDiffusionControlNetPipeline	✅	✅	✅	✅
StableDiffusionDepth2ImgPipeline	✅	✅	✅	✅
StableDiffusionDiffEditPipeline	✅	✅	✅	✅
StableDiffusionGLIGENPipeline	✅	✅	✅	✅
StableDiffusionGLIGENTextImagePipeline	✅	✅	✅	✅
StableDiffusionImageVariationPipeline	✅	✅	✅	✅
StableDiffusionImg2ImgPipeline	✅	✅	✅	✅
StableDiffusionInpaintPipeline	✅	✅	✅	✅
StableDiffusionInstructPix2PixPipeline	✅	✅	✅	✅
StableDiffusionLatentUpscalePipeline	✅	✅	✅	✅
StableDiffusionPipeline	✅	✅	✅	✅
StableDiffusionUpscalePipeline	✅	✅	✅	✅
StableDiffusionXLAdapterPipeline	✅	✅	✅	✅
StableDiffusionXLControlNetImg2ImgPipeline	✅	✅	✅	✅
StableDiffusionXLControlNetInpaintPipeline	✅	✅	✅	✅
StableDiffusionXLControlNetPipeline	✅	✅	✅	✅
StableDiffusionXLImg2ImgPipeline	✅	✅	✅	✅
StableDiffusionXLInpaintPipeline	✅	✅	✅	✅
StableDiffusionXLInstructPix2PixPipeline	✅	✅	✅	✅
StableDiffusionXLPipeline	✅	✅	✅	✅
StableVideoDiffusionPipeline	✅	❌	✅	❌	This pipeline will run out of memory under FP32; ops.bmm and ops.softmax have precision issues under FP16, so we need to upcast them to FP32 to get a good result
UnCLIPImageVariationPipeline	✅	✅	✅	✅
UnCLIPPipeline	✅	✅	✅	✅
WuerstchenPipeline	✅	✅	✅	✅	GlobalResponseNorm has precision issue under FP16, so we need to upcast it to FP32 to get a good result