diffusers/tests/pipelines/kandinsky/test_kandinsky.py

# coding=utf-8
# Copyright 2025 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import gc
import random
import unittest

import numpy as np
import pytest
import torch
from transformers import XLMRobertaTokenizerFast

from diffusers import DDIMScheduler, KandinskyPipeline, KandinskyPriorPipeline, UNet2DConditionModel, VQModel
from diffusers.pipelines.kandinsky.text_encoder import MCLIPConfig, MultilingualCLIP
from diffusers.utils import is_transformers_version

from ...testing_utils import (
    backend_empty_cache,
    enable_full_determinism,
    floats_tensor,
    load_numpy,
    require_torch_accelerator,
    slow,
    torch_device,
)
from ..test_pipelines_common import PipelineTesterMixin, assert_mean_pixel_difference


enable_full_determinism()


class Dummies:
    @property
    def text_embedder_hidden_size(self):
        return 32

    @property
    def time_input_dim(self):
        return 32

    @property
    def block_out_channels_0(self):
        return self.time_input_dim

    @property
    def time_embed_dim(self):
        return self.time_input_dim * 4

    @property
    def cross_attention_dim(self):
        return 32

    @property
    def dummy_tokenizer(self):
        tokenizer = XLMRobertaTokenizerFast.from_pretrained("YiYiXu/tiny-random-mclip-base")
        return tokenizer

    @property
    def dummy_text_encoder(self):
        torch.manual_seed(0)
        config = MCLIPConfig(
            numDims=self.cross_attention_dim,
            transformerDimensions=self.text_embedder_hidden_size,
            hidden_size=self.text_embedder_hidden_size,
            intermediate_size=37,
            num_attention_heads=4,
            num_hidden_layers=5,
            vocab_size=1005,
        )

        text_encoder = MultilingualCLIP(config)
        text_encoder = text_encoder.eval()

        return text_encoder

    @property
    def dummy_unet(self):
        torch.manual_seed(0)

        model_kwargs = {
            "in_channels": 4,
            # Out channels is double in channels because predicts mean and variance
            "out_channels": 8,
            "addition_embed_type": "text_image",
            "down_block_types": ("ResnetDownsampleBlock2D", "SimpleCrossAttnDownBlock2D"),
            "up_block_types": ("SimpleCrossAttnUpBlock2D", "ResnetUpsampleBlock2D"),
            "mid_block_type": "UNetMidBlock2DSimpleCrossAttn",
            "block_out_channels": (self.block_out_channels_0, self.block_out_channels_0 * 2),
            "layers_per_block": 1,
            "encoder_hid_dim": self.text_embedder_hidden_size,
            "encoder_hid_dim_type": "text_image_proj",
            "cross_attention_dim": self.cross_attention_dim,
            "attention_head_dim": 4,
            "resnet_time_scale_shift": "scale_shift",
            "class_embed_type": None,
        }

        model = UNet2DConditionModel(**model_kwargs)
        return model

    @property
    def dummy_movq_kwargs(self):
        return {
            "block_out_channels": [32, 64],
            "down_block_types": ["DownEncoderBlock2D", "AttnDownEncoderBlock2D"],
            "in_channels": 3,
            "latent_channels": 4,
            "layers_per_block": 1,
            "norm_num_groups": 8,
            "norm_type": "spatial",
            "num_vq_embeddings": 12,
            "out_channels": 3,
            "up_block_types": [
                "AttnUpDecoderBlock2D",
                "UpDecoderBlock2D",
            ],
            "vq_embed_dim": 4,
        }

    @property
    def dummy_movq(self):
        torch.manual_seed(0)
        model = VQModel(**self.dummy_movq_kwargs)
        return model

    def get_dummy_components(self):
        text_encoder = self.dummy_text_encoder
        tokenizer = self.dummy_tokenizer
        unet = self.dummy_unet
        movq = self.dummy_movq

        scheduler = DDIMScheduler(
            num_train_timesteps=1000,
            beta_schedule="linear",
            beta_start=0.00085,
            beta_end=0.012,
            clip_sample=False,
            set_alpha_to_one=False,
            steps_offset=1,
            prediction_type="epsilon",
            thresholding=False,
        )

        components = {
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "unet": unet,
            "scheduler": scheduler,
            "movq": movq,
        }
        return components

    def get_dummy_inputs(self, device, seed=0):
        image_embeds = floats_tensor((1, self.cross_attention_dim), rng=random.Random(seed)).to(device)
        negative_image_embeds = floats_tensor((1, self.cross_attention_dim), rng=random.Random(seed + 1)).to(device)
        if str(device).startswith("mps"):
            generator = torch.manual_seed(seed)
        else:
            generator = torch.Generator(device=device).manual_seed(seed)
        inputs = {
            "prompt": "horse",
            "image_embeds": image_embeds,
            "negative_image_embeds": negative_image_embeds,
            "generator": generator,
            "height": 64,
            "width": 64,
            "guidance_scale": 4.0,
            "num_inference_steps": 2,
            "output_type": "np",
        }
        return inputs


class KandinskyPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
    pipeline_class = KandinskyPipeline
    params = [
        "prompt",
        "image_embeds",
        "negative_image_embeds",
    ]
    batch_params = ["prompt", "negative_prompt", "image_embeds", "negative_image_embeds"]
    required_optional_params = [
        "generator",
        "height",
        "width",
        "latents",
        "guidance_scale",
        "negative_prompt",
        "num_inference_steps",
        "return_dict",
        "guidance_scale",
        "num_images_per_prompt",
        "output_type",
        "return_dict",
    ]
    test_xformers_attention = False

    supports_dduf = False

    def get_dummy_components(self):
        dummy = Dummies()
        return dummy.get_dummy_components()

    def get_dummy_inputs(self, device, seed=0):
        dummy = Dummies()
        return dummy.get_dummy_inputs(device=device, seed=seed)

    @pytest.mark.xfail(
        condition=is_transformers_version(">=", "4.56.2"),
        reason="Latest transformers changes the slices",
        strict=False,
    )
    def test_kandinsky(self):
        device = "cpu"

        components = self.get_dummy_components()

        pipe = self.pipeline_class(**components)
        pipe = pipe.to(device)

        pipe.set_progress_bar_config(disable=None)

        output = pipe(**self.get_dummy_inputs(device))
        image = output.images

        image_from_tuple = pipe(
            **self.get_dummy_inputs(device),
            return_dict=False,
        )[0]

        image_slice = image[0, -3:, -3:, -1]
        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)

        expected_slice = np.array([1.0000, 1.0000, 0.2766, 1.0000, 0.5447, 0.1737, 1.0000, 0.4316, 0.9024])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2, (
            f" expected_slice {expected_slice}, but got {image_slice.flatten()}"
        )
        assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2, (
            f" expected_slice {expected_slice}, but got {image_from_tuple_slice.flatten()}"
        )

    @require_torch_accelerator
    def test_offloads(self):
        pipes = []
        components = self.get_dummy_components()
        sd_pipe = self.pipeline_class(**components).to(torch_device)
        pipes.append(sd_pipe)

        components = self.get_dummy_components()
        sd_pipe = self.pipeline_class(**components)
        sd_pipe.enable_model_cpu_offload(device=torch_device)
        pipes.append(sd_pipe)

        components = self.get_dummy_components()
        sd_pipe = self.pipeline_class(**components)
        sd_pipe.enable_sequential_cpu_offload(device=torch_device)
        pipes.append(sd_pipe)

        image_slices = []
        for pipe in pipes:
            inputs = self.get_dummy_inputs(torch_device)
            image = pipe(**inputs).images

            image_slices.append(image[0, -3:, -3:, -1].flatten())

        assert np.abs(image_slices[0] - image_slices[1]).max() < 1e-3
        assert np.abs(image_slices[0] - image_slices[2]).max() < 1e-3


@slow
@require_torch_accelerator
class KandinskyPipelineIntegrationTests(unittest.TestCase):
    def setUp(self):
        # clean up the VRAM before each test
        super().setUp()
        gc.collect()
        backend_empty_cache(torch_device)

    def tearDown(self):
        # clean up the VRAM after each test
        super().tearDown()
        gc.collect()
        backend_empty_cache(torch_device)

    def test_kandinsky_text2img(self):
        expected_image = load_numpy(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/kandinsky/kandinsky_text2img_cat_fp16.npy"
        )

        pipe_prior = KandinskyPriorPipeline.from_pretrained(
            "kandinsky-community/kandinsky-2-1-prior", torch_dtype=torch.float16
        )
        pipe_prior.to(torch_device)

        pipeline = KandinskyPipeline.from_pretrained("kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16)
        pipeline.to(torch_device)
        pipeline.set_progress_bar_config(disable=None)

        prompt = "red cat, 4k photo"

        generator = torch.Generator(device=torch_device).manual_seed(0)
        image_emb, zero_image_emb = pipe_prior(
            prompt,
            generator=generator,
            num_inference_steps=5,
            negative_prompt="",
        ).to_tuple()

        generator = torch.Generator(device=torch_device).manual_seed(0)
        output = pipeline(
            prompt,
            image_embeds=image_emb,
            negative_image_embeds=zero_image_emb,
            generator=generator,
            num_inference_steps=100,
            output_type="np",
        )

        image = output.images[0]

        assert image.shape == (512, 512, 3)

        assert_mean_pixel_difference(image, expected_image)