bi ddlZddlZddlmZmZmZmZmZmZddl Z ddl Z ddl m Z mZmZddlmZmZddlmZddlmZmZddlmZdd lmZmZmZdd lmZd d l m!Z!d dl"m#Z#er ddl$m%cm&Z'dZ(ndZ(ejRe*Z+dZ, d#de-de-de.de.fdZ/ d$dee-deee0e jbfdeee-deee.fdZ2 d%de jfdee jhde0fdZ5d Z6Gd!d"e!eZ7y)&N)AnyCallableDictListOptionalUnion)"Qwen2_5_VLForConditionalGenerationQwen2TokenizerQwen2VLProcessor)PipelineImageInputVaeImageProcessor)QwenImageLoraLoaderMixin)AutoencoderKLQwenImageQwenImageTransformer2DModel)FlowMatchEulerDiscreteScheduler)is_torch_xla_availableloggingreplace_example_docstring) randn_tensor)DiffusionPipeline)QwenImagePipelineOutputTFa Examples: ```py >>> import torch >>> from PIL import Image >>> from diffusers import QwenImageEditPipeline >>> from diffusers.utils import load_image >>> pipe = QwenImageEditPipeline.from_pretrained("Qwen/Qwen-Image-Edit", torch_dtype=torch.bfloat16) >>> pipe.to("cuda") >>> image = load_image( ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/yarn-art-pikachu.png" ... ).convert("RGB") >>> prompt = ( ... "Make Pikachu hold a sign that says 'Qwen Edit is awesome', yarn art style, detailed, vibrant colors" ... ) >>> # Depending on the variant being used, the pipeline call will slightly vary. >>> # Refer to the pipeline documentation for more details. >>> image = pipe(image, prompt, num_inference_steps=50).images[0] >>> image.save("qwenimage_edit.png") ``` base_seq_len max_seq_len base_shift max_shiftc<||z ||z z }|||zz }||z|z}|SN) image_seq_lenrrrrmbmus p/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit.pycalculate_shiftr'Ds; Z K,$>?AQ%%A  Q B Inum_inference_stepsdevice timestepssigmasc | | td|dttj|jj j v}|std|jd|jd ||d||j}t|}||fS|dttj|jj j v}|std|jd|jd ||d||j}t|}||fS|j|fd |i||j}||fS) a Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. Args: scheduler (`SchedulerMixin`): The scheduler to get timesteps from. num_inference_steps (`int`): The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` must be `None`. device (`str` or `torch.device`, *optional*): The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. timesteps (`List[int]`, *optional*): Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, `num_inference_steps` and `sigmas` must be `None`. sigmas (`List[float]`, *optional*): Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, `num_inference_steps` and `timesteps` must be `None`. Returns: `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the second element is the number of inference steps. zYOnly one of `timesteps` or `sigmas` can be passed. Please choose one to set custom valuesr+zThe current scheduler class zx's `set_timesteps` does not support custom timestep schedules. Please check whether you are using the correct scheduler.)r+r*r,zv's `set_timesteps` does not support custom sigmas schedules. Please check whether you are using the correct scheduler.)r,r*r*r!) ValueErrorsetinspect signature set_timesteps parameterskeys __class__r+len) schedulerr)r*r+r,kwargsaccepts_timesteps accept_sigmass r&retrieve_timestepsr;Rs>!3tuu'3w/@/@AXAX/Y/d/d/i/i/k+ll .y/B/B.CDab  M)FMfM'' !)n ) ))   C(9(9):Q:Q(R(](](b(b(d$ee .y/B/B.CD_`  GvfGG'' !)n ) ))  3MFMfM'' ) ))r(encoder_output generator sample_modect|dr |dk(r|jj|St|dr|dk(r|jjSt|dr |jSt d)N latent_distsampleargmaxlatentsz3Could not access latents of provided encoder_output)hasattrr@rAmoderCAttributeError)r<r=r>s r&retrieve_latentsrGst~}-+2I))00;;  /K84K))..00  +%%%RSSr(ctj||z}||z }t|dz dz}t|dz dz}||dfS)N )mathsqrtround) target_arearatiowidthheights r&calculate_dimensionsrQsQ IIkE) *E U]F %"*  "E 6B; " $F &$ r(c/eZdZdZdZddgZdededede d e d e f fd Z d e jd e jfdZ d>deeeefdee jdee j(dee j*fdZ d?deeeefdee jdee j(dedee jdee jdefdZ d@dZedZedZde jde j:fdZdZd Z d!Z!d"Z" dAd#Z#e$d$Z%e$d%Z&e$d&Z'e$d'Z(e$d(Z)e jTe+e,dddd)ddd*dd+dddddddd,d-dddgd.fdee-deeeefd/eeeefd0e.d1eed2eed3ed4eee.d5e.dedeee j:ee j:fdee jdee jdee jd6ee jd7ee jd8eed9e/d:ee0ee1fd;ee2eee0gdfdtransformer->vaerC prompt_embedsr7vae text_encoder tokenizer processor transformerct||j||||||t|ddr"dt |j j znd|_t|ddr |j jjnd|_ t|jdz|_ ||_ d|_d|_d |_d |_y) N)rUrVrWrXrYr7rUr)vae_scale_factora<|im_start|>system Describe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|> <|im_start|>user <|vision_start|><|image_pad|><|vision_end|>{}<|im_end|> <|im_start|>assistant @)super__init__register_modulesgetattrr6rUtemperal_downsampler]configz_dimlatent_channelsrimage_processor vl_processortokenizer_max_lengthprompt_template_encode prompt_template_encode_start_idxdefault_sample_size)selfr7rUrVrWrXrYr5s r&rbzQwenImageEditPipeline.__init__s  %#  KRRVX]_cJdS)E)E%F Fjk8?eT8Rtxx44XZ 1$BWBWZ[B[\%$(!'[#02-#& r( hidden_statesmaskc|j}|jd}||}tj||j d}|S)Nrdimr)boolsumtorchsplittolist)rorprq bool_mask valid_lengthsselected split_results r&_extract_masked_hiddenz,QwenImageEditPipeline._extract_masked_hiddensHIIK ! ! ,  +{{8]-A-A-CK r(Npromptimager*dtypec|xs |j}|xs|jj}t|tr|gn|}|j }|j }|Dcgc]}|j|}}|j||ddj|} |j| j| j| j| jd} | jd} |j| | j} | Dcgc]}||d } }| Dcgc]A}t!j"|j%dt j&|j(C} }t+| Dcgc]}|j%dc}}t!j,| Dcgc]J}t!j.||j1||j%dz |j%dgLc}}t!j,| Dcgc]:}t!j.||j1||j%dz g<c}}|j||}||fScc}wcc}wcc}wcc}wcc}wcc}w) NTpt)textimagespaddingreturn_tensors) input_idsattention_mask pixel_valuesimage_grid_thwoutput_hidden_statesr)rr*r)_execution_devicerVr isinstancestrrlrmformatrXtorrrrrpr~rwonessizelongr*maxstackcat new_zeros)rorrr*rtemplatedrop_idxetxt model_inputsoutputsrpsplit_hidden_statesattn_mask_listrurTencoder_attention_masks r&_get_qwen_prompt_embedsz-QwenImageEditPipeline._get_qwen_prompt_embedss814110**00'4&&..88+12axq!22~~ & "V* ##",,'66%22'66!% $  --b1 "99-IdIde5HIq|II\opWX%**QVVAYejjRpp.AB166!9BC  Vi jQRUYY1;;{QVVAY'>q JK L j "'KY ZaUYY1;;{QVVAY'>?@ A Z" &((uV(D 444?3&JpB j [s&I" I'AI,*I1AI6?I;rnum_images_per_promptprompt_embeds_maskmax_sequence_lengthc~|xs |j}t|tr|gn|}| t|n|jd}||j |||\}}|j\} } } |j d|d}|j||z| d}|j d|d}|j||z| }||fS)a Args: prompt (`str` or `List[str]`, *optional*): prompt to be encoded image (`torch.Tensor`, *optional*): image to be encoded device: (`torch.device`): torch device num_images_per_prompt (`int`): number of images that should be generated per prompt prompt_embeds (`torch.Tensor`, *optional*): Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, text embeddings will be generated from `prompt` input argument. rrr)rrrr6shaperrepeatview) rorrr*rrTrr batch_size_seq_lens r& encode_promptz#QwenImageEditPipeline.encode_prompts21411'4&&$1$9S[}?R?RST?U  040L0LVUZ\b0c -M-%++ 7A%,,Q0EqI %**:8M+MwXZ[ /66q:OQRS/44ZBW5WY`a000r(c |jdzzdk7s|jdzzdk7r,tjdjdzd|d|d| Ltfd| Ds8t dj d | D cgc]} | j vs| c} ||t d |d |d | | t d |7t |ts't |tst dt|||t d|d|d | | t d| | t d| | dkDrt d| yycc} w)Nrrz-`height` and `width` have to be divisible by z but are z and z(. Dimensions will be resized accordinglyc3:K|]}|jvywr )_callback_tensor_inputs).0kros r& z5QwenImageEditPipeline.check_inputs..Ms#F 23A-- -F sz2`callback_on_step_end_tensor_inputs` has to be in z , but found zCannot forward both `prompt`: z and `prompt_embeds`: z2. Please make sure to only forward one of the two.zeProvide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined.z2`prompt` has to be of type `str` or `list` but is z'Cannot forward both `negative_prompt`: z and `negative_prompt_embeds`: zIf `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. Make sure to generate `prompt_embeds_mask` from the same text encoder that was used to generate `prompt_embeds`.zIf `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was used to generate `negative_prompt_embeds`.r^z9`max_sequence_length` cannot be greater than 1024 but is ) r]loggerwarningallr.rrrlisttype) rorrPrOnegative_promptrTnegative_prompt_embedsrnegative_prompt_embeds_mask"callback_on_step_end_tensor_inputsrrs ` r& check_inputsz"QwenImageEditPipeline.check_inputs;s3 T**Q. /1 4AVAVYZAZ8[_`8` NN?@U@UXY@Y?ZZcdjckkpqvpwx`a  . 9#F 7YF C DTEaEaDbbn|^pHvwbckokGkGbGpqpHoIJ   -";08N}o^00 ^ 5w  FC)@TZ\`IaQRVW]R^Q_`a a  &+A+M9/9JK*++]_   $);)CQ  " -2M2Uu   */BT/IXYlXmno o0J *=pHs EEc|j|||dzd|dzd}|jdddddd}|j||dz|dzz|dz}|S)Nrrrr )rpermutereshape)rCrnum_channels_latentsrPrOs r& _pack_latentsz#QwenImageEditPipeline._pack_latentsrsl,,z+?1aQVZ[Q[]^_//!Q1a3//*v{uz.JL`cdLder(c|j\}}}dt||dzzz}dt||dzzz}|j||dz|dz|dzdd}|jdddddd}|j ||dzd||}|S)Nrrrr rr)rintrrr)rCrPrOr]r num_patcheschannelss r&_unpack_latentsz%QwenImageEditPipeline._unpack_latents{s-4MM) Kc&k&6&:;<SZ$4q$89:,,z6Q; HPQMSTVWX//!Q1a3//*h5.A1feTr(r=c 4t|trkt|jdDcgc]2}t |j j |||dz||d4}}tj|d}n't |j j ||d}tj|j jjjd|jdddj|j|j }tj|j jj"jd|jdddj|j|j }||z |z }|Scc}w)NrrrB)r=r>rs)rrrangerrGrUencoderwrtensorrf latents_meanrrhrr*r latents_std)rorr=i image_latentsrrs r&_encode_vae_imagez'QwenImageEditPipeline._encode_vae_imagesI i &u{{1~.!q1q51A!BiXYlhpqM"IIm;M,TXX__U-CyfnoM LL55 6 T!T))1a 3 R $$m&9&9 :  LL44 5 T!T))1a 3 R $$m&9&9 :  '5D 's7Fc8|jjy)z Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. N)rUenable_slicingros r&enable_vae_slicingz(QwenImageEditPipeline.enable_vae_slicing !r(c8|jjy)z Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to computing decoding in one step. N)rUdisable_slicingrs r&disable_vae_slicingz)QwenImageEditPipeline.disable_vae_slicings   "r(c8|jjy)a Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow processing larger images. N)rU enable_tilingrs r&enable_vae_tilingz'QwenImageEditPipeline.enable_vae_tilings  r(c8|jjy)z Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to computing decoding in one step. N)rUdisable_tilingrs r&disable_vae_tilingz(QwenImageEditPipeline.disable_vae_tilingrr(c dt||jdzzz}dt||jdzzz}|d|||f} d} ||j||}|jd|jk7r|j ||} n|} || jdkDrC|| jdzdk(r.|| jdz} t j| g| zd} n^|| jdkDr4|| jdzdk7rtd| jdd|d t j| gd} | jd d\} }|j| ||| |} t|tr)t||k7rtd t|d |d | (t| |||} |j| ||||} | | fS| j||} | | fS)Nrrr*r)rr=rrsz'Cannot duplicate `image` of batch size z to z text prompts.r z/You have passed a list of generators of length z+, but requested an effective batch size of z@. Make sure the batch size matches the length of the generators.)r=r*r)rr]rrrhrrwrr.rrrr6r)rorrrrPrOrr*r=rCrradditional_image_per_promptimage_latent_heightimage_latent_widths r&prepare_latentsz%QwenImageEditPipeline.prepare_latentssDc&kd&;&;a&?@ASZD$9$9A$=>?Q 4feD  HHF%H8E{{1~!5!55 $ 6 6Ui 6 X % M//22zMDWDWXYDZ7Z^_7_.8MQ>QRS>T=UUYZdYeest!& =/q A 6C6I6I!"6M 3 !3 ..z+?ATVhM i &3y>Z+GA#i.AQR&<'gi  ?"5IfTYZG((*>RTZ\abG %%jjej 1.0 and a provided `negative_prompt`, enables true classifier-free guidance. height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): The height in pixels of the generated image. This is set to 1024 by default for the best results. width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): The width in pixels of the generated image. This is set to 1024 by default for the best results. num_inference_steps (`int`, *optional*, defaults to 50): The number of denoising steps. More denoising steps usually lead to a higher quality image at the expense of slower inference. sigmas (`List[float]`, *optional*): Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed will be used. guidance_scale (`float`, *optional*, defaults to 3.5): Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, usually at the expense of lower image quality. This parameter in the pipeline is there to support future guidance-distilled models when they come up. Note that passing `guidance_scale` to the pipeline is ineffective. To enable classifier-free guidance, please pass `true_cfg_scale` and `negative_prompt` (even an empty negative prompt like " ") should enable classifier-free guidance computations. num_images_per_prompt (`int`, *optional*, defaults to 1): The number of images to generate per prompt. generator (`torch.Generator` or `List[torch.Generator]`, *optional*): One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation deterministic. latents (`torch.Tensor`, *optional*): Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image generation. Can be used to tweak the same generation with different prompts. If not provided, a latents tensor will be generated by sampling using the supplied random `generator`. prompt_embeds (`torch.Tensor`, *optional*): Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, text embeddings will be generated from `prompt` input argument. negative_prompt_embeds (`torch.Tensor`, *optional*): Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. output_type (`str`, *optional*, defaults to `"pil"`): The output format of the generate image. Choose between [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. return_dict (`bool`, *optional*, defaults to `True`): Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. attention_kwargs (`dict`, *optional*): A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under `self.processor` in [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). callback_on_step_end (`Callable`, *optional*): A function that calls at the end of each denoising steps during the inference. The function is called with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by `callback_on_step_end_tensor_inputs`. callback_on_step_end_tensor_inputs (`List`, *optional*): The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the `._callback_tensor_inputs` attribute of your pipeline class. max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. Examples: Returns: [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`: [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated images. rirr)rrTrrrrrNF)rrrTrr*rrrrbase_image_seq_lenmax_image_seq_lenr?rffffff?)r,r%rrs)totalcondi) rptimestepguidanceencoder_hidden_states_maskencoder_hidden_states img_shapes txt_seq_lensrruncondrT)rtkeepdim)rrCrTlatent)r)r)IrrrrQr]rrrrrrr6rrrwTensorrhriresize preprocess unsqueezerrYrf in_channelsrrnplinspacer'r7getr;rorderrguidance_embedsfullfloat32expandrrvryset_begin_index progress_bar enumeraterrr cache_contextnormstepbackendsmps is_availablelocalspopupdate XLA_AVAILABLExm mark_steprrUrrrrgr*rdecode postprocessmaybe_free_model_hooksr);rorrrrrPrOr)r,rrr=rCrTrrrrrrrrr image_sizecalculated_widthcalculated_heightr multiple_ofrr* prompt_imagehas_neg_prompt do_true_cfgrrr r"r%r+num_warmup_stepsr rnegative_txt_seq_lensr rtlatent_model_inputr  noise_predneg_noise_pred comb_pred cond_norm noise_norm latents_dtypecallback_kwargsrcallback_outputsrrs; r&__call__zQwenImageEditPipeline.__call__ sl R'1&=U1X]]5:: 1EkS]^_S`cmnocpSp1q.+Q,,))++a/ ${2;&4    +'#91(C/Q 3   .!1!%  *VS"9J  Jvt$<VJ&,,Q/J''  j &E%**UV-[_[o[oJo((//7HJZ[E L((33E;LN^_EOOA&E(4 "$ . Z3NVZ3Z %q(;^ ,0,>,>'1"7 3-?- ) ) BFBTBT"&4#>&;$7CUC ? "$? $//66BBaG!%!5!5  . .         " Fd333q8%4CXCX:X\]:]^%)>)>>!CEUY]YnYnEnrsEst    TZSaS!&9"9;NOgm a(   NN ! ! % %&:C @ NN ! ! % %&94 @ NN ! ! % %lC 8 NN ! ! % %k4 8  *< NN   * & &s9~0CdnnFZFZ0ZZ\]^!)n    " " 2 2zz1#~fEMMZHw}}Q'78HH  (%'D "ASA_)--!-4;;=ei ?Z?f ' + + + 2 9 9 ;lp  &&q)   %8  9F #\!),E #1>>)*&%," ,).G]3KQR)S&88GMM!$4588G%%33F; B!%!1!1&8!)D!)3E.;#-%1)-)>)>$)"2 " "J",A/@a/@,@!AJ B))77A )-)9)9*<%-_%-7R2H'1)>-1-B-B(-*: * * &4A7Ha7H4H%IN .:P^C^1_ _I % :2t LI!&I2t!LJ!*i*.D!EJ!( ..--j!WRW-XYZ[==M1~~))668")**]";'3&(O?9-3Xa[*9';D!Q'X$.229gFG$4$8$8-$XMI**A9I/IqSTuX\XfXfXlXlNlpqNq '') LLNKE #F #P"& ( "E**7FE4CXCXYGjj0G TXX__99:a..1a8GNNGMM2   TXX__-H-H I N NqRVRZRZRaRaRgRgijlmop q t t !K +l:GHHOOGO?B1a7KE((44U 4TE ##%8O&e44g B B  9F #F #sEBd0Ad%d0+d#+E3d0 d0d d0#d- (d00d:NNNN)NNrNNr^)NNNNNNNr )5__name__ __module__ __qualname____doc__model_cpu_offload_seqrrrr r r rrbrwrr~rrrrr*rrrrr staticmethodrr Generatorrrrrrrpropertyrrrrrno_gradrEXAMPLE_DOC_STRINGr floatrurrrrD __classcell__)r5s@r&rSrSsR$=(/:'2'$'9 ' " ' $ '1'@ELL )-(,)-'+ -5c49n%-5 %-5& -5  $ -5d)-)-%&0459#''1c49n%'1 %'1& '1 # '1   - '1%U\\2'1!'1\#$(+/ 5pn  u||."#!""5&n$$&&##&&U]]_12/3(,15 # $##%(, #%&MQ*.04599=>B%* 59KO9B #&/D5*+D5c49n%D5sDI~. D5  D5  D5}D5!D5e%D5D5 #D5E%//43H"HIJD5%,,'D5  -D5%U\\2D5 !) 6!D5"&.ell%;#D5$c]%D5&'D5(#4S>2)D5*'xc40@$0F'GH+D5,-1I-D5.!/D53D5r(rS)rrrrrE)NrA)8r0rJtypingrrrrrrnumpyrrw transformersr r r rir rloadersrmodelsrr schedulersrutilsrrrutils.torch_utilsrpipeline_utilsrpipeline_outputrtorch_xla.core.xla_modelcore xla_modelr,r+ get_loggerrFrrOrrPr'rr*r;rrLrGrQrSr!r(r&r`sV == ]]D/I9OO-.4))MM   H %6       *.15%)$( 8*!#8* U3 ,- .8*S " 8* T%[ ! 8*zck TLL T-5eoo-F T\_ Tl 5-/Gl 5r(