3jSrSSKrSSKrSSKJrJrJr SSKJr SSK J r SSK J r J r JrJr SSKJrJrJr SS KJr SS KJr SS KJr S S KJrJr S SKJrJrJ r SSKJ!r! SSK"J#r#J$r$J%r%J&r&J'r'J(r(J)r)J*r*J+r+ \RX"\-5r.S\R^S\R^4Sjr0S\*\+SS4S\1\2\1-S\1S\3S\3S\RhS-4 Sjjr5SS\&\'S4S\1\2\1-S\R^S-S\1S\3S\RhS-4 Sjjr6SS\$\#\%S4S\1\2\1-S\R^\2\RnRn\RnRn4-S-S\1S\1S\3S\RhS-4 Sjjr8SFS \R^S!\RrS-S"\14S#jjr:SGS\R^S$\S!\RrS\RhS%\RvS&\3S"\14S'jjr<"S(S)\5r="S*S+\5r>"S,S-\5r?"S.S/\5r@"S0S1\5rA"S2S3\5rB"S4S5\5rC"S6S7\5rD"S8S9\5rE"S:S;\5rF"S<S=\5rG"S>S?\5rH"S@SA\5rI"SBSC\5rJ"SDSE\5rKg)Hz6 Text and VAE encoder blocks for QwenImage pipelines. N)"Qwen2_5_VLForConditionalGenerationQwen2TokenizerQwen2VLProcessor) FrozenDict)ClassifierFreeGuidance)InpaintProcessorVaeImageProcessoris_valid_imageis_valid_image_imagelist)AutoencoderKLQwenImageQwenImageControlNetModelQwenImageMultiControlNetModel)calculate_dimensions)logging) unwrap_module)ModularPipelineBlocks PipelineState) ComponentSpec InputParam OutputParam)QwenImageModularPipeline) QWENIMAGE_EDIT_PLUS_IMG_TEMPLATE#QWENIMAGE_EDIT_PLUS_PROMPT_TEMPLATE-QWENIMAGE_EDIT_PLUS_PROMPT_TEMPLATE_START_IDXQWENIMAGE_EDIT_PROMPT_TEMPLATE(QWENIMAGE_EDIT_PROMPT_TEMPLATE_START_IDX#QWENIMAGE_LAYERED_CAPTION_PROMPT_CN#QWENIMAGE_LAYERED_CAPTION_PROMPT_ENQWENIMAGE_PROMPT_TEMPLATE#QWENIMAGE_PROMPT_TEMPLATE_START_IDX hidden_statesmaskcUR5nURSS9nXn[R"XCR 5SS9nU$)Nrdimr)boolsumtorchsplittolist)r$r% bool_mask valid_lengthsselected split_results h/home/wildlama/miniconda3/lib/python3.13/site-packages/diffusers/modular_pipelines/qwenimage/encoders.py_extract_masked_hiddenr31sD IMMaM(M'H;;x)=)=)?QGL promptprompt_template_encode prompt_template_encode_start_idxtokenizer_max_lengthdevicecD[U[5(aU/OUnUnUnUV s/sHoRU 5PM n n U"XU-SSSS9RU5n U"U RU R SS9n U R Sn [XR 5nUV s/sHoUSPM nn UV s/sH@n [R"U RS5[RU RS9PMB nn [UV s/sHoRS5PM sn 5n[R"UVs/sHLn[R"UUR!UURS5- URS55/5PMN sn5n[R"UVs/sH<n[R"UUR!UURS5- 5/5PM> sn5nURUS 9nUU4$s sn fs sn fs sn fs sn fs snfs snf) NTpt) max_lengthpadding truncationreturn_tensors) input_idsattention_maskoutput_hidden_statesrdtyper:rr:) isinstancestrformattorArBr$r3r+onessizelongr:maxstackcat new_zeros) text_encoder tokenizerr6r7r8r9r:templatedrop_idxetxt txt_tokensencoder_hidden_statesr$split_hidden_statesattn_mask_list max_seq_lenu prompt_embedsencoder_attention_masks r2get_qwen_prompt_embedsra9s$FC00fXfF%H/H'- .v!??1 vC . x7RVgkbj)&&!00! *77;M0@Y@YZ1DE1DAXY<1DEXklXkSTejj%**QXXNXkNl*=>*=Qvvay*=>?KKKRefReQAq{{;#:AFF1IFG HRefM#[[GUV~!Aq{{;#:;< =~V"$$F$3M 0 001 /Fl>f Ws&HH 2AHH9AH'AHimagecT[U[5(aU/OUnUnUnUV s/sHoRU 5PM n n U"U USSS9RU5n U"U SU SU R S5U R S5SS9n U R S n [ XS5nUV s/sHoUSPM nn UV s/sH@n [R"U RS 5[RU RS 9PMB nn [UV s/sHoRS 5PM sn 5n[R"UVs/sHLn[R"UURUURS 5- URS 55/5PMN sn5n[R"UVs/sH<n[R"UURUURS 5- 5/5PM> sn5nURUS 9nUU4$s sn fs sn fs sn fs sn fs snfs snf)NTr<textimagesr>r@rArB pixel_valuesimage_grid_thwrArBrgrhrCrDrrErrG)rHrIrJrKgetr$r3r+rLrMrNr:rOrPrQrR)rS processorr6rbr7r8r:rUrVrWrX model_inputsoutputsr$r[r\r]r^r_r`s r2get_qwen_prompt_embeds_editrnas$FC00fXfF%H/H'- .v!??1 vC .    bj {+#$45!%%n5#''(89! G))"-M0M]@^_1DE1DAXY<1DEXklXkSTejj%**QXXNXkNl*=>*=Qvvay*=>?KKKRefReQAq{{;#:AFF1IFG HRefM#[[GUV~!Aq{{;#:;< =~V"$$F$3M 0 00? /&Fl>f Ws&H &H:AH HAH /AH%img_template_encodec [U[5(aU/OUn[U[5(a-Sn[U5HupXR U S-5- nM OUbUR S5nOSnUn Un UV s/sHoR X-5PM nn U"UUSSS9R U5nU"USUSUR S5UR S 5SS 9nURS n[UUS5nUV s/sHoU SPM nn UV s/sH@n [R"U RS 5[RU RS 9PMB nn [UV s/sHoRS 5PM sn 5n[R"UVs/sHLn[R "UUR#UURS 5- URS55/5PMN sn5n[R"UVs/sH<n[R "UUR#UURS 5- 5/5PM> sn5nUR US9nUU4$s sn fs sn fs sn fs sn fs snfs snf)NrTr<rdrArBrgrhrirDrrErG)rHrIlist enumeraterJrKrjr$r3r+rLrMrNr:rOrPrQrR)rSrkr6rbr7ror8r:base_img_promptiimgrUrVrWrXrlrmr$r[r\r]r^r_r`s r2 get_qwen_prompt_embeds_edit_plusrws`$FC00fXfF%&FA 99!a%@ @O'  -44Q7%H/H9? @A???. /C @    bj  {+#$45!%%n5#''(89! G))"-M0 M]@^_1DE1DAXY<1DEXklXkSTejj%**QXXNXkNl*=>*=Qvvay*=>?KKKRefReQAq{{;#:AFF1IFG HRefM#[[GUV~!Aq{{;#:;< =~V"$$F$3M 0 00; A$Fl>f Ws'<I(I-AI2(I7AI< AJencoder_output generator sample_modec[US5(a!US:XaURRU5$[US5(a US:XaURR5$[US5(a UR$[ S5e)N latent_distsampleargmaxlatentsz3Could not access latents of provided encoder_output)hasattrr|r}moderAttributeError)rxryrzs r2retrieve_latentsrs}~}--+2I))00;;  / /K84K))..00  + +%%%RSSr4vaerFlatent_channelsc [U[R5(d[S[ U5S35eUR 5S:XaUR S5nO1UR 5S:wa[SUR 5S35eURX4S9n[U[5(a][URS5Vs/sH$n[URXUS -5X'US 9PM& nn[R"USS 9nO[URU5X&S 9n[R"URR 5R#S US S S 5RUR$UR&5n [R"URR(5R#S US S S 5RUR$UR&5n X- U - nU$s snf) Nz#Expected image to be a tensor, got .rz Expected image dims 4 or 5, got )r:rFrr)ryrzr')rHr+Tensor ValueErrortyper( unsqueezerKrrrangeshaperencoderQtensorconfig latents_meanviewr:rF latents_std) rbrryr:rFrrzru image_latentsrrs r2encode_vae_imagers eU\\ * *>tE{m1MNN yy{a"  ;EIIK=JKK HHFH 0E)T""5;;q>* * SZZ!a%(89Y\_j k*   -Q7 (E):ii  SZZ,,- a!Q * M -"5"5 6  SZZ++, a!Q * M -"5"5 6 #1[@M ' s +G4c\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \S\ \4Sj5r\R""5S \S \4S j5rS rg )QwenImageEditResizeStepiac Image Resize step that resize the image to target area while maintaining the aspect ratio. Components: image_resize_processor (`VaeImageProcessor`) Inputs: image (`Image | list`): Reference image(s) for denoising. Can be a single image or list of images. Outputs: resized_image (`list`): The resized images qwenimage-editreturncg)NzZImage Resize step that resize the image to target area while maintaining the aspect ratio.selfs r2 description#QwenImageEditResizeStep.description*skr4c:[S[[SS05SS9/$Nimage_resize_processorvae_scale_factor from_configrdefault_creation_methodrr rrs r2expected_components+QwenImageEditResizeStep.expected_components.- (!!#5r":;(5    r4c0[R"S5/$NrbrrUrs r2inputsQwenImageEditResizeStep.inputs9s##G,--r4cX[S[[RRSS9/$N resized_imagezThe resized imagesname type_hintrrrrPILImagers r2intermediate_outputs,QwenImageEditResizeStep.intermediate_outputs=, $syy/0   r4 componentsstatec URU5nURn[U5(d[S[ U535e[ U5(aU/nUSR upV[SXV- 5upxn UV s/sHn URRXUS9PM n n Xl URX#5 X4$s sn f)N/Images must be image or list of images but are rheightwidth) get_block_staterbr rrr rMrrresizerset_block_state) rrr block_staterf image_width image_heightcalculated_widthcalculated_height_rbresized_imagess r2__call__ QwenImageEditResizeStep.__call__Gs**51 ""'//NtTZ|n]^ ^ & ! !XF$*1INN! 1EkS^Sm1n.Q    - - 4 4U\l 4 m  %3! U0   s>$B>rN__name__ __module__ __qualname____firstlineno____doc__ model_namepropertyrIrrrrrrrrrr+no_gradrrr__static_attributes__rr4r2rrs "J lSll T-%8  .Z(.. d;&7   ]]_!#;!M!!r4rc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \S\ \4Sj5r\S \4S j5r\R("5S \S \4S j5rSrg)QwenImageLayeredResizeStepiaa Image Resize step that resize the image to a target area (defined by the resolution parameter from user) while maintaining the aspect ratio. Components: image_resize_processor (`VaeImageProcessor`) Inputs: image (`Image | list`): Reference image(s) for denoising. Can be a single image or list of images. resolution (`int`, *optional*, defaults to 640): The target area to resize the image to, can be 1024 or 640 Outputs: resized_image (`list`): The resized images qwenimage-layeredrcg)NzImage Resize step that resize the image to a target area (defined by the resolution parameter from user) while maintaining the aspect ratio.rrs r2r&QwenImageLayeredResizeStep.descriptionvs^r4c:[S[[SS05SS9/$rrrs r2r.QwenImageLayeredResizeStep.expected_componentszrr4cN[R"S5[SS[SS9/$)Nrb resolutionz:The target area to resize the image to, can be 1024 or 640rdefaultrr)rrUintrs r2r!QwenImageLayeredResizeStep.inputss0    ( !X    r4cX[S[[RRSS9/$rrrs r2r/QwenImageLayeredResizeStep.intermediate_outputsrr4rc,US;a[SU35eg)N)r5rz&Resolution must be 1024 or 640 but is rrs r2 check_inputs'QwenImageLayeredResizeStep.check_inputss" [ (Ej\RS S )r4rrc URU5nURURS9 URn[ U5(d[ S[ U535e[U5(aU/nUSRupVURUR-n[XuU- 5upn UV s/sHn URRXUS9PM n n Xl URX#5 X4$s sn f)Nrrrr)rrrrbr rrr rMrrrrr) rrrrrfrr target_arearrrrbrs r2r#QwenImageLayeredResizeStep.__call__s**51  [%;%;<""'//NtTZ|n]^ ^ & ! !XF$*1INN! !,,{/E/EE 1EkamSm1n.Q    - - 4 4U\l 4 m  %3! U0   s0$C0rN)rrrrrrrrIrrrrrrrrr staticmethodrrr+rrrrrrr4r2rras$%J ^S^^ T-%8    Z(     d;&7  TTT ]]_!#;!M!!r4rc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \S\ \4Sj5r\R""5S \S \4S j5rS rg )QwenImageEditPlusResizeStepia Resize images for QwenImage Edit Plus pipeline. Produces two outputs: resized_image (1024x1024) for VAE encoding, resized_cond_image (384x384) for VL text encoding. Each image is resized independently based on its own aspect ratio. Components: image_resize_processor (`VaeImageProcessor`) Inputs: image (`Image | list`): Reference image(s) for denoising. Can be a single image or list of images. Outputs: resized_image (`list`): Images resized to 1024x1024 target area for VAE encoding resized_cond_image (`list`): Images resized to 384x384 target area for VL text encoding qwenimage-edit-plusrcg)NzResize images for QwenImage Edit Plus pipeline. Produces two outputs: resized_image (1024x1024) for VAE encoding, resized_cond_image (384x384) for VL text encoding. Each image is resized independently based on its own aspect ratio.rrs r2r'QwenImageEditPlusResizeStep.descriptions  Q r4c:[S[[SS05SS9/$rrrs r2r/QwenImageEditPlusResizeStep.expected_componentsrr4c0[R"S5/$rrrs r2r"QwenImageEditPlusResizeStep.inputss##G,--r4c[S[[RRSS9[S[[RRSS9/$)Nrz8Images resized to 1024x1024 target area for VAE encodingrresized_cond_imagez:Images resized to 384x384 target area for VL text encodingrrs r2r0QwenImageEditPlusResizeStep.intermediate_outputssL $syy/V  )syy/X   r4rrc URU5nURn[U5(d[S[ U535e[ U5(aU/n/n/nUHnUR up[SX- 5upn URURRX{U S95 [SX- 5upn URURRX~U S95 M XSl Xcl URX#5 X4$)Nrrri@)rrbr rrr rMrappendrrrrr)rrrrrfrresized_cond_imagesrbrr vae_width vae_heightrvl_width vl_heights r2r$QwenImageEditPlusResizeStep.__call__s**51 ""'//NtTZ|n]^ ^ & ! !XF E(- %K([Ul[Ul[TU]5 gN)r!image_caption_prompt_enr image_caption_prompt_cnsuper__init__r __class__s r2r+QwenImageLayeredGetImagePromptStep.__init__;s'J$'J$ r4rcg)NzAuto-caption step that generates a text prompt from the input image if none is provided. Uses the VL model (text_encoder) to generate a description of the image. If prompt is already provided, this step passes through unchanged.rrs r2r.QwenImageLayeredGetImagePromptStep.description@s  Q r4cB[S[5[S[5/$)NrSrk)rrrrs r2r6QwenImageLayeredGetImagePromptStep.expected_componentsHs$ .*L M +'7 8  r4c [R"SSS9[SS[RRSS9[SS[S S 9/$) Nr6F)requiredrTzIThe image to generate caption from, should be resized use the resize steprrrr use_en_promptz&Whether to use English prompt templater)rrUrrr)rs r2r)QwenImageLayeredGetImagePromptStep.inputsOsU   5  $))//g   $D    r4c"[S[SS9/$)Nr6z]The prompt or prompts to guide image generation. If not provided, updated using image captionr)rrIrs r2r7QwenImageLayeredGetImagePromptStep.intermediate_outputscs {   r4rrcURU5nURnURb URS:XdURS:XaUR(a URnO UR nUR UURSSS9RU5nURR"S 0UDSS0D6n[URU5VV s/sHupU [U5SPM n nn UR RU SSS 9S n U R5UlUR!X#5 X4$s sn nf) Nrq Tr<rdmax_new_tokensiF)skip_special_tokensclean_up_tokenization_spacesrr)r_execution_devicer6rrrrkrrKrSgenerateziprAlen batch_decodestripr) rrrrr:caption_promptrl generated_idsin_idsout_idsgenerated_ids_trimmed output_texts r2r+QwenImageLayeredGetImagePromptStep.__call__msO**51 --    %););r)A[EWEW[^E^((!%!=!=!%!=!=%//#"00# 0 bj '33<<`|`\_`M>A,BXBXZg>h%>h?6F &>h "%%..;;%4^c<K"-!2!2!4K  U0  %sD=)rr)rrrrrrrrrIrrrrrrrrrr+rrrrr __classcell__rs@r2r r #s*%J  S   T-%8   Z(  & d;&7   ]]_!#;!M!Vc!!r4r c^\rSrSrSrSrU4Sjr\S\4Sj5r \S\ \ 4Sj5r \S\ \ 4Sj5r\S\ \4S j5r\S 5r\R("5S \S \4S j5rSrU=r$)QwenImageTextEncoderStepia Text Encoder step that generates text embeddings to guide the image generation. Components: text_encoder (`Qwen2_5_VLForConditionalGeneration`): The text encoder to use tokenizer (`Qwen2Tokenizer`): The tokenizer to use guider (`ClassifierFreeGuidance`) Inputs: prompt (`str`): The prompt or prompts to guide image generation. negative_prompt (`str`, *optional*): The prompt or prompts not to guide the image generation. max_sequence_length (`int`, *optional*, defaults to 1024): Maximum sequence length for prompt encoding. Outputs: prompt_embeds (`Tensor`): The prompt embeddings. prompt_embeds_mask (`Tensor`): The encoder attention mask. negative_prompt_embeds (`Tensor`): The negative prompt embeddings. negative_prompt_embeds_mask (`Tensor`): The negative prompt embeddings mask. qwenimagec\>[Ul[UlSUl[ TU]5 g)Nr5)r"r7r#r8r9rrrs r2r!QwenImageTextEncoderStep.__init__s&&?#0S-$(! r4rcg)NzOText Encoder step that generates text embeddings to guide the image generation.rrs r2r$QwenImageTextEncoderStep.descriptions`r4c r[S[SS9[S[SS9[S[[ SS05S S 9/$) NrSzThe text encoder to use)rrTzThe tokenizer to useguiderguidance_scale@rr)rrrrrrs r2r,QwenImageTextEncoderStep.expected_componentssH .*LZs t +~CY Z &!#3S"9:(5    r4c[R"S5[R"S5[R"SSS9/$)Nr6negative_promptmax_sequence_lengthr5)rrrs r2rQwenImageTextEncoderStep.inputss<    )    1 2    5t D  r4c[R"S5[R"S5[R"S5[R"S5/$Nr_prompt_embeds_masknegative_prompt_embedsnegative_prompt_embeds_maskrrUrs r2r-QwenImageTextEncoderStep.intermediate_outputsI   1  !5 6  !9 :  !> ?   r4c>[U[5(d,[U[5(d[S[ U535eUbA[U[5(d,[U[5(d[S[ U535eUbUS:a[SU35egg)N2`prompt` has to be of type `str` or `list` but is ;`negative_prompt` has to be of type `str` or `list` but is r5z9`max_sequence_length` cannot be greater than 1024 but is rHrIrrrr)r6rArBs r2r%QwenImageTextEncoderStep.check_inputss&#&&z&$/G/GQRVW]R^Q_`a a  '4455Z[_`o[pZqrs s  */BT/IXYlXmno o0J *r4rrc URU5nURnURURURUR 5 [ URURURURURURUS9uUl Ul URSS2SUR 24Ul URSS2SUR 24Ul SUlSUlUR (aUR=(d Sn[ URURUURURURUS9uUlUlURSS2SUR 24UlURSS2SUR 24UlUR#X#5 X4$)N)r6r7r8r9r:rq)rr%rr6rArBrarSrTr7r8r9r_rFrGrHrequires_unconditional_embedsrrrrrr:rAs r2r!QwenImageTextEncoderStep.__call__s**51 -- +,,k.I.I;KjKjkDZ  # #  %%#'#>#>-1-R-R!%!:!:E A !;#A%0$=$=aAb;CbCbAb>b$c !)4)G)GKl[MlMlKlHl)m &-1 *26 /  3 3)99?ROZp''$$&'+'B'B151V1V%)%>%>[ WK . 0W2=1S1S4[44442K .7B6]6]4[44447K 3 U0  r4)r7r8r9rrrrrrrrrIrrrrrrrrrrrr+rrrrrr2r3s@r2r5r5s4J aSaa  T-%8     Z(   d;&7   p p ]]_(!#;(!M(!(!r4r5c^\rSrSrSrSrU4Sjr\S\4Sj5r \S\ \ 4Sj5r \S\ \ 4Sj5r\S\ \4S j5r\S 5r\R("5S \S \4S j5rSrU=r$)QwenImageEditTextEncoderStepia Text Encoder step that processes both prompt and image together to generate text embeddings for guiding image generation. Components: text_encoder (`Qwen2_5_VLForConditionalGeneration`) processor (`Qwen2VLProcessor`) guider (`ClassifierFreeGuidance`) Inputs: prompt (`str`): The prompt or prompts to guide image generation. negative_prompt (`str`, *optional*): The prompt or prompts not to guide the image generation. resized_image (`Image`): The image prompt to encode, should be resized using resize step Outputs: prompt_embeds (`Tensor`): The prompt embeddings. prompt_embeds_mask (`Tensor`): The encoder attention mask. negative_prompt_embeds (`Tensor`): The negative prompt embeddings. negative_prompt_embeds_mask (`Tensor`): The negative prompt embeddings mask. r6cN>[Ul[Ul[TU]5 gr )rr7rr8rrrs r2r%QwenImageEditTextEncoderStep.__init__4s&D#0X- r4rcg)NzyText Encoder step that processes both prompt and image together to generate text embeddings for guiding image generation.rrs r2r(QwenImageEditTextEncoderStep.description9sKr4c v[S[5[S[5[S[[ SS05SS9/$NrSrkr<r=r>rrrrrrrrs r2r0QwenImageEditTextEncoderStep.expected_components=C .*L M +'7 8 &!#3S"9:(5    r4c [R"S5[R"S5[SS[RRSS9/$)Nr6rArTz?The image prompt to encode, should be resized using resize steprrrUrrrs r2r#QwenImageEditTextEncoderStep.inputsJsG    )    1 2 $))//]    r4c[R"S5[R"S5[R"S5[R"S5/$rErIrs r2r1QwenImageEditTextEncoderStep.intermediate_outputsWrKr4c[U[5(d,[U[5(d[S[ U535eUbC[U[5(d-[U[5(d[S[ U535egggNrMrNrOr6rAs r2r)QwenImageEditTextEncoderStep.check_inputs`&#&&z&$/G/GQRVW]R^Q_`a a  '4455Z[_`o[pZqrs s65 (r4rrc dURU5nURURUR5 URn[ UR URURURURURUS9uUl Ul SUl SUlUR(acUR=(d Sn[ UR URUURURURUS9uUl UlUR!X#5 X4$)N)r6rbr7r8r:r!)rrr6rAr%rnrSrkrr7r8r_rFrGrHrRrrSs r2r%QwenImageEditTextEncoderStep.__call__ls**51  +,,k.I.IJ--D_  # #  %%++#'#>#>-1-R-RE A !;#A.2 *26 /  3 3)99@SOZu''$$&!//'+'B'B151V1V[ WK . 0W U0  r4)r7r8rUr3s@r2rWrWs6J KSKK  T-%8      Z(     d;&7   t t ]]_ !#; !M ! !r4rWc^\rSrSrSrSrU4Sjr\S\4Sj5r \S\ \ 4Sj5r \S\ \ 4Sj5r\S\ \4S j5r\S 5r\R("5S \S \4S j5rSrU=r$) QwenImageEditPlusTextEncoderStepia Text Encoder step for QwenImage Edit Plus that processes prompt and multiple images together to generate text embeddings for guiding image generation. Components: text_encoder (`Qwen2_5_VLForConditionalGeneration`) processor (`Qwen2VLProcessor`) guider (`ClassifierFreeGuidance`) Inputs: prompt (`str`): The prompt or prompts to guide image generation. negative_prompt (`str`, *optional*): The prompt or prompts not to guide the image generation. resized_cond_image (`Tensor`): The image(s) to encode, can be a single image or list of images, should be resized to 384x384 using resize step Outputs: prompt_embeds (`Tensor`): The prompt embeddings. prompt_embeds_mask (`Tensor`): The encoder attention mask. negative_prompt_embeds (`Tensor`): The negative prompt embeddings. negative_prompt_embeds_mask (`Tensor`): The negative prompt embeddings mask. rcd>[Ul[Ul[Ul[ TU]5 gr )rr7rrorr8rrrs r2r)QwenImageEditPlusTextEncoderStep.__init__s&&I##C 0]- r4rcg)NzText Encoder step for QwenImage Edit Plus that processes prompt and multiple images together to generate text embeddings for guiding image generation.rrs r2r,QwenImageEditPlusTextEncoderStep.descriptions  H r4c v[S[5[S[5[S[[ SS05SS9/$r]r^rs r2r4QwenImageEditPlusTextEncoderStep.expected_componentsr`r4c [R"S5[R"S5[SS[RSS9/$)Nr6rArTzoThe image(s) to encode, can be a single image or list of images, should be resized to 384x384 using resize stepr)rrUr+rrs r2r'QwenImageEditPlusTextEncoderStep.inputssF    )    1 2 ),,N    r4c[R"S5[R"S5[R"S5[R"S5/$rErIrs r2r5QwenImageEditPlusTextEncoderStep.intermediate_outputsrKr4c[U[5(d,[U[5(d[S[ U535eUbC[U[5(d-[U[5(d[S[ U535egggrgrOrhs r2r-QwenImageEditPlusTextEncoderStep.check_inputsrjr4rrc URU5nURURUR5 URn[ UR URURURURURURUS9uUl Ul SUlSUlUR (anUR=(d Sn[ UR URUURURURURUS9uUlUlUR#X#5 X4$)N)r6rbr7ror8r:r!)rrr6rAr%rwrSrkrr7ror8r_rFrGrHrRrrSs r2r)QwenImageEditPlusTextEncoderStep.__call__s)**51  +,,k.I.IJ--Dd  # #  %%00#'#>#> $ 8 8-1-R-R E A !;#A.2 *26 /  3 3)99@SO0++((*%88+/+F+F(,(@(@595Z5Z!  XK . 0W U0  r4)ror7r8rUr3s@r2rnrns8'J  S    T-%8      Z(     d;&7   t t ]]_$!#;$!M$!$!r4rnc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \S\ \4Sj5r\S 5r\R&"5S \S \4S j5rS rg)&QwenImageInpaintProcessImagesInputStepia Image Preprocess step for inpainting task. This processes the image and mask inputs together. Images will be resized to the given height and width. Components: image_mask_processor (`InpaintProcessor`) Inputs: mask_image (`Image`): Mask image for inpainting. image (`Image | list`): Reference image(s) for denoising. Can be a single image or list of images. height (`int`, *optional*): The height in pixels of the generated image. width (`int`, *optional*): The width in pixels of the generated image. padding_mask_crop (`int`, *optional*): Padding for mask cropping in inpainting. Outputs: processed_image (`Tensor`): The processed image processed_mask_image (`Tensor`): The processed mask image mask_overlay_kwargs (`dict`): The kwargs for the postprocess step to apply the mask overlay r6rcg)NzImage Preprocess step for inpainting task. This processes the image and mask inputs together. Images will be resized to the given height and width.rrs r2r2QwenImageInpaintProcessImagesInputStep.description9ser4c:[S[[SS05SS9/$Nimage_mask_processorrrrrrr rrs r2r:QwenImageInpaintProcessImagesInputStep.expected_components=- & !#5r":;(5    r4c[R"S5[R"S5[R"S5[R"S5[R"S5/$)N mask_imagerbrrpadding_mask_croprrs r2r-QwenImageInpaintProcessImagesInputStep.inputsHsU    -    (    )    (    3 4   r4c[S[RSS9[S[RSS9[S[SS9/$Nprocessed_imageThe processed imagerprocessed_mask_imagezThe processed mask imagemask_overlay_kwargsz=The kwargs for the postprocess step to apply the mask overlayrr+rdictrs r2r;QwenImageInpaintProcessImagesInputStep.intermediate_outputsRsM &,,1  +,,6  *[   r4cUbXS--S:wa[SUS-SU35eUb XS--S:wa[SUS-SU35eggNrrzHeight must be divisible by z but is zWidth must be divisible by rrrrs r2r3QwenImageInpaintProcessImagesInputStep.check_inputsf{  &q,@"AQ"F;!?1!D:;Ka;O:PPXY^X_`a a"E r4rrcURU5nURURURURS9 UR=(d UR nUR=(d UR nURRURURUUURS9uUl Ul UlURX#5 X4$)Nrrbr%rrr)rrrrrdefault_height default_widthr preprocessrbrrrrrrrrrrrrs r2r/QwenImageInpaintProcessImagesInputStep.__call__ns**51  %%[->->Q[QlQl  ##@z'@'@!!=Z%=%=  + + 6 6!'' ++"-"?"? 7  g #[%E{Gf U0  r4rNrrrrrrrrIrrrrrrrrrrrr+rrrrrrr4r2r~r~s8J eSee T-%8   Z(   d;&7  &bb ]]_!#;!M!!r4r~c\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \S\ \4Sj5r\R""5S \S \4S j5rS rg )*QwenImageEditInpaintProcessImagesInputStepia Image Preprocess step for inpainting task. This processes the image and mask inputs together. Images should be resized first. Components: image_mask_processor (`InpaintProcessor`) Inputs: mask_image (`Image`): Mask image for inpainting. resized_image (`Image`): The resized image. should be generated using a resize step padding_mask_crop (`int`, *optional*): Padding for mask cropping in inpainting. Outputs: processed_image (`Tensor`): The processed image processed_mask_image (`Tensor`): The processed mask image mask_overlay_kwargs (`dict`): The kwargs for the postprocess step to apply the mask overlay rrcg)Nz}Image Preprocess step for inpainting task. This processes the image and mask inputs together. Images should be resized first.rrs r2r6QwenImageEditInpaintProcessImagesInputStep.descriptionsOr4c:[S[[SS05SS9/$rrrs r2r>QwenImageEditInpaintProcessImagesInputStep.expected_componentsrr4c[R"S5[SS[RRSS9[R"S5/$)NrrT:The resized image. should be generated using a resize steprrrbrs r2r1QwenImageEditInpaintProcessImagesInputStep.inputssI    - $))//X      3 4  r4c[S[RSS9[S[RSS9[S[SS9/$rrrs r2r?QwenImageEditInpaintProcessImagesInputStep.intermediate_outputssH .%,,Ti j +,,6  *[   r4rrcURU5nURSRupEURR URUR UUUR S9uUlUlUl URX#5 X4$)Nrr) rrrMrrrrrrrrrrrrrrs r2r3QwenImageEditInpaintProcessImagesInputStep.__call__s**51 #11!499   + + 6 6!// ++"-"?"? 7  g #[%E{Gf U0  r4rNrrr4r2rrs0"J OSOO T-%8    Z(      d;&7     ]]_!#;!M!!r4rc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \S\ \4Sj5r\S 5r\R&"5S \S \4S j5rS rg)QwenImageProcessImagesInputStepia Image Preprocess step. will resize the image to the given height and width. Components: image_processor (`VaeImageProcessor`) Inputs: image (`Image | list`): Reference image(s) for denoising. Can be a single image or list of images. height (`int`, *optional*): The height in pixels of the generated image. width (`int`, *optional*): The width in pixels of the generated image. Outputs: processed_image (`Tensor`): The processed image r6rcg)NzKImage Preprocess step. will resize the image to the given height and width.rrs r2r+QwenImageProcessImagesInputStep.descriptions\r4c:[S[[SS05SS9/$Nimage_processorrrrrrrs r2r3QwenImageProcessImagesInputStep.expected_components- !!!#5r":;(5    r4c[R"S5[R"S5[R"S5/$)Nrbrrrrs r2r&QwenImageProcessImagesInputStep.inputss8    (    )    (  r4c6[S[RSS9/$Nrrrrr+rrs r2r4QwenImageProcessImagesInputStep.intermediate_outputs# &,,1   r4cUbXS--S:wa[SUS-SU35eUb XS--S:wa[SUS-SU35eggrrrs r2r,QwenImageProcessImagesInputStep.check_inputsrr4rrczURU5nURURURURS9 UR=(d UR nUR=(d UR nURRURUUS9Ul URX#5 X4$)Nrrbrr) rrrrrrrrrrbrrrs r2r(QwenImageProcessImagesInputStep.__call__"s**51  %%[->->Q[QlQl  ##@z'@'@!!=Z%=%=&0&@&@&K&K##'L' # U0  r4rNrrr4r2rrs&J ]S]] T-%8   Z(   d;&7  bb ]]_!#;!M!!r4rc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \S\ \4Sj5r\R""5S \S \4S j5rS rg )#QwenImageEditProcessImagesInputStepi7a/ Image Preprocess step. Images needs to be resized first. Components: image_processor (`VaeImageProcessor`) Inputs: resized_image (`list`): The resized image. should be generated using a resize step Outputs: processed_image (`Tensor`): The processed image rrcg)Nz8Image Preprocess step. Images needs to be resized first.rrs r2r/QwenImageEditProcessImagesInputStep.descriptionIsIr4c:[S[[SS05SS9/$rrrs r2r7QwenImageEditProcessImagesInputStep.expected_componentsMrr4cZ[SS[[RRSS9/$NrTrrrrrrrrs r2r*QwenImageEditProcessImagesInputStep.inputsX/ $syy/X    r4c6[S[RSS9/$rrrs r2r8QwenImageEditProcessImagesInputStep.intermediate_outputscrr4rrcURU5nURSRupEURR URUUS9UlUR X#5 X4$)Nrr)rrrMrrrrrs r2r,QwenImageEditProcessImagesInputStep.__call__msq**51 #11!499 &0&@&@&K&K++'L' # U0  r4rNrrr4r2rr7s "J JSJJ T-%8   Z(   d;&7   ]]_ !#; !M ! !r4rc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \S\ \4Sj5r\R""5S \S \4S j5rS rg )'QwenImageEditPlusProcessImagesInputStepi~as Image Preprocess step. Images can be resized first. If a list of images is provided, will return a list of processed images. Components: image_processor (`VaeImageProcessor`) Inputs: resized_image (`list`): The resized image. should be generated using a resize step Outputs: processed_image (`Tensor`): The processed image rrcg)Nz|Image Preprocess step. Images can be resized first. If a list of images is provided, will return a list of processed images.rrs r2r3QwenImageEditPlusProcessImagesInputStep.descriptionsNr4c:[S[[SS05SS9/$rrrs r2r;QwenImageEditPlusProcessImagesInputStep.expected_componentsrr4cZ[SS[[RRSS9/$rrrs r2r.QwenImageEditPlusProcessImagesInputStep.inputsrr4c6[S[RSS9/$rrrs r2rUc[SS[RSS9nUc[R"S5n[ U[5(d[ S[U535e[ U[5(d[ S[U535eXlX l URUl URUl [TU]95 g) aInitialize a VAE encoder step for converting images to latent representations. Handles both single images and lists of images. When input is a list, outputs a list of latents. When input is a single tensor, outputs a single latent tensor. Args: input (InputParam, optional): Input parameter for the processed image. Defaults to "processed_image". output (OutputParam, optional): Output parameter for the image latents. Defaults to "image_latents". NrTzThe image tensor to encoderrz input must be InputParam but is z"output must be OutputParam but is )rr+rrrUrHrr_input_outputr_image_input_name_image_latents_output_namerr)rrrrs r2r QwenImageVaeEncoderStep.__init__s =&[wE > ))/:F%,,?U }MN N&+..A$v,PQ Q  !&*0++' r4rc<SURSURS3$)NzVAE Encoder step that converts z into latent representations zI. Handles both single images and lists of images with varied resolutions.)rrrs r2r#QwenImageVaeEncoderStep.descriptions?.d.D.D-EEbcgdCdCcDDV V r4c$[S[5/$)Nr)rr rs r2r+QwenImageVaeEncoderStep.expected_componentsse%;<==r4cFUR[R"S5/$)Nry)rrrUrs r2rQwenImageVaeEncoderStep.inputss# KK    ,  r4cUR/$r )rrs r2r,QwenImageVaeEncoderStep.intermediate_outputss ~r4rrc URU5nURnURRn[ X0R 5n[ U[5nU(dU/n/nUH>n UR[U URURUUURS95 M@ U(dUSn[X0RU5 URX#5 X4$)N)rbrryr:rFrr)rr%rrFgetattrrrHrrrrrynum_channels_latentssetattrrr) rrrrr:rFrbrrrvs r2r QwenImageVaeEncoderStep.__call__s**51 --$$ %;%;<"5$/ GE C   ")33!$.$C$C  )!,M <T-%8>> Z(  d;&7 ]]_!#;!M!Vc!!r4rc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \S\ \4Sj5r\S 5r\R&"5S \S \S\4S j5rS rg)!QwenImageControlNetVaeEncoderStepiAa VAE Encoder step that converts `control_image` into latent representations control_image_latents. Components: vae (`AutoencoderKLQwenImage`) controlnet (`QwenImageControlNetModel`) control_image_processor (`VaeImageProcessor`) Inputs: control_image (`Image`): Control image for ControlNet conditioning. height (`int`, *optional*): The height in pixels of the generated image. width (`int`, *optional*): The width in pixels of the generated image. generator (`Generator`, *optional*): Torch generator for deterministic generation. Outputs: control_image_latents (`Tensor`): The latents representing the control image r6rcg)NzbVAE Encoder step that converts `control_image` into latent representations control_image_latents. rrs r2r-QwenImageControlNetVaeEncoderStep.descriptionZstr4c z[S[5[S[5[S[[ SS05SS9/nU$)Nr controlnetcontrol_image_processorrrrr)rr rr r)rrs r2r5QwenImageControlNetVaeEncoderStep.expected_components^sH %!7 8 ,(@ A )!!#5r":;(5    r4c[R"S5[R"S5[R"S5[R"S5/nU$)N control_imagerrryr)rrs r2r(QwenImageControlNetVaeEncoderStep.inputslsJ    0    )    (    ,    r4c6[S[RSS9/$)Ncontrol_image_latentsz*The latents representing the control image)rrrrs r2r6QwenImageControlNetVaeEncoderStep.intermediate_outputsvs# ',,H   r4cUbXS--S:wa[SUS-SU35eUb XS--S:wa[SUS-SU35eggrrrs r2r.QwenImageControlNetVaeEncoderStep.check_inputsrr4rrc RURU5nURURURUR5 UR nUR RnUR=(d URnUR=(d URn[UR5n[U[5(a1[UR[5(dUR/Ul[U[5(a~/UlURHfn UR"R%U UUS9n ['U UR UR(UUUR*SS9n UR R-U 5 Mh O[U[.5(aXUR"R%URUUS9n ['U UR UR(UUUR*SS9UlO[1S[3U535eUR5X#5 X4$)Nrr})rbrryr:rFrrzz[Expected controlnet to be a QwenImageControlNetModel or QwenImageMultiControlNetModel, got )rrrrrr%rrFrrrrrHrrrrrrrrryrrrrrr) rrrrr:rFrrrcontrol_image_control_image_latents_rs r2r*QwenImageControlNetVaeEncoderStep.__call__s**51  +,,k.?.?A\A\]--$$##@z'@'@!!=Z%=%=":#8#89 j"? @ @T_TmTmosItIt)4)B)B(CK % j"? @ @02K -"-";";!+!C!C!N!N(!"O" *:(")33!$.$C$C (*&11889OP!#<$ $< = =&>>II!//JM 1A#NN%// * ? ?$1K -mnrs}n~mA  U0  r4rNrrr4r2rrAs,J uSuu T-%8  Z( d;&7  bb ]]_:!#;:!M:!Vc:!:!r4rc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \R"5S\S\4S j5rS rg ) "QwenImageLayeredPermuteLatentsStepiaz Permute image latents from (B, C, 1, H, W) to (B, 1, C, H, W) for Layered packing. Inputs: image_latents (`Tensor`): image latents used to guide the image generation. Can be generated from vae_encoder step. Outputs: image_latents (`Tensor`): The latent representation of the input image. (permuted from [B, C, 1, H, W] to [B, 1, C, H, W]) rrcg)NzRPermute image latents from (B, C, 1, H, W) to (B, 1, C, H, W) for Layered packing.rrs r2r.QwenImageLayeredPermuteLatentsStep.descriptionscr4c0[R"S5/$)Nrrrs r2r)QwenImageLayeredPermuteLatentsStep.inputss    0  r4c.[R"SSS9/$)Nrz0permuted from [B, C, 1, H, W] to [B, 1, C, H, W])noterIrs r2r7QwenImageLayeredPermuteLatentsStep.intermediate_outputss  7i j  r4rcURU5nURnURSSSSS5UlURX#5 X4$)Nrrrrr)rrpermuter)rrrrrs r2r+QwenImageLayeredPermuteLatentsStep.__call__sO**51 ++$+OOAq!Q$B ! U0  r4rN)rrrrrrrrIrrrrrrrr+rrrrrr4r2rrs %J dSdd Z(   d;&7   ]]_!-!M!!r4r)Nr})rr~)Lrrr+ transformersrrrconfiguration_utilsrguidersrrr r r r modelsr rr+pipelines.qwenimage.pipeline_qwenimage_editrutilsrutils.torch_utilsrmodular_pipelinerrmodular_pipeline_utilsrrrrprompt_templatesrrrrrr r!r"r# get_loggerrloggerrr3rIrrrr:rarnrrw GeneratorrrFrrrrr r5rWrnr~rrrrrrrrr4r2r'sI ]]--lleeO.CKK6      H %%,,ell#";,O $"&%1 $s)O%1 %1 '* %1  %1 LL4 %1V#!%"@,T"&,1 $s)O,1 <<$  ,1  ,1 '* ,1 LL4 ,1d#JN"E?,Y"&41 $s)O41 <<$syy ?@ @4 G 41  41  41'*41 LL4 41r`h TLL T-2__t-C TY\ T(( <<( (( LL ( ;; (  ((DE!3E!RX!!6X!x]!"7]!Li!)>i!d}!4}!Bw!#8w!v@!'<@!Ri!-Bi!ZX!1FX!xP!&;P!hC!*?C!NN!.CN!nh!3h!XB!(=B!V(!)>(!r4