o i1@s<dZddlZddlmZmZmZddlmZmZm Z m Z ddl m Z ddl mZddlmZeedd dd d Z d edBd efddZdeded efddZde dedBdedBdejdededejdBdejdBdejdBdejdBded eejdBejdBffdd ZGd!d"d"ZGd#d$d$ZGd%d&d&ZdS)'uKFlat denoiser classes — transformer received at call time, not stored. Three implementations of the :class:`~ltx_pipelines.utils.types.Denoiser` protocol: * :class:`SimpleDenoiser` — single transformer call, no guidance. * :class:`GuidedDenoiser` — static guiders, handles CFG + STG + isolated modality. * :class:`FactoryGuidedDenoiser` — resolves guiders per-step from sigma. ``GuidedDenoiser`` and ``FactoryGuidedDenoiser`` share the core multi-pass logic via the module-level :func:`_guided_denoise` function, which batches all guidance passes into a single transformer call. N)MultiModalGuiderMultiModalGuiderFactoryMultiModalGuiderParams)BatchedPerturbationConfig PerturbationPerturbationConfigPerturbationType)X0Model) LatentState)modality_from_latent_stateg?) cfg_scale stg_scalemodality_scale)paramsguiderreturncCs|dur|StS)zIReturn the guider as-is, or a positive-only guider for absent modalities.N)_POSITIVE_ONLY_GUIDER)rrN/home/ubuntu/LTX-2/packages/ltx-pipelines/src/ltx_pipelines/utils/denoisers.py_ensure_guidersrstatencs\dtjdtjffdd }t||j||j||j||j|jdur*||jdSddS)uRepeat a ``LatentState`` *n* times along the batch dimension. ``(B, ...) → (n*B, ...)`` by tiling the whole tensor n times, so the ordering is ``[item0, item1, ..., item0, item1, ...]`` — matching ``torch.cat`` of n per-pass contexts. trcs dg|}|d<||S)Nr)dimrepeat)rrepeatsrrr_repeat*s z_repeat_state.._repeatN)latent denoise_mask positions clean_latentattention_mask)torchTensorr r r!r"r#r$)rrrrrr _repeat_state#sr' transformer video_state audio_statesigma video_guider audio_guider v_context a_contextlast_denoised_videolast_denoised_audio step_indexc& sP|| } || } | r| r|| fS|dur|durtd|dur*|dur*tdtttjdBtjdBtf} d||tfg}|sI|r|rV|j durVtd|rc|j durctd|j durk|j n|}|j duru|j n|}| d||tfg}| r| t t j|jjd| r| t t j|jjd|r| d||t|f|s|r| d ||tt t jddt t jddgfd d |D}d d |D}t|d tdtjffdd }d}|durtjdd |Ddd}tt||||| d}d}|dur1tjdd |Ddd}tt||||| d}|||t|d\}}|durHt|ndg}|durYt|ndg}tt|t||dddd}|d\}}|dd\}}|dd\} }!|d d\}"}#| r|n| ||| |"}$| r| n| |||!|#}%|$|%fS)uZCore guided denoising — batches all guidance passes into one transformer call. Collects per-pass contexts first, then builds a single batched Modality per present modality via :func:`modality_from_latent_state`. When wrapped with :class:`~ltx_core.batch_split.BatchSplitAdapter`, the transformer may split this batch into sequential chunks internally. Guiders must not be ``None``. For absent modalities, callers should pass :data:`_POSITIVE_ONLY_GUIDER` (via :func:`_ensure_guider`) so that only the conditioned pass runs and ``calculate()`` returns cond unchanged. Nz2v_context is required when video_state is providedz2a_context is required when audio_state is providedcondz8Negative context is required for unconditioned denoisinguncond)typeblocksptbmodcSsg|]\}}}}|qSrr).0name_rrr z#_guided_denoise..cSsg|]\}}}}|qSrr)r9r;r7rrrr<r=rrcs|jjdS)z._batched_sigmacSsg|]\}}}}|qSrr)r9r;vcrrrr<r=r)r)enabledcSsg|]\}}}}|qSrr)r9r;acrrrr<r=videoaudio perturbationsr T)strict)r r )!should_skip_step ValueErrortuplestrr%r&remptydo_unconditional_generationnegative_contextappenddo_perturbed_generationrrSKIP_VIDEO_SELF_ATTNr stg_blocksSKIP_AUDIO_SELF_ATTNdo_isolated_modality_generationSKIP_A2V_CROSS_ATTNSKIP_V2A_CROSS_ATTNlenr catr r'rlistchunkdictzipget calculate)&r(r)r*r+r,r-r.r/r0r1r2v_skipa_skip_passpassesv_nega_negstg_perturbations pass_names ptb_configsrA batched_video batched_audioall_vall_asplits_vsplits_arcond_vcond_auncond_vuncond_aptb_vptb_amod_vmod_adenoised_videodenoised_audiorr@r_guided_denoise8s        "" r{c@sreZdZdZdejdBdejdBddfddZded edBd edBd ejd e de ejdBejdBff d dZ dS)SimpleDenoiserz^Single transformer call, no guidance. Passes ``None`` Modality for absent modalities. r.Nr/rcCs||_||_dSN)r.r/)selfr.r/rrr__init__s zSimpleDenoiser.__init__r(r)r*sigmasr2c CsJ||}|durt||j|nd}|durt||j|nd}|||ddS)NrE)r r.r/) r~r(r)r*rr2r+ pos_video pos_audiorrr__call__szSimpleDenoiser.__call__) __name__ __module__ __qualname____doc__r%r&rr r intrLrrrrrr|s, r|c@eZdZdZ  ddejdBdejdBdedBdedBddf dd Zd ed e dBd e dBd ejde de ejdBejdBff ddZ dS)GuidedDenoiseruStatic guiders — handles CFG + STG + isolated modality. Context/guider can be ``None`` for absent modalities (a positive-only guider is substituted at call time). Nr.r/r,r-rcCs(||_||_||_||_d|_d|_dSr})r.r/r,r-_last_denoised_video_last_denoised_audio)r~r.r/r,r-rrrrs  zGuidedDenoiser.__init__r(r)r*rr2c CsNt|||||t|jt|j|j|j|j|j|d \}}||_||_||fSN) r(r)r*r+r,r-r.r/r0r1r2)r{rr,r-r.r/rr)r~r(r)r*rr2ryrzrrrrs   zGuidedDenoiser.__call__NN) rrrrr%r&rrr r rrLrrrrrrs8  rc@r)FactoryGuidedDenoiserzLResolves guiders per-step from sigma, then delegates to shared guided logic.Nr.r/video_guider_factoryaudio_guider_factoryrcCs.||_||_||_||_d|_d|_d|_dSr})r.r/rrrr_sigma_vals_cached)r~r.r/rrrrrrs zFactoryGuidedDenoiser.__init__r(r)r*rr2c Cs|jdur||_|j|}t|jr|j|nd}t|js'|jr0|jp,|j|nd}t||||||||j |j |j |j |d \} } | |_ | |_ | | fSr) rdetachcputolistrrbuild_from_sigmarr{r.r/rr) r~r(r)r*rr2 sigma_valr,r-ryrzrrrr s:    zFactoryGuidedDenoiser.__call__r) rrrrr%r&rrr r rrLrrrrrrs8 r)rr%ltx_core.components.guidersrrrltx_core.guidance.perturbationsrrrrltx_core.model.transformerr ltx_core.typesr ltx_pipelines.utils.helpersr rrrr'r&rLr{r|rrrrrrsR         z.