o GiF@@sdZddlmZddlZddlmZmZddlmZm Z ddl m Z dd l m Z e eZeGd d d eZGd d d e eZdS)aM LTXEulerAncestralRFScheduler This scheduler implements a K-diffusion style Euler-Ancestral sampler specialized for flow / CONST parameterization, closely mirroring ComfyUI's `sample_euler_ancestral_RF` implementation used for LTX-Video. Reference implementation (ComfyUI): comfy.k_diffusion.sampling.sample_euler_ancestral_RF ) dataclassN) ConfigMixinregister_to_config) BaseOutputlogging) randn_tensor)SchedulerMixinc@seZdZUdZejed<dS)"LTXEulerAncestralRFSchedulerOutputz Output class for the scheduler's `step` function output. Args: prev_sample (`torch.FloatTensor`): Updated sample for the next step in the denoising process. prev_sampleN)__name__ __module__ __qualname____doc__torch FloatTensor__annotations__rrj/home/ubuntu/.local/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ltx_euler_ancestral_rf.pyr &s r c@sneZdZdZdgZdZe   d.dededefd d Z e d efd d Z e d efddZ d/defddZ d0deejBdejdBd efddZdeejBfddZ     d1dedBdeejBdBdeeejBdBdeeejBdBdedBf d d!Zd"ejd#ejd ejfd$d%Z  &d2d'ejdeejBd#ejd(ejdBd)ed eeejBf d*d+Zd efd,d-ZdS)3LTXEulerAncestralRFScheduleraI Euler-Ancestral scheduler for LTX-Video (RF / CONST parametrization). This scheduler is intended for models where the network is trained with a CONST-like parameterization (as in LTXV / FLUX). It approximates ComfyUI's `sample_euler_ancestral_RF` sampler and is useful when reproducing ComfyUI workflows inside diffusers. The scheduler can either: - reuse the [`FlowMatchEulerDiscreteScheduler`] sigma / timestep logic when only `num_inference_steps` is provided (default diffusers-style usage), or - follow an explicit ComfyUI-style sigma schedule when `sigmas` (or `timesteps`) are passed to [`set_timesteps`]. Args: num_train_timesteps (`int`, defaults to 1000): Included for config compatibility; not used to build the schedule. eta (`float`, defaults to 1.0): Stochasticity parameter. `eta=0.0` yields deterministic DDIM-like sampling; `eta=1.0` matches ComfyUI's default RF behavior. s_noise (`float`, defaults to 1.0): Global scaling factor for the stochastic noise term. FlowMatchEulerDiscreteSchedulerr ?num_train_timestepsetas_noisecCs"d|_d|_d|_d|_d|_dSN)num_inference_stepssigmas timesteps _step_index _begin_index)selfrrrrrr__init__Ns  z%LTXEulerAncestralRFScheduler.__init__returncC|jSr)r!r#rrr step_index\sz'LTXEulerAncestralRFScheduler.step_indexcCr&)z The index for the first timestep. It can be set from a pipeline with `set_begin_index` to support image-to-image like workflows that start denoising part-way through the schedule. r"r'rrr begin_index`sz(LTXEulerAncestralRFScheduler.begin_indexrr*cCs ||_dS)z Included for API compatibility; not strictly needed here but kept to allow pipelines that call `set_begin_index`. Nr))r#r*rrrset_begin_indexhs z,LTXEulerAncestralRFScheduler.set_begin_indexNtimestepschedule_timestepscCsx|dur|jdur td|j}t|tjr||j}||k}t|dkr*dnd}t|dkr6td|| S)a Map a (continuous) `timestep` value to an index into `self.timesteps`. This follows the convention used in other discrete schedulers: if the same timestep value appears multiple times in the schedule (which can happen when starting in the middle of the schedule), the *second* occurrence is used for the first `step` call so that no sigma is accidentally skipped. N8Timesteps have not been set. Call `set_timesteps` first.r rzaPassed `timestep` is not in `self.timesteps`. Make sure to use values from `scheduler.timesteps`.) r ValueError isinstancerTensortodevicenonzerolenitem)r#r,r-indicesposrrrindex_for_timestepos      z/LTXEulerAncestralRFScheduler.index_for_timestepcCsR|jdur td|jdur#t|tjr||jj}|||_ dS|j |_ dS)zO Initialize the internal step index based on a given timestep. Nr.) r r/r*r0rr1r2r3r9r!r")r#r,rrr_init_step_indexs    z-LTXEulerAncestralRFScheduler._init_step_indexrr3rr muc Ks|durB|durB|durtdddlm}||j}|j||d|dd|j|_|jj|d|_|j j|d|_ d|_ d|_ dS|durH|}t |t rVtj|tjd} nt |tjrd|jtjd} n tdt|d | jdkr~td t| jd | d d krtd | d |dur| |} | |_tt|jdd} | | |_ |dur|t|dkrtd|t|dt|d|_d|_ d|_ dS)a Set the sigma / timestep schedule for sampling. When `sigmas` or `timesteps` are provided explicitly, they are used as the RF sigma schedule (ComfyUI-style) and are expected to include the terminal 0.0. When both are `None`, the scheduler reuses the [`FlowMatchEulerDiscreteScheduler`] logic to generate sigmas from `num_inference_steps` and the stored config (including any resolution-dependent shifting, Karras/beta schedules, etc.). Args: num_inference_steps (`int`, *optional*): Number of denoising steps. If provided together with explicit `sigmas`/`timesteps`, they are expected to be consistent and are otherwise ignored with a warning. device (`str` or `torch.device`, *optional*): Device to move the internal tensors to. sigmas (`list[float]` or `torch.Tensor`, *optional*): Explicit sigma schedule, e.g. `[1.0, 0.99, ..., 0.0]`. timesteps (`list[float]` or `torch.Tensor`, *optional*): Optional alias for `sigmas`. If `sigmas` is None and `timesteps` is provided, timesteps are treated as sigmas. mu (`float`, *optional*): Optional shift parameter used when delegating to [`FlowMatchEulerDiscreteScheduler.set_timesteps`] and `config.use_dynamic_shifting` is `True`. NzzLTXEulerAncestralRFScheduler.set_timesteps requires either explicit `sigmas`/`timesteps` or a `num_inference_steps` value.r )r)rr3rr;r )r3)dtypez-`sigmas` must be a list or torch.Tensor, got .z(`sigmas` must be a 1D tensor, got shape gư>zThe last sigma in the schedule is not zero (%.6f). For best compatibility with ComfyUI's RF sampler, the terminal sigma should be 0.0.rrz{Provided `num_inference_steps=%d` does not match `len(sigmas)-1=%d`. Overriding `num_inference_steps` with `len(sigmas)-1`.)r/$scheduling_flow_match_euler_discreter from_configconfig set_timestepsrrr2r r!r"r0listrtensorfloat32r1 TypeErrortypendimtupleshapeabsr6loggerwarningfloatgetattrr5) r#rr3rr r;kwargsrbase_scheduler sigmas_tensor num_trainrrrrBs`!          z*LTXEulerAncestralRFScheduler.set_timestepssigmasamplecCs2|j|jkr|jg|jdR}|j|jks|S)zN Helper to broadcast a scalar sigma to the shape of `sample`. r )rHviewrJ)r#rTrUrrr_sigma_broadcast s  z-LTXEulerAncestralRFScheduler._sigma_broadcastT model_output generator return_dictcCs&t|ttjtjfrtd|jdus|jdurtd|jdur&| ||j}|t |jdkr5|}n| tj }| tj } |j|} |j|d} | | d|} | | d|} || | }| dkrp|}nt|jj}t|jj}d| | d|}| |}d| }d|}| |d|}| |d|}| |d|}|| }||d||}|dkr|dkr| d|d|d|dd jdd }t|j||j|jd }||d ||||}| |j}t|jdt |jd|_|s|fSt|d S) aQ Perform a single Euler-Ancestral RF update step. Args: model_output (`torch.FloatTensor`): Raw model output at the current step. Interpreted under the CONST parametrization as `v_t`, with denoised state reconstructed as `x0 = x_t - sigma_t * v_t`. timestep (`float` or `torch.Tensor`): The current sigma value (must match one entry in `self.timesteps`). sample (`torch.FloatTensor`): Current latent sample `x_t`. generator (`torch.Generator`, *optional*): Optional generator for reproducible noise. return_dict (`bool`): If `True`, return a `LTXEulerAncestralRFSchedulerOutput`; otherwise return a tuple where the first element is the updated sample. zPassing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to `LTXEulerAncestralRFScheduler.step()` is not supported. Make sure to pass one of the `scheduler.timesteps` values as `timestep`.NzGScheduler has not been initialized. Call `set_timesteps` before `step`.r g:0yE>rgrg-q=)min)rYr3r<)r )r0intr IntTensor LongTensorr/rr r!r:r5r2rErWrVrKr6rNrArrclampsqrtrrJr3r<r[r )r#rXr,rUrYrZir sample_fmodel_output_frT sigma_nextsigma_b sigma_next_bdenoisedxrrdownstep_ratio sigma_down alpha_ip1 alpha_down sigma_down_b alpha_ip1_b alpha_down_b sigma_ratio renoise_coeffnoiserrrsteps\        "  z!LTXEulerAncestralRFScheduler.stepcCstt|jddS)Nrr)r\rOrAr'rrr__len__~sz$LTXEulerAncestralRFScheduler.__len__)rrr)rr)NNNNN)NT)r rrr _compatiblesorderrr\rNr$propertyr(r*r+rr1r9r:strr3rCrBrWr Generatorboolr rIrsrtrrrrr3s|   !  o   ir)r dataclassesrrconfiguration_utilsrrutilsrrutils.torch_utilsrscheduling_utilsr get_loggerr rLr rrrrrs