[LTX-2] Add conditioning_frame_rate to decouple the positional time base from the playback frame rate

src/diffusers/modular_pipelines/ltx2/before_denoise.py

@@ -752,6 +752,15 @@ def inputs(self) -> list[InputParam]:
             InputParam(
                 "frame_rate", type_hint=float, default=24.0, description="Frames per second of the generated video."
             ),
+            InputParam(
+                "conditioning_frame_rate",
+                type_hint=float,
+                description=(
+                    "Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults "
+                    "to `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture "
+                    "rate differs from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`."
+                ),
+            ),
             InputParam("audio_num_frames", type_hint=int, required=True),
             InputParam.template("num_images_per_prompt", name="num_videos_per_prompt"),
             InputParam(
@@ -790,7 +799,12 @@ def __call__(self, components, state: PipelineState) -> tuple[LTX2ModularPipelin
         latent_num_frames = (block_state.num_frames - 1) // components.vae_temporal_compression_ratio + 1
 
         block_state.video_coords = components.transformer.rope.prepare_video_coords(
-            batch_size, latent_num_frames, latent_height, latent_width, device, fps=block_state.frame_rate
+            batch_size,
+            latent_num_frames,
+            latent_height,
+            latent_width,
+            device,
+            fps=block_state.conditioning_frame_rate or block_state.frame_rate,
         )
         block_state.audio_coords = components.transformer.audio_rope.prepare_audio_coords(
             batch_size, block_state.audio_num_frames, device
@@ -863,6 +877,15 @@ def inputs(self) -> list[InputParam]:
             InputParam(
                 "frame_rate", type_hint=float, default=24.0, description="Frames per second of the generated video."
             ),
+            InputParam(
+                "conditioning_frame_rate",
+                type_hint=float,
+                description=(
+                    "Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults "
+                    "to `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture "
+                    "rate differs from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`."
+                ),
+            ),
             InputParam(
                 "noise_scale",
                 type_hint=float,
@@ -1022,7 +1045,7 @@ def __call__(self, components, state: PipelineState) -> tuple[LTX2ModularPipelin
                 keyframe_latent_width=kf_latent_width,
                 pixel_frame_idx=(latent_idx - 1) * frame_scale_factor + 1,
                 num_pixel_frames=num_pixel_frames,
-                fps=block_state.frame_rate,
+                fps=block_state.conditioning_frame_rate or block_state.frame_rate,
                 patch_size=spatial_patch,
                 patch_size_t=temporal_patch,
                 scale_factors=scale_factors,
@@ -1159,6 +1182,15 @@ def inputs(self) -> list[InputParam]:
             InputParam(
                 "frame_rate", type_hint=float, default=24.0, description="Frames per second of the generated video."
             ),
+            InputParam(
+                "conditioning_frame_rate",
+                type_hint=float,
+                description=(
+                    "Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults "
+                    "to `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture "
+                    "rate differs from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`."
+                ),
+            ),
             InputParam(
                 "noise_scale",
                 type_hint=float,
@@ -1317,7 +1349,7 @@ def __call__(self, components, state: PipelineState) -> tuple[LTX2ModularPipelin
                 keyframe_latent_width=kf_latent_width,
                 pixel_frame_idx=(latent_idx - 1) * frame_scale_factor + 1,
                 num_pixel_frames=num_pixel_frames,
-                fps=block_state.frame_rate,
+                fps=block_state.conditioning_frame_rate or block_state.frame_rate,
                 patch_size=spatial_patch,
                 patch_size_t=temporal_patch,
                 scale_factors=scale_factors,
@@ -1706,6 +1738,15 @@ def inputs(self) -> list[InputParam]:
             InputParam(
                 "frame_rate", type_hint=float, default=24.0, description="Frames per second of the generated video."
             ),
+            InputParam(
+                "conditioning_frame_rate",
+                type_hint=float,
+                description=(
+                    "Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults "
+                    "to `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture "
+                    "rate differs from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`."
+                ),
+            ),
             InputParam("audio_num_frames", type_hint=int, required=True),
             InputParam(
                 "appended_coords",
@@ -1750,7 +1791,12 @@ def __call__(self, components, state: PipelineState) -> tuple[LTX2ModularPipelin
         latent_num_frames = (block_state.num_frames - 1) // components.vae_temporal_compression_ratio + 1
 
         video_coords = components.transformer.rope.prepare_video_coords(
-            batch_size, latent_num_frames, latent_height, latent_width, device, fps=block_state.frame_rate
+            batch_size,
+            latent_num_frames,
+            latent_height,
+            latent_width,
+            device,
+            fps=block_state.conditioning_frame_rate or block_state.frame_rate,
         )
         block_state.video_coords = torch.cat([video_coords, block_state.appended_coords.to(video_coords.dtype)], dim=2)
         block_state.audio_coords = components.transformer.audio_rope.prepare_audio_coords(

src/diffusers/modular_pipelines/ltx2/denoise.py

@@ -319,6 +319,15 @@ def inputs(self) -> list[InputParam]:
             InputParam(
                 "frame_rate", type_hint=float, default=24.0, description="Frames per second of the generated video."
             ),
+            InputParam(
+                "conditioning_frame_rate",
+                type_hint=float,
+                description=(
+                    "Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults "
+                    "to `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture "
+                    "rate differs from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`."
+                ),
+            ),
             InputParam(
                 "use_cross_timestep",
                 type_hint=bool,
@@ -361,7 +370,7 @@ def __call__(
             num_frames=latent_num_frames,
             height=latent_height,
             width=latent_width,
-            fps=block_state.frame_rate,
+            fps=block_state.conditioning_frame_rate or block_state.frame_rate,
             use_cross_timestep=block_state.use_cross_timestep,
             attention_kwargs=block_state.attention_kwargs,
             perturbation_mask=None,

src/diffusers/modular_pipelines/ltx2/encoders.py

@@ -1197,6 +1197,15 @@ def inputs(self) -> list[InputParam]:
             InputParam(
                 "frame_rate", type_hint=float, default=24.0, description="Frames per second of the generated video."
             ),
+            InputParam(
+                "conditioning_frame_rate",
+                type_hint=float,
+                description=(
+                    "Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults "
+                    "to `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture "
+                    "rate differs from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`."
+                ),
+            ),
             InputParam.template("generator"),
         ]
 
@@ -1285,7 +1294,7 @@ def __call__(self, components, state: PipelineState) -> tuple[LTX2ModularPipelin
                 height=ref_latent_height,
                 width=ref_latent_width,
                 device=device,
-                fps=block_state.frame_rate,
+                fps=block_state.conditioning_frame_rate or block_state.frame_rate,
             )
             if downscale_factor != 1:
                 ref_coords[:, 1, :, :] = ref_coords[:, 1, :, :] * downscale_factor

src/diffusers/modular_pipelines/ltx2/modular_blocks_ltx2.py

@@ -322,6 +322,10 @@ class LTX2CoreDenoiseStep(SequentialPipelineBlocks):
               Frames per second of the generated video.
           audio_latents (`Tensor`, *optional*):
               Optional pre-encoded audio latents; random noise is used when not provided.
+          conditioning_frame_rate (`float`, *optional*):
+              Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults to
+              `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture rate differs
+              from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`.
           dtype (`dtype`):
               The dtype the model inputs are cast to.
           **denoiser_input_fields (`None`, *optional*):
@@ -424,6 +428,10 @@ class LTX2Image2VideoCoreDenoiseStep(SequentialPipelineBlocks):
               Frames per second of the generated video.
           audio_latents (`Tensor`, *optional*):
               Optional pre-encoded audio latents; random noise is used when not provided.
+          conditioning_frame_rate (`float`, *optional*):
+              Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults to
+              `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture rate differs
+              from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`.
           dtype (`dtype`):
               The dtype the model inputs are cast to.
           **denoiser_input_fields (`None`, *optional*):
@@ -520,6 +528,10 @@ class LTX2ConditionCoreDenoiseStep(SequentialPipelineBlocks):
               `LTX2AutoDurationStep`).
           frame_rate (`float`, *optional*, defaults to 24.0):
               Frames per second of the generated video.
+          conditioning_frame_rate (`float`, *optional*):
+              Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults to
+              `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture rate differs
+              from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`.
           noise_scale (`float`, *optional*):
               Initial noise level for the un-conditioned tokens. `None` (default) resolves to `sigmas[0]` when custom
               `sigmas` are supplied, else 1.0.
@@ -676,6 +688,10 @@ class LTX2AutoReferenceEncoderStep(ConditionalPipelineBlocks):
               `LTX2AutoDurationStep`).
           frame_rate (`float`, *optional*, defaults to 24.0):
               Frames per second of the generated video.
+          conditioning_frame_rate (`float`, *optional*):
+              Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults to
+              `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture rate differs
+              from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`.
           generator (`Generator`, *optional*):
               Torch generator for deterministic generation.
 
@@ -831,6 +847,10 @@ class LTX2InContextCoreDenoiseStep(SequentialPipelineBlocks):
               `LTX2AutoDurationStep`).
           frame_rate (`float`, *optional*, defaults to 24.0):
               Frames per second of the generated video.
+          conditioning_frame_rate (`float`, *optional*):
+              Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults to
+              `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture rate differs
+              from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`.
           noise_scale (`float`, *optional*):
               Initial noise level for the un-conditioned tokens. `None` (default) resolves to `sigmas[0]` when custom
               `sigmas` are supplied, else 1.0.
@@ -962,6 +982,10 @@ class LTX2AutoCoreDenoiseStep(ConditionalPipelineBlocks):
               `LTX2AutoDurationStep`).
           frame_rate (`float`, *optional*, defaults to 24.0):
               Frames per second of the generated video.
+          conditioning_frame_rate (`float`, *optional*):
+              Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults to
+              `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture rate differs
+              from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`.
           noise_scale (`float`, *optional*):
               Initial noise level for the un-conditioned tokens. `None` (default) resolves to `sigmas[0]` when custom
               `sigmas` are supplied, else 1.0.
@@ -1297,6 +1321,10 @@ class LTX2Blocks(SequentialPipelineBlocks):
               which keeps the provided latents.
           audio_latents (`Tensor`, *optional*):
               Optional pre-encoded audio latents; random noise is used when not provided.
+          conditioning_frame_rate (`float`, *optional*):
+              Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults to
+              `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture rate differs
+              from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`.
           **denoiser_input_fields (`None`, *optional*):
               conditional model inputs for the denoiser: e.g. prompt_embeds, negative_prompt_embeds, etc.
           use_cross_timestep (`bool`, *optional*, defaults to True):
@@ -1412,6 +1440,10 @@ class LTX2ImageToVideoBlocks(SequentialPipelineBlocks):
               VAE-encoded reference-image latents used for image-to-video conditioning.
           audio_latents (`Tensor`, *optional*):
               Optional pre-encoded audio latents; random noise is used when not provided.
+          conditioning_frame_rate (`float`, *optional*):
+              Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults to
+              `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture rate differs
+              from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`.
           **denoiser_input_fields (`None`, *optional*):
               conditional model inputs for the denoiser: e.g. prompt_embeds, negative_prompt_embeds, etc.
           use_cross_timestep (`bool`, *optional*, defaults to True):
@@ -1511,6 +1543,10 @@ class LTX2ConditionBlocks(SequentialPipelineBlocks):
               The number of images to generate per prompt.
           latents (`Tensor`, *optional*):
               Pre-generated noisy latents for image generation.
+          conditioning_frame_rate (`float`, *optional*):
+              Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults to
+              `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture rate differs
+              from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`.
           noise_scale (`float`, *optional*):
               Initial noise level for the un-conditioned tokens. `None` (default) resolves to `sigmas[0]` when custom
               `sigmas` are supplied, else 1.0.
@@ -1633,6 +1669,10 @@ class LTX2InContextBlocks(SequentialPipelineBlocks):
               `conditioning_attention_strength`.
           frame_rate (`float`, *optional*, defaults to 24.0):
               Frames per second of the generated video.
+          conditioning_frame_rate (`float`, *optional*):
+              Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults to
+              `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture rate differs
+              from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`.
           num_videos_per_prompt (`int`, *optional*, defaults to 1):
               The number of images to generate per prompt.
           reference_latents (`Tensor`, *optional*):
@@ -1794,6 +1834,10 @@ class LTX2AutoBlocks(SequentialPipelineBlocks):
               Optional pixel-space mask of shape (1, 1, F, H, W) with values in [0, 1] giving spatially varying
               attention strength. Downsampled to the reference's latent grid and multiplied by
               `conditioning_attention_strength`.
+          conditioning_frame_rate (`float`, *optional*):
+              Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults to
+              `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture rate differs
+              from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`.
           num_videos_per_prompt (`int`, *optional*, defaults to 1):
               The number of images to generate per prompt.
           condition_latents (`list`, *optional*):

src/diffusers/modular_pipelines/ltx2/modular_blocks_ltx25.py

@@ -290,6 +290,10 @@ class LTX25AutoBlocks(SequentialPipelineBlocks):
               Optional pixel-space mask of shape (1, 1, F, H, W) with values in [0, 1] giving spatially varying
               attention strength. Downsampled to the reference's latent grid and multiplied by
               `conditioning_attention_strength`.
+          conditioning_frame_rate (`float`, *optional*):
+              Frame rate the model is conditioned on (the time axis of the positional embeddings). Defaults to
+              `frame_rate`. Set it apart from `frame_rate` for adapters trained on footage whose capture rate differs
+              from its playback rate, e.g. slow-motion LoRAs: `frame_rate / speed`.
           num_videos_per_prompt (`int`, *optional*, defaults to 1):
               The number of images to generate per prompt.
           condition_latents (`list`, *optional*):