From 05f0e34547366b1e1909cc6aab224fac719bcb2c Mon Sep 17 00:00:00 2001
From: christopher5106 <christopher.bourez@gmail.com>
Date: Tue, 6 Oct 2026 17:44:36 +0200
Subject: [PATCH] [LTX-2] Add `conditioning_frame_rate` to decouple the
 positional time base from the playback frame rate

LTX-2 places video tokens on a time axis in seconds (`pixel_frame / fps`
in the RoPE coordinates), derived from `frame_rate`. Adapters such as the
LTX-2.5 Slow-Motion-Control LoRA are conditioned on a different rate
(`fps / speed`) than the video plays at. Add an optional
`conditioning_frame_rate` to the LTX-2 classic pipelines and modular
blocks: it drives the video, keyframe and reference coordinates and the
transformer's `fps`, while the audio length, the duration head and the
returned video keep following `frame_rate`. `None` resolves to
`frame_rate`, leaving current outputs unchanged.
---
 .../modular_pipelines/ltx2/before_denoise.py  | 54 +++++++++++++++++--
 .../modular_pipelines/ltx2/denoise.py         | 11 +++-
 .../modular_pipelines/ltx2/encoders.py        | 11 +++-
 .../ltx2/modular_blocks_ltx2.py               | 44 +++++++++++++++
 .../ltx2/modular_blocks_ltx25.py              |  4 ++
 src/diffusers/pipelines/ltx2/pipeline_ltx2.py | 21 ++++++--
 .../pipelines/ltx2/pipeline_ltx2_condition.py | 23 ++++++--
 .../pipelines/ltx2/pipeline_ltx2_ic_lora.py   | 23 ++++++--
 .../ltx2/pipeline_ltx2_image2video.py         | 21 ++++++--
 .../ltx2/test_modular_pipeline_ltx25.py       | 19 +++++++
 tests/pipelines/ltx2/test_ltx2.py             | 26 +++++++++
 11 files changed, 233 insertions(+), 24 deletions(-)

diff --git a/src/diffusers/modular_pipelines/ltx2/before_denoise.py b/src/diffusers/modular_pipelines/ltx2/before_denoise.py
index 1878c01305fb..5f2bb7e6fd8d 100644
--- a/src/diffusers/modular_pipelines/ltx2/before_denoise.py
+++ b/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(
diff --git a/src/diffusers/modular_pipelines/ltx2/denoise.py b/src/diffusers/modular_pipelines/ltx2/denoise.py
index 2fb4951377a1..4ebcb9585625 100644
--- a/src/diffusers/modular_pipelines/ltx2/denoise.py
+++ b/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,
diff --git a/src/diffusers/modular_pipelines/ltx2/encoders.py b/src/diffusers/modular_pipelines/ltx2/encoders.py
index ae3d29b5cfe3..47ff0943cc98 100644
--- a/src/diffusers/modular_pipelines/ltx2/encoders.py
+++ b/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
diff --git a/src/diffusers/modular_pipelines/ltx2/modular_blocks_ltx2.py b/src/diffusers/modular_pipelines/ltx2/modular_blocks_ltx2.py
index 86428328a5a6..adea71e92baf 100644
--- a/src/diffusers/modular_pipelines/ltx2/modular_blocks_ltx2.py
+++ b/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*):
diff --git a/src/diffusers/modular_pipelines/ltx2/modular_blocks_ltx25.py b/src/diffusers/modular_pipelines/ltx2/modular_blocks_ltx25.py
index 7c77aba94a74..919935e63942 100644
--- a/src/diffusers/modular_pipelines/ltx2/modular_blocks_ltx25.py
+++ b/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*):
diff --git a/src/diffusers/pipelines/ltx2/pipeline_ltx2.py b/src/diffusers/pipelines/ltx2/pipeline_ltx2.py
index 4e5ced0b4ec8..49a91f2efbcc 100644
--- a/src/diffusers/pipelines/ltx2/pipeline_ltx2.py
+++ b/src/diffusers/pipelines/ltx2/pipeline_ltx2.py
@@ -935,6 +935,7 @@ def __call__(
         min_seconds: float = 1.0,
         max_seconds: float = 20.0,
         frame_rate: float = 24.0,
+        conditioning_frame_rate: float | None = None,
         num_inference_steps: int = 30,
         sigmas: list[float] | None = None,
         timesteps: list[int] = None,
@@ -998,6 +999,11 @@ def __call__(
                 present. Ignored when `num_frames` is set explicitly. Must be strictly greater than `min_seconds`.
             frame_rate (`float`, *optional*, defaults to `24.0`):
                 The frames per second (FPS) of the generated video.
+            conditioning_frame_rate (`float`, *optional*):
+                The frame rate the model is conditioned on, i.e. 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. a slow-motion LoRA at `frame_rate / speed`. Audio duration and the
+                predicted number of frames keep following `frame_rate`.
             num_inference_steps (`int`, *optional*, defaults to 30):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                 expense of slower inference.
@@ -1150,6 +1156,8 @@ def __call__(
         audio_modality_scale = audio_modality_scale or modality_scale
         audio_guidance_rescale = audio_guidance_rescale or guidance_rescale
 
+        conditioning_frame_rate = conditioning_frame_rate if conditioning_frame_rate is not None else frame_rate
+
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(
             prompt=prompt,
@@ -1367,7 +1375,12 @@ def __call__(
         # 6. Prepare micro-conditions
         # Pre-compute video and audio positional ids as they will be the same at each step of the denoising loop
         video_coords = self.transformer.rope.prepare_video_coords(
-            latents.shape[0], latent_num_frames, latent_height, latent_width, latents.device, fps=frame_rate
+            latents.shape[0],
+            latent_num_frames,
+            latent_height,
+            latent_width,
+            latents.device,
+            fps=conditioning_frame_rate,
         )
         audio_coords = self.transformer.audio_rope.prepare_audio_coords(
             audio_latents.shape[0], audio_num_frames, audio_latents.device
@@ -1408,7 +1421,7 @@ def __call__(
                         num_frames=latent_num_frames,
                         height=latent_height,
                         width=latent_width,
-                        fps=frame_rate,
+                        fps=conditioning_frame_rate,
                         audio_num_frames=audio_num_frames,
                         video_coords=video_coords,
                         audio_coords=audio_coords,
@@ -1480,7 +1493,7 @@ def __call__(
                             num_frames=latent_num_frames,
                             height=latent_height,
                             width=latent_width,
-                            fps=frame_rate,
+                            fps=conditioning_frame_rate,
                             audio_num_frames=audio_num_frames,
                             video_coords=video_pos_ids,
                             audio_coords=audio_pos_ids,
@@ -1520,7 +1533,7 @@ def __call__(
                             num_frames=latent_num_frames,
                             height=latent_height,
                             width=latent_width,
-                            fps=frame_rate,
+                            fps=conditioning_frame_rate,
                             audio_num_frames=audio_num_frames,
                             video_coords=video_pos_ids,
                             audio_coords=audio_pos_ids,
diff --git a/src/diffusers/pipelines/ltx2/pipeline_ltx2_condition.py b/src/diffusers/pipelines/ltx2/pipeline_ltx2_condition.py
index e1bff845302c..f27f4b9d39c1 100644
--- a/src/diffusers/pipelines/ltx2/pipeline_ltx2_condition.py
+++ b/src/diffusers/pipelines/ltx2/pipeline_ltx2_condition.py
@@ -1355,6 +1355,7 @@ def __call__(
         min_seconds: float = 1.0,
         max_seconds: float = 20.0,
         frame_rate: float = 24.0,
+        conditioning_frame_rate: float | None = None,
         num_inference_steps: int = 30,
         sigmas: list[float] | None = None,
         timesteps: list[float] | None = None,
@@ -1420,6 +1421,11 @@ def __call__(
                 present. Ignored when `num_frames` is set explicitly. Must be strictly greater than `min_seconds`.
             frame_rate (`float`, *optional*, defaults to `24.0`):
                 The frames per second (FPS) of the generated video.
+            conditioning_frame_rate (`float`, *optional*):
+                The frame rate the model is conditioned on, i.e. 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. a slow-motion LoRA at `frame_rate / speed`. Audio duration and the
+                predicted number of frames keep following `frame_rate`.
             num_inference_steps (`int`, *optional*, defaults to 30):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                 expense of slower inference.
@@ -1574,6 +1580,8 @@ def __call__(
         audio_modality_scale = audio_modality_scale or modality_scale
         audio_guidance_rescale = audio_guidance_rescale or guidance_rescale
 
+        conditioning_frame_rate = conditioning_frame_rate if conditioning_frame_rate is not None else frame_rate
+
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(
             prompt=prompt,
@@ -1721,7 +1729,7 @@ def __call__(
             height=height,
             width=width,
             num_frames=num_frames,
-            frame_rate=frame_rate,
+            frame_rate=conditioning_frame_rate,
             noise_scale=noise_scale,
             dtype=torch.float32,
             device=device,
@@ -1792,7 +1800,12 @@ def __call__(
         # 6. Prepare micro-conditions
         # Pre-compute video and audio positional ids as they will be the same at each step of the denoising loop
         video_coords = self.transformer.rope.prepare_video_coords(
-            latents.shape[0], latent_num_frames, latent_height, latent_width, latents.device, fps=frame_rate
+            latents.shape[0],
+            latent_num_frames,
+            latent_height,
+            latent_width,
+            latents.device,
+            fps=conditioning_frame_rate,
         )
         audio_coords = self.transformer.audio_rope.prepare_audio_coords(
             audio_latents.shape[0], audio_num_frames, audio_latents.device
@@ -1840,7 +1853,7 @@ def __call__(
                         num_frames=latent_num_frames,
                         height=latent_height,
                         width=latent_width,
-                        fps=frame_rate,
+                        fps=conditioning_frame_rate,
                         audio_num_frames=audio_num_frames,
                         video_coords=video_coords,
                         audio_coords=audio_coords,
@@ -1916,7 +1929,7 @@ def __call__(
                             num_frames=latent_num_frames,
                             height=latent_height,
                             width=latent_width,
-                            fps=frame_rate,
+                            fps=conditioning_frame_rate,
                             audio_num_frames=audio_num_frames,
                             video_coords=video_pos_ids,
                             audio_coords=audio_pos_ids,
@@ -1958,7 +1971,7 @@ def __call__(
                             num_frames=latent_num_frames,
                             height=latent_height,
                             width=latent_width,
-                            fps=frame_rate,
+                            fps=conditioning_frame_rate,
                             audio_num_frames=audio_num_frames,
                             video_coords=video_pos_ids,
                             audio_coords=audio_pos_ids,
diff --git a/src/diffusers/pipelines/ltx2/pipeline_ltx2_ic_lora.py b/src/diffusers/pipelines/ltx2/pipeline_ltx2_ic_lora.py
index 2924a086c721..905f7d4d910a 100644
--- a/src/diffusers/pipelines/ltx2/pipeline_ltx2_ic_lora.py
+++ b/src/diffusers/pipelines/ltx2/pipeline_ltx2_ic_lora.py
@@ -1787,6 +1787,7 @@ def __call__(
         width: int = 768,
         num_frames: int = 121,
         frame_rate: float = 24.0,
+        conditioning_frame_rate: float | None = None,
         num_inference_steps: int = 30,
         sigmas: list[float] | None = None,
         timesteps: list[float] | None = None,
@@ -1864,6 +1865,11 @@ def __call__(
                 The number of video frames to generate. Must satisfy `(n - 1) % 8 == 0`.
             frame_rate (`float`, *optional*, defaults to `24.0`):
                 The frames per second (FPS) of the generated video.
+            conditioning_frame_rate (`float`, *optional*):
+                The frame rate the model is conditioned on, i.e. 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. a slow-motion LoRA at `frame_rate / speed`. Audio duration and the
+                predicted number of frames keep following `frame_rate`.
             num_inference_steps (`int`, *optional*, defaults to 30):
                 The number of denoising steps.
             sigmas (`List[float]`, *optional*):
@@ -1964,6 +1970,8 @@ def __call__(
         audio_modality_scale = audio_modality_scale or modality_scale
         audio_guidance_rescale = audio_guidance_rescale or guidance_rescale
 
+        conditioning_frame_rate = conditioning_frame_rate if conditioning_frame_rate is not None else frame_rate
+
         # 1. Check inputs
         self.check_inputs(
             prompt=prompt,
@@ -2101,7 +2109,7 @@ def __call__(
             height=height,
             width=width,
             num_frames=num_frames,
-            frame_rate=frame_rate,
+            frame_rate=conditioning_frame_rate,
             noise_scale=noise_scale,
             dtype=torch.float32,
             device=device,
@@ -2195,7 +2203,12 @@ def __call__(
 
         # 7. Prepare positional coordinates
         video_coords = self.transformer.rope.prepare_video_coords(
-            latents.shape[0], latent_num_frames, latent_height, latent_width, latents.device, fps=frame_rate
+            latents.shape[0],
+            latent_num_frames,
+            latent_height,
+            latent_width,
+            latents.device,
+            fps=conditioning_frame_rate,
         )
         if appended_coords is not None:
             video_coords = torch.cat([video_coords, appended_coords], dim=2)
@@ -2258,7 +2271,7 @@ def __call__(
                         num_frames=latent_num_frames,
                         height=latent_height,
                         width=latent_width,
-                        fps=frame_rate,
+                        fps=conditioning_frame_rate,
                         audio_num_frames=audio_num_frames,
                         video_coords=video_coords,
                         audio_coords=audio_coords,
@@ -2347,7 +2360,7 @@ def __call__(
                             num_frames=latent_num_frames,
                             height=latent_height,
                             width=latent_width,
-                            fps=frame_rate,
+                            fps=conditioning_frame_rate,
                             audio_num_frames=audio_num_frames,
                             video_coords=video_pos_ids,
                             audio_coords=audio_pos_ids,
@@ -2396,7 +2409,7 @@ def __call__(
                             num_frames=latent_num_frames,
                             height=latent_height,
                             width=latent_width,
-                            fps=frame_rate,
+                            fps=conditioning_frame_rate,
                             audio_num_frames=audio_num_frames,
                             video_coords=video_pos_ids,
                             audio_coords=audio_pos_ids,
diff --git a/src/diffusers/pipelines/ltx2/pipeline_ltx2_image2video.py b/src/diffusers/pipelines/ltx2/pipeline_ltx2_image2video.py
index 36b7effdfd6c..fde5ca960fb1 100644
--- a/src/diffusers/pipelines/ltx2/pipeline_ltx2_image2video.py
+++ b/src/diffusers/pipelines/ltx2/pipeline_ltx2_image2video.py
@@ -990,6 +990,7 @@ def __call__(
         min_seconds: float = 1.0,
         max_seconds: float = 20.0,
         frame_rate: float = 24.0,
+        conditioning_frame_rate: float | None = None,
         num_inference_steps: int = 30,
         sigmas: list[float] | None = None,
         timesteps: list[int] | None = None,
@@ -1056,6 +1057,11 @@ def __call__(
                 present. Ignored when `num_frames` is set explicitly. Must be strictly greater than `min_seconds`.
             frame_rate (`float`, *optional*, defaults to `24.0`):
                 The frames per second (FPS) of the generated video.
+            conditioning_frame_rate (`float`, *optional*):
+                The frame rate the model is conditioned on, i.e. 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. a slow-motion LoRA at `frame_rate / speed`. Audio duration and the
+                predicted number of frames keep following `frame_rate`.
             num_inference_steps (`int`, *optional*, defaults to 30):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                 expense of slower inference.
@@ -1212,6 +1218,8 @@ def __call__(
         audio_modality_scale = audio_modality_scale or modality_scale
         audio_guidance_rescale = audio_guidance_rescale or guidance_rescale
 
+        conditioning_frame_rate = conditioning_frame_rate if conditioning_frame_rate is not None else frame_rate
+
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(
             prompt=prompt,
@@ -1442,7 +1450,12 @@ def __call__(
         # 6. Prepare micro-conditions
         # Pre-compute video and audio positional ids as they will be the same at each step of the denoising loop
         video_coords = self.transformer.rope.prepare_video_coords(
-            latents.shape[0], latent_num_frames, latent_height, latent_width, latents.device, fps=frame_rate
+            latents.shape[0],
+            latent_num_frames,
+            latent_height,
+            latent_width,
+            latents.device,
+            fps=conditioning_frame_rate,
         )
         audio_coords = self.transformer.audio_rope.prepare_audio_coords(
             audio_latents.shape[0], audio_num_frames, audio_latents.device
@@ -1484,7 +1497,7 @@ def __call__(
                         num_frames=latent_num_frames,
                         height=latent_height,
                         width=latent_width,
-                        fps=frame_rate,
+                        fps=conditioning_frame_rate,
                         audio_num_frames=audio_num_frames,
                         video_coords=video_coords,
                         audio_coords=audio_coords,
@@ -1558,7 +1571,7 @@ def __call__(
                             num_frames=latent_num_frames,
                             height=latent_height,
                             width=latent_width,
-                            fps=frame_rate,
+                            fps=conditioning_frame_rate,
                             audio_num_frames=audio_num_frames,
                             video_coords=video_pos_ids,
                             audio_coords=audio_pos_ids,
@@ -1599,7 +1612,7 @@ def __call__(
                             num_frames=latent_num_frames,
                             height=latent_height,
                             width=latent_width,
-                            fps=frame_rate,
+                            fps=conditioning_frame_rate,
                             audio_num_frames=audio_num_frames,
                             video_coords=video_pos_ids,
                             audio_coords=audio_pos_ids,
diff --git a/tests/modular_pipelines/ltx2/test_modular_pipeline_ltx25.py b/tests/modular_pipelines/ltx2/test_modular_pipeline_ltx25.py
index 69080095c627..86afd81ffa49 100644
--- a/tests/modular_pipelines/ltx2/test_modular_pipeline_ltx25.py
+++ b/tests/modular_pipelines/ltx2/test_modular_pipeline_ltx25.py
@@ -22,6 +22,7 @@
 from diffusers.pipelines.ltx2.pipeline_ltx2_condition import LTX2VideoCondition
 from diffusers.pipelines.ltx2.pipeline_ltx2_ic_lora import LTX2ReferenceCondition
 
+from ...testing_utils import assert_tensors_close
 from ..testing_utils import (
     BaseModularPipelineTesterConfig,
     ModularLoadingTesterMixin,
@@ -182,6 +183,24 @@ def test_auto_duration_predicts_a_grid_valid_frame_count(self):
         assert (num_frames - 1) % pipe.vae_temporal_compression_ratio == 0
         assert 0 < num_frames <= round(2.0 * inputs["frame_rate"])
 
+    def test_conditioning_frame_rate_rescales_only_the_video_time_axis(self):
+        pipe = self.get_pipeline().to("cpu")
+        frame_rate = self.get_dummy_inputs()["frame_rate"]
+        outputs = ["videos", "audio", "video_coords", "audio_coords"]
+
+        default = pipe(**self.get_dummy_inputs(), output=outputs)
+        same = pipe(**self.get_dummy_inputs(), conditioning_frame_rate=frame_rate, output=outputs)
+        slow = pipe(**self.get_dummy_inputs(), conditioning_frame_rate=2 * frame_rate, output=outputs)
+
+        assert_tensors_close(same["videos"], default["videos"], atol=1e-6, rtol=0)
+        # A doubled conditioning rate halves the time axis of the video positions and nothing else.
+        assert_tensors_close(slow["video_coords"][:, 0], default["video_coords"][:, 0] / 2, atol=1e-6, rtol=1e-6)
+        assert torch.equal(slow["video_coords"][:, 1:], default["video_coords"][:, 1:])
+        assert torch.equal(slow["audio_coords"], default["audio_coords"])
+        assert slow["videos"].shape == default["videos"].shape
+        assert slow["audio"].shape == default["audio"].shape
+        assert not torch.allclose(slow["videos"], default["videos"])
+
 
 class TestLTX25Text2VideoModularPipelineLoading(LTX25Text2VideoModularPipelineTesterConfig, ModularLoadingTesterMixin):
     pass
diff --git a/tests/pipelines/ltx2/test_ltx2.py b/tests/pipelines/ltx2/test_ltx2.py
index 0ad9da2b46a3..0b7c0101bf92 100644
--- a/tests/pipelines/ltx2/test_ltx2.py
+++ b/tests/pipelines/ltx2/test_ltx2.py
@@ -244,6 +244,32 @@ def test_invalid_duration_bounds_raise(self):
         with pytest.raises(ValueError, match="min_seconds"):
             pipe(**inputs)
 
+    def test_conditioning_frame_rate_changes_the_video_only(self):
+        pipe = self.get_pipeline()
+        frame_rate = self.get_dummy_inputs()["frame_rate"]
+
+        default = pipe(**self.get_dummy_inputs())
+        same = pipe(**{**self.get_dummy_inputs(), "conditioning_frame_rate": frame_rate})
+        slow = pipe(**{**self.get_dummy_inputs(), "conditioning_frame_rate": 2 * frame_rate})
+
+        # Defaulting to `frame_rate` leaves the pipeline unchanged.
+        assert_tensors_close(same.frames, default.frames, atol=1e-6, rtol=0)
+        assert_tensors_close(same.audio, default.audio, atol=1e-6, rtol=0)
+        # A different conditioning rate changes the video, not the output shapes or the audio length.
+        assert slow.frames.shape == default.frames.shape
+        assert slow.audio.shape == default.audio.shape
+        assert not torch.allclose(slow.frames, default.frames)
+
+    def test_duration_head_follows_frame_rate_not_conditioning_frame_rate(self):
+        pipe = self.get_pipeline_with_duration_head()
+        inputs = self.get_dummy_inputs()
+        inputs.pop("num_frames")
+        inputs.update(min_seconds=1.0, max_seconds=2.0)
+
+        default = pipe(**inputs).frames
+        slow = pipe(**{**inputs, "conditioning_frame_rate": 5 * inputs["frame_rate"]}).frames
+        assert slow.shape[1] == default.shape[1]
+
 
 class TestLTX2PipelineMemory(LTX2PipelineTesterConfig, LTX2MemoryTesterMixin):
     """Memory optimization tests (CPU offload, group offload, layerwise casting) for the LTX2 pipeline."""