@@ -77,6 +77,26 @@ pipe = QwenImage21Pipeline.from_pretrained("Qwen/Qwen-Image-2.1", transformer=tr
 )
 ```
 
+## Sampling sigmas
+
+Model authors can configure a default sampling grid with `sample_sigmas` in the pipeline config. When you load a
+released checkpoint, its default grid and scheduler settings are restored automatically.
+
+To experiment with a different grid at runtime, pass `sigmas` to the pipeline call:
+
+```python
+# Use the checkpoint's default sampling grid.
+image = pipe(prompt).images[0]
+
+# Override the default grid for this call.
+image = pipe(prompt, sigmas=[1.0, 0.8, 0.5, 0.2]).images[0]
+```
+
+The custom grid above illustrates the API; generation quality depends on the checkpoint and grid. Sigma lists exclude
+the terminal sigma, which the scheduler appends. Explicit `sigmas` override the configured `sample_sigmas`, and either
+list determines the number of steps instead of `num_inference_steps`. If neither is provided, the pipeline uses
+`num_inference_steps` to generate the schedule. The scheduler applies its configured processing to either grid.
+
 ## QwenImage21Pipeline
 
 [[autodoc]] QwenImage21Pipeline
 
@@ -174,6 +174,10 @@ class QwenImage21Pipeline(DiffusionPipeline, QwenImageLoraLoaderMixin):
             Processor that builds the chat template and tokenizes prompt and condition images.
         transformer ([`QwenImage21Transformer2DModel`]):
             The single-stream block-causal transformer that denoises the latents.
+        sample_sigmas (`list[float]`, *optional*):
+            Default sampling sigmas configured by the model author, excluding the terminal sigma. Their length
+            determines the number of denoising steps. To experiment with a different grid at runtime, pass `sigmas` to
+            `__call__`.
     """
 
     model_cpu_offload_seq = "text_encoder->transformer->vae"
@@ -186,6 +190,7 @@ def __init__(
         text_encoder: Qwen3VLForConditionalGeneration,
         processor: Qwen3VLProcessor,
         transformer: QwenImage21Transformer2DModel,
+        sample_sigmas: list[float] | None = None,
     ):
         super().__init__()
 
@@ -196,6 +201,7 @@ def __init__(
             transformer=transformer,
             scheduler=scheduler,
         )
+        self.register_to_config(sample_sigmas=sample_sigmas)
         # The VAE compresses 16x spatially and the transformer consumes latents unpatched, so one token covers a 16x16
         # pixel tile.
         self.vae_scale_factor = 16
@@ -547,9 +553,12 @@ def __call__(
             width (`int`, *optional*):
                 Width in pixels of the generated image. Derived from the condition image's aspect ratio if omitted.
             num_inference_steps (`int`, *optional*, defaults to 40):
-                Number of denoising steps.
+                Number of denoising steps. Ignored when `sigmas` or the pipeline's configured `sample_sigmas` is used;
+                the length of that schedule determines the number of steps.
             sigmas (`list[float]`, *optional*):
-                Custom sigmas for the denoising schedule.
+                Sampling sigmas to try for this call, excluding the terminal sigma. Overrides the pipeline's configured
+                `sample_sigmas` and determines the number of denoising steps. The scheduler applies its configured
+                processing to these values.
             num_images_per_prompt (`int`, *optional*, defaults to 1):
                 Number of images generated per prompt.
             generator (`torch.Generator` or `list[torch.Generator]`, *optional*):
@@ -720,7 +729,10 @@ def __call__(
         ] * batch_size
 
         # 4. Prepare timesteps
-        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
+        if sigmas is None:
+            sigmas = self.config.sample_sigmas
+            if sigmas is None:
+                sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
         mu = calculate_shift(
             latents.shape[1],
             self.scheduler.config.get("base_image_seq_len", 256),
 
@@ -182,6 +182,74 @@ def get_dummy_inputs(self):
 
 
 class TestQwenImage21Pipeline(QwenImage21PipelineTesterConfig, PipelineTesterMixin):
+    @pytest.mark.parametrize("num_inference_steps", [None, 2])
+    def test_sample_sigmas_from_pipeline_config(self, num_inference_steps):
+        pipe = self.get_pipeline()
+        sample_sigmas = [1.0, 0.978453, 0.954180, 0.926626, 0.895080, 0.845148, 0.704534, 0.414568]
+        pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(
+            pipe.scheduler.config,
+            use_dynamic_shifting=False,
+            shift=1.0,
+            shift_terminal=None,
+        )
+        pipe = self.pipeline_class(**pipe.components, sample_sigmas=sample_sigmas)
+        inputs = self.get_dummy_inputs()
+        inputs["output_type"] = "latent"
+        if num_inference_steps is None:
+            inputs.pop("num_inference_steps")
+        else:
+            inputs["num_inference_steps"] = num_inference_steps
+        seen_timesteps = []
+
+        def callback(pipe, step, timestep, callback_kwargs):
+            seen_timesteps.append(timestep.clone())
+            return callback_kwargs
+
+        pipe(**inputs, callback_on_step_end=callback)
+        expected_sigmas = torch.tensor(sample_sigmas + [0.0], device=pipe.scheduler.sigmas.device)
+        expected_timesteps = expected_sigmas[:-1] * pipe.scheduler.config.num_train_timesteps
+        assert torch.equal(pipe.scheduler.sigmas, expected_sigmas)
+        assert torch.equal(pipe.scheduler.timesteps, expected_timesteps)
+        assert torch.equal(torch.stack(seen_timesteps), expected_timesteps)
+        assert pipe.scheduler.num_inference_steps == len(sample_sigmas)
+
+    def test_explicit_sigmas_override_pipeline_config(self):
+        pipe = self.get_pipeline()
+        pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(
+            pipe.scheduler.config, use_dynamic_shifting=True, shift_terminal=0.02
+        )
+        inputs = self.get_dummy_inputs()
+        inputs.update(sigmas=[1.0, 0.6, 0.2], output_type="latent")
+        expected = pipe(**inputs).images
+        expected_sigmas = pipe.scheduler.sigmas.clone()
+        pipe.register_to_config(sample_sigmas=[1.0, 0.8])
+        inputs["generator"] = self.get_generator(0)
+        actual = pipe(**inputs).images
+        assert torch.equal(pipe.scheduler.sigmas, expected_sigmas)
+        assert torch.equal(actual, expected)
+
+    def test_sample_sigmas_save_load(self, tmp_path):
+        pipe = self.get_pipeline()
+        sample_sigmas = [1.0, 0.8, 0.4]
+        pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(
+            pipe.scheduler.config, use_dynamic_shifting=False, shift=1.0, shift_terminal=None
+        )
+        pipe = self.pipeline_class(**pipe.components, sample_sigmas=sample_sigmas)
+        inputs = self.get_dummy_inputs()
+        inputs["output_type"] = "latent"
+        expected = pipe(**inputs).images
+        pipe.save_pretrained(tmp_path)
+        restored = self.pipeline_class.from_pretrained(tmp_path)
+        assert restored.config.sample_sigmas == sample_sigmas
+        assert "sample_sigmas" not in restored.scheduler.config
+        inputs["generator"] = self.get_generator(0)
+        actual = restored(**inputs).images
+        assert torch.equal(actual, expected)
+        overridden = self.pipeline_class.from_pretrained(tmp_path, sample_sigmas=[1.0, 0.5])
+        inputs["generator"] = self.get_generator(0)
+        overridden(**inputs)
+        assert torch.equal(overridden.scheduler.sigmas.cpu(), torch.tensor([1.0, 0.5, 0.0]))
+
     @PROCESSOR_REQUIRED_AT_INIT
     def test_encode_prompt_works_in_isolation(self):
         super().test_encode_prompt_works_in_isolation()