diff --git a/modules/api/video.py b/modules/api/video.py
index 37cdf3be0..2b34357d2 100644
--- a/modules/api/video.py
+++ b/modules/api/video.py
@@ -27,10 +27,10 @@ class ReqVideo(BaseModel):
     seed: int = Field(default=-1, title="Seed", description="Generation seed; -1 for random")
     guidance_scale: float = Field(default=-1.0, title="Guidance scale", description="CFG scale; -1 keeps the model default")
     guidance_true: float = Field(default=-1.0, title="True guidance", description="True CFG scale; -1 keeps the model default")
-    init_image: str | None = Field(default=None, title="Init image", description="Base64, data URI, or upload reference for the first-frame image")
+    init_image: str | None = Field(default=None, title="Init image", description="Base64 or data URI for the first-frame image; an upload reference resolves only where an extension provides the upload store")
     init_strength: float = Field(default=0.8, ge=0.0, le=1.0, title="Init strength", description="Denoising strength for the init image")
-    last_image: str | None = Field(default=None, title="Last image", description="Base64, data URI, or upload reference for the last-frame image")
-    references: list[str] = Field(default=[], title="References", description="Reference images for a reference workflow, in the order the model reads them; base64, data URIs, or upload references. At most 9, each within a 1:4 to 4:1 aspect ratio. Rejected on models that do not condition on references")
+    last_image: str | None = Field(default=None, title="Last image", description="Base64 or data URI for the last-frame image; an upload reference resolves only where an extension provides the upload store")
+    references: list[str] = Field(default=[], title="References", description="Reference images for a reference workflow, in the order the model reads them; base64 or data URIs, or upload references where an extension provides the upload store. Images only: the video core also conditions on video and audio references, which this endpoint cannot carry. At most 9, each within a 1:4 to 4:1 aspect ratio. Rejected on models that do not condition on references")
     vae_type: str = Field(default="Default", title="VAE type", description="Decode variant: Default, Tiny, Remote, or Upscale")
     vae_tile_frames: int = Field(default=16, ge=1, le=64, title="VAE tile frames", description="Frames per VAE decode tile")
     audio: bool = Field(default=True, title="Audio", description="Generate audio on models that support it")
@@ -142,11 +142,14 @@ class APIVideo:
         `send_thumbnail`. Artifacts above the base64 size cap return `video` empty with
         `video_path` set; fetch those via `GET /sdapi/v1/video/file`.
 
-        `init_image` and `last_image` accept base64 data, data URIs, or upload references.
+        `init_image` and `last_image` accept base64 data or data URIs. An `upload:` reference
+        resolves only where an extension registers an upload store; without one it is rejected.
         Models whose workflow is `ref2va` condition on `references` instead: an ordered list of
         images the prompt addresses as `<Picture 1>`, `<Picture 2>` and so on, following list
         order. A single reference may also be passed as `init_image`. Reference images do not
-        set the output canvas, and `last_image` is ignored.
+        set the output canvas, and `last_image` is ignored. The workflow also conditions on video
+        and audio references, addressed as `<Video i>` and `<Audio i>`, but they decode from files
+        rather than from the wire, so this endpoint carries images alone.
 
         Progress is reported on `GET /sdapi/v1/progress`; `POST /sdapi/v1/interrupt` cancels.
         Switching checkpoints via `override_settings` is not supported here; use
diff --git a/modules/processing_args.py b/modules/processing_args.py
index f0e27a8ed..94a66153f 100644
--- a/modules/processing_args.py
+++ b/modules/processing_args.py
@@ -488,6 +488,8 @@ def set_pipeline_args(p, model, prompts:list, negative_prompts:list, prompts_2:l
             clean[k] = v.shape
         elif isinstance(v, list) and len(v) > 0 and (isinstance(v[0], torch.Tensor) or isinstance(v[0], np.ndarray)):
             clean[k] = [x.shape for x in v]
+        elif isinstance(v, list) and len(v) > 0 and hasattr(v[0], 'kind'): # media references carry decoded frames and waveforms
+            clean[k] = [getattr(x, 'kind', type(x).__name__) for x in v]
         elif not debug_enabled and k.endswith('_embeds'):
             del clean[k]
             clean['prompt'] = 'embeds'
diff --git a/modules/video_models/video_run.py b/modules/video_models/video_run.py
index 861b743e4..cd52c950e 100644
--- a/modules/video_models/video_run.py
+++ b/modules/video_models/video_run.py
@@ -9,7 +9,7 @@ from modules.paths import resolve_output_path
 
 
 debug = log.trace if os.environ.get('SD_VIDEO_DEBUG', None) is not None else lambda *args, **kwargs: None
-MAX_IMAGE_REFERENCES = 9 # mirrors the reference setup block's own limit; reading it off the pipe would deep-copy the block tree per access
+REFERENCE_WORKFLOWS = ('ref2va',) # workflows that condition on references; the resolver and the limits it enforces live with the architecture
 
 
 class VideoError(Exception):
@@ -62,29 +62,27 @@ def resolve_model(engine: str | None, model: str | None) -> tuple[models_def.Mod
     return selected, False
 
 
+def reference_caps(workflow: str | None):
+    """Reference limits of a workflow, None when it conditions on none. The seam a client reads
+    instead of mirroring the numbers."""
+    if workflow not in REFERENCE_WORKFLOWS:
+        return None
+    from modules.minimax import minimax_references
+    return minimax_references.get_reference_caps(workflow)
+
+
 def validate_references(selected: models_def.Model, references: list | None, init_image) -> list | None:
-    """Return the ordered reference images for a reference workflow, None for every other model.
+    """Return the ordered references a reference workflow conditions on, None for every other model.
     Reference conditioning is exclusive to ref2va: its partition holds no keyframe transformer, and
     a mismatched request would only fail once the pipeline reached a component it never loaded.
     Checks run before the model load so a rejected request costs nothing."""
     workflow = getattr(selected, 'workflow', None)
-    if workflow != 'ref2va':
+    if workflow not in REFERENCE_WORKFLOWS:
         if references:
-            raise VideoError(f'reference images require a ref2va model: model="{selected.name}" workflow={workflow}', 400)
+            raise VideoError(f'reference media requires a reference workflow: model="{selected.name}" workflow={workflow} supported={list(REFERENCE_WORKFLOWS)}', 400)
         return None
-    refs = list(references) if references else ([init_image] if init_image is not None else [])
-    if len(refs) == 0:
-        raise VideoError('No reference image provided. The ref2va workflow conditions on reference images, so at least one is required.', 400)
-    if len(refs) > MAX_IMAGE_REFERENCES:
-        raise VideoError(f'too many reference images: count={len(refs)} max={MAX_IMAGE_REFERENCES}', 400)
-    for index, image in enumerate(refs):
-        size = getattr(image, 'size', None)
-        if size is None or len(size) != 2:
-            raise VideoError(f'reference {index + 1} is not an image: type={type(image).__name__}', 400)
-        width, height = size
-        if width > 4 * height or height > 4 * width: # the same bound the pipeline enforces, raised here where it is free
-            raise VideoError(f'reference {index + 1} aspect ratio out of range: size={width}x{height} supported=1:4..4:1', 400)
-    return refs
+    from modules.minimax import minimax_references
+    return minimax_references.resolve(workflow, references, init_image)
 
 
 def run(selected: models_def.Model, *,
@@ -188,10 +186,8 @@ def run(selected: models_def.Model, *,
         # unresized since the pipeline defines its own canvas placement per anchor
         p.video_still = int(frames) <= 1
         if refs is not None:
-            from diffusers.modular_pipelines.minimax_h3 import MiniMaxH3ImageReference
-            # references outrank the keyframe inputs in every block, so those stay unset; the reference
-            # encoder reads the image array as (height, width, 3) and never converts it itself
-            p.task_args['references'] = [MiniMaxH3ImageReference(image=image.convert('RGB')) for image in refs]
+            # references outrank the keyframe inputs in every block, so those stay unset
+            p.task_args['references'] = refs
             if last_image is not None:
                 log.warning(f'Video: op=reference model="{selected.name}" last frame not supported, ignoring')
         else:
diff --git a/test/test-video-references.py b/test/test-video-references.py
index fae7d0596..2290e7b3e 100644
--- a/test/test-video-references.py
+++ b/test/test-video-references.py
@@ -556,6 +556,47 @@ def test_short_video_is_rejected_after_the_decode():
     return expect_error(lambda: refs.resolve('ref2va', [image(), fn]), 'video too short')
 
 
+# ============================================================
+# Core seam
+# ============================================================
+
+def test_core_serves_the_caps():
+    from modules.video_models import video_run
+    assert video_run.reference_caps('ref2va') is caps(), 'the core served different caps than the resolver'
+    assert video_run.reference_caps('fl2va') is None, 'fl2va reported reference limits'
+    assert video_run.reference_caps(None) is None, 'a missing workflow reported reference limits'
+
+
+def test_core_gate_and_caps_table_agree():
+    from modules.video_models import video_run
+    assert set(video_run.REFERENCE_WORKFLOWS) == set(refs.REFERENCE_CAPS), f'{video_run.REFERENCE_WORKFLOWS} vs {sorted(refs.REFERENCE_CAPS)}'
+
+
+def test_core_gate_names_workflows_the_registry_carries():
+    from modules.video_models import models_def, video_run
+    known = {row.workflow for rows in models_def.models.values() for row in rows if getattr(row, 'workflow', None)}
+    assert set(video_run.REFERENCE_WORKFLOWS) <= known, f'{video_run.REFERENCE_WORKFLOWS} not among {sorted(known)}'
+
+
+def test_core_rejects_references_on_a_keyframe_model():
+    from modules.video_models import models_def, video_run
+    row = models_def.Model(name='test keyframe', workflow='fl2va')
+    expect_error(lambda: video_run.validate_references(row, [image()], None), 'requires a reference workflow')
+    assert video_run.validate_references(row, None, None) is None, 'a keyframe model resolved references'
+    assert video_run.validate_references(row, [], image()) is None, 'a keyframe model claimed its init image'
+
+
+def test_core_resolves_a_reference_model():
+    from modules.video_models import models_def, video_run
+    row = models_def.Model(name='test reference', workflow='ref2va')
+    expect_error(lambda: video_run.validate_references(row, None, None), 'No reference media provided')
+    if not has_diffusers():
+        return 'diffusers not installed'
+    built = video_run.validate_references(row, None, image()) # the init image stands in for a single reference
+    assert len(built) == 1 and built[0].kind == 'image', f'{built}'
+    return True
+
+
 # ============================================================
 # Runner
 # ============================================================
@@ -655,6 +696,17 @@ def run_all():
         ]:
             run_test(cat, fn)
 
+        log.warning('=== core seam ===')
+        cat = category('core')
+        for fn in [
+            test_core_serves_the_caps,
+            test_core_gate_and_caps_table_agree,
+            test_core_gate_names_workflows_the_registry_carries,
+            test_core_rejects_references_on_a_keyframe_model,
+            test_core_resolves_a_reference_model,
+        ]:
+            run_test(cat, fn)
+
     log.warning('=== Results ===')
     total_passed = 0
     total_failed = 0