Restore Continuum image reference inputs
This commit is contained in:
@@ -93,14 +93,11 @@ the chunk count, the shared chunk duration, and a per-chunk reference bundle.
|
||||
Reference images are inserted as `@R1` through `@R9`; routing is based on the
|
||||
shared directive, soundscape, music, and each chunk's full prompt text.
|
||||
|
||||
The builder exposes nine optional `H3_REF_MODS` sockets, labelled `RefMod 1`
|
||||
through `RefMod 9`. Each connected socket occupies the matching `@R` slot in
|
||||
socket order, so the chips and generated prompt definitions are deterministic;
|
||||
there is no image-socket offset or chained-bundle ambiguity. Visual RefMods are
|
||||
automatically described in the generated shared directive and emitted through
|
||||
the `ref_mods` output for downstream RefMod handling. The typed
|
||||
`reference_images` output remains as an empty compatibility bundle for older
|
||||
workflows; this builder no longer accepts image sockets.
|
||||
The builder exposes nine optional `IMAGE` sockets, labelled `Reference Image 1`
|
||||
through `Reference Image 9`. Each connected socket occupies the matching `@R`
|
||||
slot, and the generated Continuum reference bundle contains those images with
|
||||
per-chunk routing. RefMod inputs and direct RefMod conditioning are deliberately
|
||||
not part of this node.
|
||||
|
||||
The prompt builder intentionally uses the `ReferenceImagesV39Bundle` class
|
||||
from the installed upstream H3 Continuum pack. Install that pack alongside
|
||||
|
||||
@@ -12,15 +12,6 @@ import sys
|
||||
REFERENCE_IMAGES_V39_TYPE = "H3_CONTINUUM_REFERENCE_IMAGES_V39"
|
||||
|
||||
|
||||
class DumasH3RefModBundle(list):
|
||||
"""H3_REF_MODS list with optional per-chunk routing metadata."""
|
||||
|
||||
def __init__(self, values=(), *, chunk_refs=None):
|
||||
super().__init__(values)
|
||||
self.chunk_refs = tuple(tuple(int(index) for index in refs)
|
||||
for refs in (chunk_refs or ()))
|
||||
|
||||
|
||||
def _continuum_reference_bundle_type():
|
||||
"""Find the upstream H3 bundle class without copying its implementation."""
|
||||
candidates = [
|
||||
@@ -42,34 +33,6 @@ def _continuum_reference_bundle_type():
|
||||
)
|
||||
|
||||
|
||||
def _visual_refmods(mods):
|
||||
"""Return positive-weight visual RefMods in stable picker order."""
|
||||
result = []
|
||||
seen = set()
|
||||
for mod, strength in (mods or []):
|
||||
if getattr(mod, "kind", None) == "audio" or float(strength) <= 0:
|
||||
continue
|
||||
key = str(getattr(mod, "path", "") or "") or (getattr(mod, "name", "refmod"), id(mod))
|
||||
if key in seen:
|
||||
continue
|
||||
seen.add(key)
|
||||
result.append((mod, float(strength)))
|
||||
return result
|
||||
|
||||
|
||||
def _refmod_directive(refmods, start_slot, slots=None):
|
||||
"""Build conservative default definitions for RefMods in the prompt."""
|
||||
lines = []
|
||||
for offset, (mod, _strength) in enumerate(refmods):
|
||||
slot = slots[offset] if slots is not None else start_slot + offset
|
||||
name = str(getattr(mod, "name", "visual reference") or "visual reference").strip()
|
||||
lines.append(
|
||||
f"@R{slot} is the {name} visual reference; preserve its identity, "
|
||||
"appearance, and important physical details when it is used."
|
||||
)
|
||||
return "\n".join(lines)
|
||||
|
||||
|
||||
MAX_CLIPS = 16
|
||||
DEFAULT_CLIP_SECONDS = 4.0
|
||||
|
||||
@@ -200,8 +163,8 @@ class DumasH3SequencePromptBuilder:
|
||||
"the node emits one absolute Timeline prompt for Sequence Prompt."
|
||||
)
|
||||
CATEGORY = "Dumas/MiniMax H3/Prompt"
|
||||
RETURN_TYPES = ("STRING", "INT", "FLOAT", REFERENCE_IMAGES_V39_TYPE, "H3_REF_MODS", "INT")
|
||||
RETURN_NAMES = ("sequence_prompt", "chunk_count", "chunk_duration", "reference_images", "ref_mods", "refmod_offset")
|
||||
RETURN_TYPES = ("STRING", "INT", "FLOAT", REFERENCE_IMAGES_V39_TYPE)
|
||||
RETURN_NAMES = ("sequence_prompt", "chunk_count", "chunk_duration", "reference_images")
|
||||
FUNCTION = "build"
|
||||
|
||||
@classmethod
|
||||
@@ -218,9 +181,9 @@ class DumasH3SequencePromptBuilder:
|
||||
},
|
||||
"optional": {
|
||||
**{
|
||||
f"ref_mod_{index}": ("H3_REF_MODS", {
|
||||
"display_name": f"RefMod {index} (@R{index})",
|
||||
"tooltip": f"RefMod socket {index}; use @R{index} in a clip prompt.",
|
||||
f"reference_image_{index}": ("IMAGE", {
|
||||
"display_name": f"Reference Image {index} (@R{index})",
|
||||
"tooltip": f"Fixed Reference Image {index}; use @R{index} in a clip prompt.",
|
||||
}) for index in range(1, 10)
|
||||
},
|
||||
},
|
||||
@@ -249,73 +212,28 @@ class DumasH3SequencePromptBuilder:
|
||||
"clips": [clip for clip in sequence["clips"] if clip.get("enabled", True)],
|
||||
}
|
||||
selectors = []
|
||||
routing_directive = str(shared_directive or "").strip()
|
||||
directive = routing_directive
|
||||
directive = str(shared_directive or "").strip()
|
||||
sound = str(soundscape or "").strip() or "N/A"
|
||||
music_text = str(music or "").strip() or "N/A"
|
||||
# Each socket is one deterministic @R slot. This avoids trying to
|
||||
# inspect the contents of a chained bundle in the browser and makes
|
||||
# the socket order match the prompt chips exactly.
|
||||
refmods = []
|
||||
for index in range(1, 10):
|
||||
refmods.extend(_visual_refmods(kwargs.get(f"ref_mod_{index}")))
|
||||
if len(refmods) > 9:
|
||||
raise ValueError("The prompt builder supports at most nine RefMod slots.")
|
||||
authored_text_by_chunk = []
|
||||
for clip in active_sequence["clips"]:
|
||||
authored_text_by_chunk.append("\n".join([
|
||||
routing_directive,
|
||||
*(str(slot.get("text", "")) for slot in clip["slots"]),
|
||||
f"overall_soundscape: {sound}",
|
||||
f"non_diegetic_music: {music_text}",
|
||||
]))
|
||||
# RefMod slots are defined by socket order, but only referenced slots
|
||||
# should be passed to the sampler. Keep the original @R number in the
|
||||
# generated definition so a sparse use such as @R2 remains accurate.
|
||||
referenced_slots = {
|
||||
index for index in range(1, len(refmods) + 1)
|
||||
if any(re.search(rf"@R{index}(?![A-Za-z0-9_])", text, re.IGNORECASE)
|
||||
for text in authored_text_by_chunk)
|
||||
}
|
||||
selected_refmods = [
|
||||
refmod for index, refmod in enumerate(refmods, start=1)
|
||||
if index in referenced_slots
|
||||
]
|
||||
selected_slots = [index for index in range(1, len(refmods) + 1) if index in referenced_slots]
|
||||
selected_position = {slot: position for position, slot in enumerate(selected_slots)}
|
||||
chunk_refs = []
|
||||
for text in authored_text_by_chunk:
|
||||
chunk_refs.append(tuple(
|
||||
selected_position[index]
|
||||
for index in selected_slots
|
||||
if re.search(rf"@R{index}(?![A-Za-z0-9_])", text, re.IGNORECASE)
|
||||
))
|
||||
generated = _refmod_directive(selected_refmods, 1, selected_slots)
|
||||
if generated:
|
||||
directive = "\n\n".join(part for part in (directive, generated) if part)
|
||||
for reference_index in range(1, 10):
|
||||
chunks = []
|
||||
for chunk_index, clip in enumerate(active_sequence["clips"], start=1):
|
||||
# Match against the same per-chunk content emitted by
|
||||
# format_sequence_prompt, not only the editable shot lines.
|
||||
text = "\n".join([
|
||||
routing_directive,
|
||||
directive,
|
||||
*(str(slot.get("text", "")) for slot in clip["slots"]),
|
||||
f"overall_soundscape: {sound}",
|
||||
f"non_diegetic_music: {music_text}",
|
||||
])
|
||||
available = reference_index <= len(refmods)
|
||||
if available and re.search(
|
||||
if kwargs.get(f"reference_image_{reference_index}") is not None and re.search(
|
||||
rf"@(R{reference_index}|image{reference_index})(?![A-Za-z0-9_])",
|
||||
text,
|
||||
re.IGNORECASE,
|
||||
):
|
||||
chunks.append(chunk_index)
|
||||
selectors.append(",".join(map(str, chunks)) or "off")
|
||||
# Keep the typed reference-images output for compatibility with the
|
||||
# Continuum bridge, but it is intentionally empty: this builder now
|
||||
# owns RefMod sockets rather than image sockets.
|
||||
images = (None,) * 9
|
||||
images = tuple(kwargs.get(f"reference_image_{index}") for index in range(1, 10))
|
||||
reference_bundle = _continuum_reference_bundle_type()(
|
||||
images=images,
|
||||
reference_use="Per chunk",
|
||||
@@ -331,8 +249,6 @@ class DumasH3SequencePromptBuilder:
|
||||
len(active_sequence["clips"]),
|
||||
float(default_chunk_seconds),
|
||||
reference_bundle,
|
||||
DumasH3RefModBundle(selected_refmods, chunk_refs=chunk_refs),
|
||||
0,
|
||||
)
|
||||
|
||||
|
||||
|
||||
+9
-101
@@ -84,7 +84,7 @@ def _namespace_class(node_class, *, node_id, category, display_name):
|
||||
namespaced.CATEGORY = category
|
||||
namespaced.NODE_CATEGORY = category
|
||||
if node_id == "DumasH3Continuum_H3ContinuumSamplerV39":
|
||||
namespaced = _add_direct_refmod_input(namespaced)
|
||||
namespaced = _clean_v39_sampler_input(namespaced)
|
||||
namespaced.__module__ = __name__
|
||||
namespaced.__qualname__ = node_class.__qualname__
|
||||
namespaced.__name__ = node_class.__name__
|
||||
@@ -92,13 +92,13 @@ def _namespace_class(node_class, *, node_id, category, display_name):
|
||||
return namespaced
|
||||
|
||||
|
||||
def _add_direct_refmod_input(node_class):
|
||||
"""Add the DUMAS direct RefMod path to the vendored V3.9 sampler.
|
||||
def _clean_v39_sampler_input(node_class):
|
||||
"""Keep the DUMAS V3.9 manual-dimension surface without RefMods.
|
||||
|
||||
Continuum's normal image-reference bundle remains untouched. When a
|
||||
RefMod bundle is connected, the same model-scoped OUTER_SAMPLE bridge used
|
||||
by the maintained RefMod pack is attached to a cloned model for this run.
|
||||
This keeps RefMods on the native H3 ref2va path and avoids global state.
|
||||
The sampler uses Continuum's native V3.9 ``reference_images`` bundle. We
|
||||
retain the DUMAS cleanup that makes linked width/height authoritative and
|
||||
removes duplicate legacy audio controls, without adding a second RefMod
|
||||
conditioning path.
|
||||
"""
|
||||
original_input_types = node_class.INPUT_TYPES
|
||||
|
||||
@@ -107,116 +107,24 @@ def _add_direct_refmod_input(node_class):
|
||||
schema = original_input_types()
|
||||
schema = {name: dict(value) for name, value in schema.items()}
|
||||
required = dict(schema.get("required", {}))
|
||||
# V3.8 exposes both its newer resolution preset controls and the
|
||||
# original aspect/size controls. The DUMAS V3.9 node is intentionally
|
||||
# a clean manual-dimension node: width and height are the one source of
|
||||
# truth and can be converted to linked inputs in ComfyUI.
|
||||
for name in ("aspect", "preset", "custom_mp", "size_source"):
|
||||
required.pop(name, None)
|
||||
schema["required"] = required
|
||||
optional = dict(schema.get("optional", {}))
|
||||
# The bundled audio reference socket supersedes the two legacy scalar
|
||||
# audio inputs; exposing both makes it unclear which route is active.
|
||||
for name in ("reference_audio_1", "reference_audio_vae"):
|
||||
optional.pop(name, None)
|
||||
optional["refmods"] = (
|
||||
"H3_REF_MODS",
|
||||
{
|
||||
"display_name": "RefMods (Optional)",
|
||||
"tooltip": (
|
||||
"Optional MiniMax H3 RefMod bundle. RefMods are injected directly "
|
||||
"into this sampler for every generated chunk."
|
||||
),
|
||||
},
|
||||
)
|
||||
optional["refmod_retention"] = (
|
||||
"FLOAT",
|
||||
{
|
||||
"default": 1.0,
|
||||
"min": 0.0,
|
||||
"max": 1.0,
|
||||
"step": 0.01,
|
||||
"display": "number",
|
||||
"display_name": "RefMod Retention",
|
||||
"tooltip": "Master RefMod strength. 1.0 preserves the selected bundle fully.",
|
||||
},
|
||||
)
|
||||
optional["refmod_max_total_tokens"] = (
|
||||
"INT",
|
||||
{
|
||||
"default": 0,
|
||||
"min": 0,
|
||||
"max": 1048576,
|
||||
"step": 1,
|
||||
"display_name": "RefMod Token Budget",
|
||||
"tooltip": "0 uses the RefMod bundle's normal budget; otherwise enforce this maximum.",
|
||||
},
|
||||
)
|
||||
schema["optional"] = optional
|
||||
return schema
|
||||
|
||||
def run(self, refmods=None, refmod_retention=1.0,
|
||||
refmod_max_total_tokens=0, **kwargs):
|
||||
# The clean DUMAS surface always uses the explicitly supplied width /
|
||||
# height pair. Keep the upstream runtime's legacy arguments internal.
|
||||
def run(self, **kwargs):
|
||||
kwargs["size_source"] = "Manual"
|
||||
kwargs.pop("aspect", None)
|
||||
kwargs.pop("preset", None)
|
||||
kwargs.pop("custom_mp", None)
|
||||
if refmods:
|
||||
model = kwargs.get("model")
|
||||
if model is None:
|
||||
raise ValueError(
|
||||
"DUMAS H3 Continuum Sampler V3.9 requires a model when RefMods are connected."
|
||||
)
|
||||
try:
|
||||
pack = importlib.import_module(
|
||||
f"{__package__}.vendor.dumas_minimax_h3_refmod.nodes"
|
||||
if __package__ else "vendor.dumas_minimax_h3_refmod.nodes"
|
||||
)
|
||||
bridge = importlib.import_module(
|
||||
f"{__package__}.vendor.dumas_minimax_h3_refmod.continuum_bridge"
|
||||
if __package__ else "vendor.dumas_minimax_h3_refmod.continuum_bridge"
|
||||
)
|
||||
blocks = pack._ref_blocks(
|
||||
refmods,
|
||||
refmod_retention,
|
||||
seed=-1,
|
||||
scramble_mode="shuffle",
|
||||
max_total_tokens=int(refmod_max_total_tokens),
|
||||
)
|
||||
chunk_blocks = None
|
||||
chunk_refs = getattr(refmods, "chunk_refs", None)
|
||||
if chunk_refs:
|
||||
# Build the per-chunk payloads from the same selected
|
||||
# RefMods. Ordinary H3_REF_MODS lists have no routing
|
||||
# metadata and retain the upstream all-chunks behavior.
|
||||
chunk_blocks = []
|
||||
for positions in chunk_refs:
|
||||
chunk_mods = [refmods[index] for index in positions]
|
||||
chunk_blocks.append(pack._ref_blocks(
|
||||
chunk_mods,
|
||||
refmod_retention,
|
||||
seed=-1,
|
||||
scramble_mode="shuffle",
|
||||
max_total_tokens=int(refmod_max_total_tokens),
|
||||
))
|
||||
kwargs["model"] = bridge.attach(
|
||||
model, blocks, enabled=True, chunk_blocks=chunk_blocks,
|
||||
)
|
||||
print(
|
||||
"[DUMAS H3 Continuum Sampler V3.9] direct RefMods active "
|
||||
f"({len(blocks)} block(s))"
|
||||
)
|
||||
except Exception as exc:
|
||||
raise RuntimeError(
|
||||
"Unable to attach RefMods directly to DUMAS H3 Continuum "
|
||||
f"Sampler V3.9: {exc}"
|
||||
) from exc
|
||||
return super(node_class, self).run(**kwargs)
|
||||
|
||||
node_class.INPUT_TYPES = INPUT_TYPES
|
||||
node_class.run = run
|
||||
node_class.INPUT_TYPES = INPUT_TYPES
|
||||
return node_class
|
||||
|
||||
|
||||
|
||||
@@ -77,9 +77,8 @@ class H3NodeTests(unittest.TestCase):
|
||||
"Dumas/MiniMax H3/Prompt",
|
||||
)
|
||||
optional = PROMPT.DumasH3SequencePromptBuilder.INPUT_TYPES()["optional"]
|
||||
self.assertEqual(optional["ref_mod_1"][0], "H3_REF_MODS")
|
||||
self.assertEqual(optional["ref_mod_9"][0], "H3_REF_MODS")
|
||||
self.assertNotIn("reference_image_1", optional)
|
||||
self.assertEqual(optional["reference_image_1"][0], "IMAGE")
|
||||
self.assertEqual(optional["reference_image_9"][0], "IMAGE")
|
||||
reference_module = types.ModuleType("upstream.v3.reference_images_v39")
|
||||
reference_module.ReferenceImagesV39Bundle = FakeReferenceBundle
|
||||
sys.modules[reference_module.__name__] = reference_module
|
||||
@@ -90,8 +89,8 @@ class H3NodeTests(unittest.TestCase):
|
||||
'{"clips":[{"slots":[{"start":0,"end":4,"text":"Walk."}]},'
|
||||
'{"slots":[{"start":0,"end":4,"text":"Turn."}]}]}',
|
||||
"Use @R2 ambience.", "Music.",
|
||||
ref_mod_1=[(FakeRefMod("environment"), 1.0)],
|
||||
ref_mod_2=[(FakeRefMod("hero"), 1.0)],
|
||||
reference_image_1="environment",
|
||||
reference_image_2="hero",
|
||||
)
|
||||
self.assertEqual(result[1], 2)
|
||||
self.assertEqual(result[2], 4.0)
|
||||
@@ -101,7 +100,7 @@ class H3NodeTests(unittest.TestCase):
|
||||
finally:
|
||||
sys.modules.pop(reference_module.__name__, None)
|
||||
|
||||
def test_prompt_builder_defines_connected_refmods_and_emits_them(self):
|
||||
def test_prompt_builder_emits_connected_images(self):
|
||||
reference_module = types.ModuleType("upstream.v3.reference_images_v39")
|
||||
reference_module.ReferenceImagesV39Bundle = FakeReferenceBundle
|
||||
sys.modules[reference_module.__name__] = reference_module
|
||||
@@ -109,11 +108,9 @@ class H3NodeTests(unittest.TestCase):
|
||||
result = PROMPT.DumasH3SequencePromptBuilder().build(
|
||||
2, 4, "", '{"clips":[{"slots":[{"start":0,"end":4,"text":"Use @R1."}]},'
|
||||
'{"slots":[{"start":0,"end":4,"text":"Continue with @R1."}]}]}',
|
||||
"", "", ref_mod_1=[(FakeRefMod("hero_ref"), 1.0)],
|
||||
"", "", reference_image_1="hero_ref",
|
||||
)
|
||||
self.assertEqual(result[4][0][0].name, "hero_ref")
|
||||
self.assertEqual(result[5], 0)
|
||||
self.assertIn("@R1 is the hero_ref visual reference", result[0])
|
||||
self.assertEqual(result[3].images[0], "hero_ref")
|
||||
self.assertEqual(result[3].selectors[0], "1,2")
|
||||
finally:
|
||||
sys.modules.pop(reference_module.__name__, None)
|
||||
@@ -128,7 +125,7 @@ class H3NodeTests(unittest.TestCase):
|
||||
'{"enabled":true,"slots":[{"start":0,"end":4,"text":"Use @R1."}]},'
|
||||
'{"enabled":false,"slots":[{"start":0,"end":4,"text":"Do not emit @R1."}]},'
|
||||
'{"enabled":true,"slots":[{"start":0,"end":4,"text":"Continue @R1."}]}]}',
|
||||
"", "", ref_mod_1=[(FakeRefMod("hero"), 1.0)],
|
||||
"", "", reference_image_1="hero",
|
||||
)
|
||||
self.assertEqual(result[1], 2)
|
||||
self.assertEqual(result[3].selectors[0], "1,2")
|
||||
@@ -139,7 +136,7 @@ class H3NodeTests(unittest.TestCase):
|
||||
finally:
|
||||
sys.modules.pop(reference_module.__name__, None)
|
||||
|
||||
def test_only_authored_refmod_tags_are_emitted(self):
|
||||
def test_image_slots_are_preserved_even_when_not_referenced(self):
|
||||
reference_module = types.ModuleType("upstream.v3.reference_images_v39")
|
||||
reference_module.ReferenceImagesV39Bundle = FakeReferenceBundle
|
||||
sys.modules[reference_module.__name__] = reference_module
|
||||
@@ -148,14 +145,13 @@ class H3NodeTests(unittest.TestCase):
|
||||
2, 4, "", '{"clips":['
|
||||
'{"slots":[{"start":0,"end":4,"text":"Use @R2."}]},'
|
||||
'{"slots":[{"start":0,"end":4,"text":"No extra reference."}]}]}',
|
||||
"", "", ref_mod_1=[(FakeRefMod("unused"), 1.0)],
|
||||
ref_mod_2=[(FakeRefMod("used"), 1.0)],
|
||||
ref_mod_3=[(FakeRefMod("also_unused"), 1.0)],
|
||||
"", "", reference_image_1="unused",
|
||||
reference_image_2="used",
|
||||
reference_image_3="also_unused",
|
||||
)
|
||||
self.assertEqual([mod.name for mod, _strength in result[4]], ["used"])
|
||||
self.assertEqual(result[4].chunk_refs, ((0,), ()))
|
||||
self.assertIn("@R2 is the used visual reference", result[0])
|
||||
self.assertNotIn("unused visual reference", result[0])
|
||||
self.assertEqual(result[3].images[1], "used")
|
||||
self.assertEqual(result[3].images[0], "unused")
|
||||
self.assertEqual(result[3].images[2], "also_unused")
|
||||
finally:
|
||||
sys.modules.pop(reference_module.__name__, None)
|
||||
|
||||
|
||||
@@ -49,8 +49,8 @@ class _Sampler(_SamplerBase):
|
||||
|
||||
|
||||
class VendoredNamespaceTests(unittest.TestCase):
|
||||
def test_v39_sampler_exposes_direct_refmod_socket(self):
|
||||
sampler = vendored._add_direct_refmod_input(_Sampler)
|
||||
def test_v39_sampler_keeps_clean_image_reference_surface(self):
|
||||
sampler = vendored._clean_v39_sampler_input(_Sampler)
|
||||
inputs = sampler.INPUT_TYPES()
|
||||
self.assertNotIn("aspect", inputs["required"])
|
||||
self.assertNotIn("preset", inputs["required"])
|
||||
@@ -60,9 +60,9 @@ class VendoredNamespaceTests(unittest.TestCase):
|
||||
self.assertIn("height", inputs["required"])
|
||||
self.assertNotIn("reference_audio_1", inputs["optional"])
|
||||
self.assertNotIn("reference_audio_vae", inputs["optional"])
|
||||
self.assertEqual(inputs["optional"]["refmods"][0], "H3_REF_MODS")
|
||||
self.assertIn("refmod_retention", inputs["optional"])
|
||||
self.assertIn("refmod_max_total_tokens", inputs["optional"])
|
||||
self.assertNotIn("refmods", inputs["optional"])
|
||||
self.assertNotIn("refmod_retention", inputs["optional"])
|
||||
self.assertNotIn("refmod_max_total_tokens", inputs["optional"])
|
||||
result = sampler().run(
|
||||
model=object(), width=128, height=64, size_source="Legacy",
|
||||
aspect="Auto", preset="Draft", custom_mp=0.3,
|
||||
|
||||
Reference in New Issue
Block a user