Add prompt-driven reference image routing
This commit is contained in:
@@ -30,6 +30,17 @@ outside the clip tabs. The editor state is stored in a serializable JSON
|
||||
widget, making saved workflows portable even when the optional browser panel
|
||||
is unavailable.
|
||||
|
||||
The node can also accept up to nine optional reference-image inputs. Each
|
||||
input has a fixed identity (`Reference Image 1` is always `@R1`, through
|
||||
`@R9`); when an image is connected, the editor shows a button for inserting
|
||||
that tag at the current caret position. The fourth output is the same typed
|
||||
`H3_CONTINUUM_REFERENCE_IMAGES_V39` bundle used by the V3.9 sampler. It is
|
||||
automatically set to `Per chunk`, and only references explicitly mentioned in
|
||||
each chunk's prompt are assigned to that chunk. References that are connected
|
||||
but not mentioned remain off for that chunk. Connect this output directly to
|
||||
the sampler's `Reference Images (Optional)` input instead of adding a separate
|
||||
Reference Images helper node.
|
||||
|
||||
Example output:
|
||||
|
||||
```text
|
||||
|
||||
+42
-3
@@ -3,8 +3,14 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import re
|
||||
from typing import Any, Mapping
|
||||
|
||||
from .v3.reference_images_v39 import (
|
||||
REFERENCE_IMAGES_V39_TYPE,
|
||||
ReferenceImagesV39Bundle,
|
||||
)
|
||||
|
||||
|
||||
MAX_CLIPS = 16
|
||||
DEFAULT_CLIP_SECONDS = 4.0
|
||||
@@ -135,8 +141,8 @@ class DumasH3SequencePromptBuilder:
|
||||
"the node emits one absolute Timeline prompt for Sequence Prompt."
|
||||
)
|
||||
CATEGORY = "DUMAS H3/Prompt"
|
||||
RETURN_TYPES = ("STRING", "INT", "FLOAT")
|
||||
RETURN_NAMES = ("sequence_prompt", "chunk_count", "chunk_duration")
|
||||
RETURN_TYPES = ("STRING", "INT", "FLOAT", REFERENCE_IMAGES_V39_TYPE)
|
||||
RETURN_NAMES = ("sequence_prompt", "chunk_count", "chunk_duration", "reference_images")
|
||||
FUNCTION = "build"
|
||||
|
||||
@classmethod
|
||||
@@ -150,15 +156,47 @@ class DumasH3SequencePromptBuilder:
|
||||
"timeline_json": ("STRING", {"default": state, "multiline": True}),
|
||||
"soundscape": ("STRING", {"default": "", "multiline": True}),
|
||||
"music": ("STRING", {"default": "", "multiline": True}),
|
||||
},
|
||||
"optional": {
|
||||
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)
|
||||
}
|
||||
}
|
||||
|
||||
def build(self, clip_count, default_chunk_seconds, shared_directive, timeline_json, soundscape, music):
|
||||
def build(
|
||||
self,
|
||||
clip_count,
|
||||
default_chunk_seconds,
|
||||
shared_directive,
|
||||
timeline_json,
|
||||
soundscape,
|
||||
music,
|
||||
**kwargs,
|
||||
):
|
||||
sequence = normalize_sequence(
|
||||
timeline_json,
|
||||
clip_count=int(clip_count),
|
||||
default_seconds=float(default_chunk_seconds),
|
||||
)
|
||||
selectors = []
|
||||
for reference_index in range(1, 10):
|
||||
tag = f"R{reference_index}"
|
||||
chunks = []
|
||||
for chunk_index, clip in enumerate(sequence["clips"], start=1):
|
||||
text = "\n".join(str(slot.get("text", "")) for slot in clip["slots"])
|
||||
if kwargs.get(f"reference_image_{reference_index}") is not None and re.search(
|
||||
rf"@{tag}(?![A-Za-z0-9_])", text, re.IGNORECASE
|
||||
):
|
||||
chunks.append(chunk_index)
|
||||
selectors.append(",".join(map(str, chunks)) or "off")
|
||||
images = tuple(kwargs.get(f"reference_image_{index}") for index in range(1, 10))
|
||||
reference_bundle = ReferenceImagesV39Bundle(
|
||||
images=images,
|
||||
reference_use="Per chunk",
|
||||
selectors=tuple(selectors),
|
||||
)
|
||||
return (
|
||||
format_sequence_prompt(
|
||||
shared_directive=shared_directive,
|
||||
@@ -168,6 +206,7 @@ class DumasH3SequencePromptBuilder:
|
||||
),
|
||||
len(sequence["clips"]),
|
||||
float(default_chunk_seconds),
|
||||
reference_bundle,
|
||||
)
|
||||
|
||||
|
||||
|
||||
@@ -61,11 +61,42 @@ def test_normalizer_rejects_invalid_slot_ranges(bad):
|
||||
normalize_sequence(bad, clip_count=1, default_seconds=4)
|
||||
|
||||
|
||||
def test_node_schema_has_single_string_output_and_serialized_editor_state():
|
||||
schema = DumasH3SequencePromptBuilder.INPUT_TYPES()["required"]
|
||||
assert DumasH3SequencePromptBuilder.RETURN_TYPES == ("STRING", "INT", "FLOAT")
|
||||
def test_node_schema_has_prompt_outputs_reference_bundle_and_serialized_editor_state():
|
||||
inputs = DumasH3SequencePromptBuilder.INPUT_TYPES()
|
||||
schema = inputs["required"]
|
||||
assert DumasH3SequencePromptBuilder.RETURN_TYPES == (
|
||||
"STRING", "INT", "FLOAT", "H3_CONTINUUM_REFERENCE_IMAGES_V39"
|
||||
)
|
||||
assert "reference_image_1" in inputs["optional"]
|
||||
assert schema["timeline_json"][1]["multiline"] is True
|
||||
result = DumasH3SequencePromptBuilder().build(1, 4, "Directive", json.dumps(default_sequence(1, 4)), "N/A", "N/A")
|
||||
assert result[0].startswith("[0-4s]\nDirective")
|
||||
assert result[0].count("overall_soundscape:") == 1
|
||||
assert result[1:] == (1, 4.0)
|
||||
assert result[1] == 1
|
||||
assert result[2] == 4.0
|
||||
assert result[3].reference_use == "Per chunk"
|
||||
assert result[3].selectors == ("off",) * 9
|
||||
|
||||
|
||||
def test_builder_routes_connected_references_only_to_mentioning_chunks():
|
||||
value = {
|
||||
"clips": [
|
||||
{"slots": [{"start": 0, "end": 4, "text": "@R1 walks toward @R2."}]},
|
||||
{"slots": [{"start": 0, "end": 4, "text": "The camera follows @r2."}]},
|
||||
{"slots": [{"start": 0, "end": 4, "text": "A wide shot with no reference."}]},
|
||||
]
|
||||
}
|
||||
result = DumasH3SequencePromptBuilder().build(
|
||||
3,
|
||||
4,
|
||||
"",
|
||||
json.dumps(value),
|
||||
"",
|
||||
"",
|
||||
reference_image_1=object(),
|
||||
reference_image_2=object(),
|
||||
)
|
||||
bundle = result[3]
|
||||
assert bundle.reference_use == "Per chunk"
|
||||
assert bundle.selectors == ("1", "1,2", "off", "off", "off", "off", "off", "off", "off")
|
||||
assert bundle.images[0] is not None and bundle.images[1] is not None
|
||||
|
||||
@@ -241,6 +241,31 @@ function configure(node) {
|
||||
shotChips.append(chip);
|
||||
});
|
||||
panel.append(shotChips);
|
||||
const referenceRow = el("div"); referenceRow.style.cssText = "display:flex;gap:4px;overflow:hidden;margin:4px 0 7px;align-items:center";
|
||||
const connectedReferences = Array.from({ length: 9 }, (_, offset) => offset + 1)
|
||||
.filter((index) => node.inputs?.some((input) => input.name === `reference_image_${index}` && input.link != null));
|
||||
if (connectedReferences.length) {
|
||||
const referenceLabel = el("span", "Refs:"); referenceLabel.style.cssText = "color:#8fa898;font:10px monospace;margin-right:2px"; referenceRow.append(referenceLabel);
|
||||
connectedReferences.forEach((index) => {
|
||||
const button = el("button", `@R${index}`); button.type = "button";
|
||||
button.title = `Insert reference image ${index} as @R${index}`;
|
||||
button.style.cssText = "padding:3px 8px;border:1px solid #4f9bb0;border-radius:4px;background:#172d33;color:#8bd5e7;font:11px monospace;cursor:pointer";
|
||||
button.onclick = (event) => {
|
||||
event.preventDefault(); event.stopPropagation();
|
||||
if (!prompt) return;
|
||||
const start = prompt.selectionStart ?? prompt.value.length;
|
||||
const end = prompt.selectionEnd ?? start;
|
||||
const token = `@R${index} `;
|
||||
prompt.value = prompt.value.slice(0, start) + token + prompt.value.slice(end);
|
||||
const caret = start + token.length;
|
||||
prompt.focus(); prompt.setSelectionRange(caret, caret);
|
||||
promptText = prompt.value; clip.prompt = prompt.value;
|
||||
syncPrompt(clip, clip.prompt); persistState();
|
||||
};
|
||||
referenceRow.append(button);
|
||||
});
|
||||
}
|
||||
panel.append(referenceRow);
|
||||
const promptLabel = el("div", "PROMPT SCRIPT"); promptLabel.style.cssText = "color:#8fa898;font-size:10px;letter-spacing:.08em;margin:4px 0 3px"; panel.append(promptLabel);
|
||||
prompt = el("textarea"); prompt.rows = 10; prompt.value = promptText; prompt.placeholder = "[Shot 1] At 00:00.000, describe the opening action...\n\nType @ at the start of a new line to add the next shot."; prompt.style.cssText = "width:100%;box-sizing:border-box;color:#e8e8e8;background:#171717;border:1px solid #4e4e4e;border-radius:4px;padding:8px;font:12px/1.45 monospace;margin-top:2px;resize:vertical;outline:none";
|
||||
prompt.oninput = () => { promptText = prompt.value; clip.prompt = prompt.value; };
|
||||
@@ -303,6 +328,12 @@ function configure(node) {
|
||||
setState(); render(); return result;
|
||||
})(secondsWidget.callback);
|
||||
stateWidget.callback = ((previous) => function(...args) { const result = previous?.apply(this, args); render(); return result; })(stateWidget.callback);
|
||||
const previousConnections = node.onConnectionsChange;
|
||||
node.onConnectionsChange = function(...args) {
|
||||
const result = previousConnections?.apply(this, args);
|
||||
render();
|
||||
return result;
|
||||
};
|
||||
node.__dumasSequencePromptBuilder = refresh;
|
||||
refresh();
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user