A port of PromptForge's Cast list (13-14 Oct 2026), widened to the place, the
style and each pose: one revamp of how the model sees and describes the
pictures. In labelled REF2VA (and for characters in I2VA / FL2VA / L2VA) the
panel's "Cast and scene" section shows one editable card per character, and
one for the place, the style and each Pose picture, 600 characters at most.
Reading (POST /dasiwa/h3/forge/cast, nodes/h3_cast.py):
- One vision call per card with the model picked in Forge. A character is read
from all their pictures and the frames they stand in, each from their people
box where it shows them alone (boxes overlapping more than 35% of the smaller
are read whole, by position); a picture of several with no order and no
boxes is read once as a group, each person their own line. The place, the
style and a pose are read from their own pictures.
- The system prompt is the bundle's new picture_description key (PromptForge's
prompts/_shared/picture-description.md, carried as that one key). The
message ports describeMessage: which person this is and whom to leave out,
the brief, a note typed on one of several pictures said for that picture
only, the setting only for a place behind the characters, and stance only
for a pose.
- Cleaned in code: a leaked </think>, code fences and a signed-off UNAVAILABLE
go; a character card never opens with "Character 2 is..." (H3 cannot place
it); a pose card keeps only the stance (PromptForge's poseOnly).
- Reads happen only in the open panel, never in the queue: empty cards fill
when the panel opens, Find people reads everyone it boxes, Scan people
reads everyone in a picture again from the boxes as they are, and Scan
picture reads a place, style or pose card again.
- The model stays loaded across the reads and unloads 30 s after the last
one, at once when the panel closes (POST /dasiwa/h3/forge/release), when a
draft starts with another model, and before the Director runs. Find people
and reads wait their turn behind a call already running rather than
refusing; the panel holds its card queue while Find people or a draft runs.
Models in ComfyUI/models/llm still load per call (the loader frees them).
Cards:
- Saved in the Director node's properties (dasiwaH3ForgeCast): characters by
number, then "place", "style" and "pose:<item id>". They travel with the
workflow; the Director's Clear removes them with the media.
- An untouched card is read again when its pictures, box, order or notes
change; a card the person edited is only marked out of date. A mode switch
is not a change. Renumbering a box moves its card when no other picture
holds those numbers. Clicking a box shows its card.
At Generate the cards are used as written:
- Labelled REF2VA: a card is that subject's look in Subject definitions, never
cut; a character's Retention kept list is made from it. The cast line says
"looks:" (a pose line "pose:"), which the labelled system prompt already
tells the writer to keep, whether or not the writer sees the pictures. What
has a card is not a Descriptions target for a seeing writer.
- I2VA / FL2VA / L2VA: character cards reach the writer beside where each one
stands ("The character on the right in the first frame: ..."), never as
"Character 2".
- Anything without a card is drafted exactly as before. Editing a card drops
the current draft.
Also: the cards sit in columns in the wider panel (people-boxes PR).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
