akatz-labs / h3-character-swap-tool

# H3 Character Swap
A local video workbench for preparing performances, correcting masks and
rendering character swaps. Build and run the browser UI and GPU worker directly from source on a Linux
x86-64 NVIDIA machine. Docker images are the prebuilt alternative.
https://github.com/user-attachments/assets/e9ed875e-bfe9-42b9-8667-c895ae79f9bc
## Shots and generation segments
Each shot contains editable segments, with one diffusion pass per segment.
Choose 5-, 10- or 15-second segments, split at the playhead, move a boundary,
or merge adjacent segments. Preview a new layout before applying it; previous
layouts retain their directions and takes.
Select a segment in Render to edit its dialogue, speaker, performance prompt,
seed and scene reference. Dialogue starts from the transcript for that segment's
source range. Update prompt rebuilds the performance prompt from the dialogue,
speaker and references; direct edits are sent to H3 exactly as shown. Each
segment saves its own text. The dialogue and prompt editors sit below the player;
generation settings stay in the right column. Anchors contains the character and
scene images.
Drag the Resize video bar beneath the canvas to give playback more room. Its
height is remembered in this browser; arrow keys also resize it, and double-click
resets it. Playback and mask controls stay grouped on wide screens, with stable
space for expansion save messages.
Continue a previous segment with a chosen amount of generated history, or start
fresh at a cut. A continuation can also use an additional reference image.
Generate one segment, or queue it while its predecessor is still rendering.
Queueing freezes that segment’s saved prompt and settings; it automatically
starts with the exact preceding take when ready, even after navigating away.
The popover below Generate offers Queue all segments and Queue from here.
While a render is active, these add the remaining segments without duplicating
active work or current takes. A failed or cancelled predecessor stops dependent
jobs with an explanation. Progress cards appear only for unfinished segments
covered by that job, with future segments shown as queued.
Selecting a different preceding take marks dependent
continuations as needing refresh. Once every selected take is current, the app
assembles the shot with the original source audio. Lossless continuation masters
preserve history across application and worker restarts. Existing whole-shot
takes remain available.
New shots default to Character Swap LoRA V1.5 on MiniMax H3. V1 and V1.5 are
LoRA adapters, not separate base models. Setup offers both as required downloads
from Hugging Face,
with pinned revisions, sizes and SHA-256 hashes in the
resource catalog. Existing verified files are
reused, and existing shots keep their saved LoRA selection. V1 remains selectable
in Advanced settings. Internal checkpoint IDs and paths remain compatible with
previous installations.
## ComfyUI workflow
The standalone workflow and API guid