45 lines
1.7 KiB
Markdown
45 lines
1.7 KiB
Markdown
# Connected-chunk continuation
|
||
|
||
This workflow continues a previous MiniMax H3 clip at 960×544 for 10 seconds,
|
||
creating a 20-second stitched sequence when the 10-second baseline source is
|
||
used.
|
||
It uses the MM-R2V-68 generation settings and adds a squirrel to the bright
|
||
park-and-robins scene.
|
||
|
||
## Before running
|
||
|
||
1. Put the source clip in the ComfyUI input directory as
|
||
`continuation_source.mp4`.
|
||
2. The supplied source is the 10-second MM-R2V-68 baseline at 24 FPS. The
|
||
final 22 frames are loaded as the motion tail (`batch_index: 218`,
|
||
`length: 22`).
|
||
3. If the source duration differs, change `batch_index` to
|
||
`(source_frames - 22)`.
|
||
4. The connected-chunk node uses the visual tail directly and provides the
|
||
continuation latent to the sampler. The server's video loader does not
|
||
expose a latent output, so the optional latent-guide node is not used in
|
||
this tested template.
|
||
|
||
## Fixed generation settings
|
||
|
||
- Resolution: 960×544
|
||
- Continuation duration: 10 seconds (20 seconds total with the supplied source)
|
||
- SLA: 0.85, block 32
|
||
- Comfy Kitchen attention
|
||
- Low-VRAM Attention: 4 head chunks
|
||
- Chunk FeedForward: 2 chunks
|
||
- Turbo LoRA
|
||
- 8 steps, Res Multistep sampler, Simple scheduler, CFG 1
|
||
- Seed: 20260927
|
||
|
||
The continuation prompt introduces a friendly grey squirrel while preserving
|
||
the character, park, lighting, and motion from the source clip.
|
||
|
||
## UI workflow
|
||
|
||
`connected-chunk-continuation-ui.json` is the web-editor version of the same
|
||
graph. Load it into ComfyUI, upload/select `continuation_source.mp4`, and
|
||
adjust `ImageFromBatch.batch_index` when using a source with a different frame
|
||
count. The source audio is connected to the continuation node so the original
|
||
soundtrack is retained.
|