ComfyUI Node
JR MiniMax H3 Tail Frame Latent (Experimental)
Experimental endpoint conversion: decodes the true final video frame and encodes it once as a reusable H3 image-keyframe latent. Not lossless or zero-encode. Keeps source resolution. Appended tail_context_latent preserves the original decoder window for video upscale/refinement, without re-encoding that output. Decode it then select the last image; not a Director keyframe. Use the finished final-pass video latent; never an intermediate noisy sampling state.
JR MiniMax H3 Tail Frame Latent (Experimental)
- video_latent
- vae
- decoded_frames
- tail_latent
- tail_image
- status
- tail_context_latent
◄decode_modetail_context►
CategoryJR MiniMax H3/Latent
Inputs (4)
| Name | Type | Default | Description |
|---|---|---|---|
| video_latent | LATENT | Final sampling output -> Split AV Latent -> video_latent. | |
| vae | VAE | Matching H3 video VAE with encoder. Native VAE recommended for the first test. | |
| decode_mode | COMBO | tail_context | Tail context uses the reviewed H3 final 7-token window; unknown contracts fall back to full decode. |
| decoded_framesopt | IMAGE | Optional matching decoded/trimmed video frames at source resolution. Uses their last frame, skips decode. No MP4/JPEG round trip. |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| tail_latent | LATENT | — |
| tail_image | IMAGE | — |
| status | STRING | — |
| tail_context_latent | LATENT | — |