ComfyUI Node
JR MiniMax H3 Regenerate
Video-only H3 regeneration: native VIDEO + video VAE or an existing H3 LATENT. Use MODEL from Regenerate LoRA then optional Unified. Inherit preserves Unified's Sage/FFN. VIDEO returns video LATENT plus original AUDIO bypass. AV input retains the exact audio latent. VIDEO is sampled at 24 fps and end-padded to H3's frame grid; trim on decode if needed.
JR MiniMax H3 Regenerate
- model
- clip
- latent
- video
- vae
- audio_vae
- h3_latent
- status
- original_audio
◄source_mode▾►
◄sigma_schedulelow_denoise_shift6►
◄denoise0.30►
◄steps3►
◄seed1766580932►
◄promptFaithful reconstruction of the input video; preserve motion, composition, colors and fine details.►
◄attentioninherit►
◄resize_modescale►
◄scale2.00►
◄target_megapixels3.69►
◄upscaler_modelauto►
◄chunk_mode5s►
CategoryJR MiniMax H3/Sampling
Inputs (18)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| clip | CLIP | — | |
| source_mode | COMBO | 2 options: H3 latent, video | |
| sigma_schedule | COMBO | low_denoise_shift6 | Internal exact schedules. Low denoise uses 0.15–0.35. Original release fixes its 2/3-step sigmas and ignores denoise. |
| denoise | FLOAT | 0.300.15–0.35 | Starting sigma, not a fraction of discarded sampler steps. |
| steps | INT | 31–50 | — |
| seed | INT | 17665809320–18446744073709550000 | — |
| prompt | STRING | Faithful reconstruction of the input video; preserve motion, composition, colors and fine details. | — |
| attention | COMBO | inherit | For upstream Unified select inherit. cuda++ is the installed KJ SageAttention 2++ path. |
| resize_mode | COMBO | scale | 3 options: keep, scale, megapixels |
| scale | FLOAT | 2.001–4 | — |
| target_megapixels | FLOAT | 3.690.01–16 | — |
| upscaler_model | COMBO | auto | 1 options: auto |
| latentopt | LATENT | — | |
| videoopt | VIDEO | — | |
| vaeopt | VAE | — | |
| audio_vaeopt | VAE | Legacy socket; unused. Original audio bypasses the node. | |
| chunk_modeopt | COMBO | 5s | Sequential hard-prefix windows: 5s=124 frames, 8s=192, 10s=243 at 24 fps. 39-frame overlap. Fixed seed and global noise. off processes the full clip. |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| h3_latent | LATENT | — |
| status | STRING | — |
| original_audio | AUDIO | — |