ComfyUI Node

TS H3 Regenerate

Latent upscale and 2/3-step Euler refinement with NVIDIA's Open-H3-Regenerate recipe. Audio is held clean. Optional first/last frame guidance uses a normal H3 video VAE at final resolution.

By AlexYez·Created 2 years ago·Updated 2 days ago· 21
TS H3 Regenerate
  • model
  • conditioning
  • latent
  • audio_vae
  • first_frame
  • video_vae
  • last_frame
  • latent
  • report
◄upscale_model▾►
◄width1920►
◄height1088►
◄denoise1.00►
◄steps3►
◄seed0►
◄audio_modekeep source audio (lipsync)►
◄precisionbf16►
◄chunk_length0►
◄temporal_overlap17►
◄reference_modeoff►
◄chunk_modeindependent (crossfade)►
CategoryTS/Video

Inputs (19)

NameTypeDefaultDescription
modelMODELOutput of TS H3 Regenerate Model.
conditioningCONDITIONINGUse the original positive conditioning. Full Reference keeps native image, video and audio references from MiniMax H3 Reference to Video.
latentLATENTDenoised H3 video or audio+video latent.
upscale_modelCOMBO1 options: (no upscale models found in: /tmp/ComfyUI/models/latent_upscale_models)
widthINT192064–4096—
heightINT108864–4096—
denoiseFLOAT1.000–11 = release sigma schedule; lower = less redraw, shift-6 grid from a lower starting sigma; 0 = upscale only.
stepsCOMBO32 options: 3, 2
seedINT00–18446744073709550000—
audio_modeCOMBOkeep source audio (lipsync)Source audio is held clean for lipsync. Reference mode encodes silence; mux the original soundtrack after decoding.
precisionCOMBObf16Latent upscaler precision. NVIDIA reference uses bf16.
chunk_lengthINT00–1000000 = whole clip, release recipe. Nonzero = lower-memory chunks, in multiples of 17 frames. The stitching mode controls how chunks share context.
temporal_overlapINT170–100000—
audio_vaeoptVAEH3 audio VAE; required only for silence reference mode.
reference_modeoptCOMBOoffFirst/last uses the normal H3 video VAE and keyframe images at final size. Full Reference uses encoded media from the original Ref2VA conditioning with a clean Ref2VA Regenerate model. Reference resolution and length affect VRAM use.
first_frameoptIMAGEStart image prepared directly from the source at final output size, before strong draft downscaling. Only the first image is used; it is fitted to width/height without cropping.
video_vaeoptVAENormal H3 video VAE for encoding the reference, including its compatible INT8 version. The fast decode-only VAE cannot be used here.
chunk_modeoptCOMBOindependent (crossfade)Lock overlap fixes the whole overlap. Blend overlap fixes its early part and gradually transitions to the redrawn tail. Both continuation modes require nonzero overlap for multiple chunks.
last_frameoptIMAGEEnd image for first and last frames mode. The first image in the batch is center-cropped to the final canvas if necessary, then encoded as a guide at the clip's last frame.

Outputs (2)

NameTypeDescription
latentLATENT—
reportSTRING—