ComfyUI Node
IronLeo78 2in1 Revitalizer
All-in-one: loads a first-pass latent + conditioning saved by the IronLeo78 MTV, upscales the video latent (Minimax H3 Latent Upscaler 3D), refines it (guider + sampler + sigmas), decodes video and audio, builds the video and saves it. by IronLeo78
IronLeo78 2in1 Revitalizer
- video
- images
- audio
- latent
- status
◄latent▾►
◄conditioning(match the latent)►
◄freeze_audiotrue►
◄unet_name(none found in models/diffusion_models)►
◄vae_video(none found in models/vae)►
◄vae_audio(none found in models/vae)►
◄lora_onfalse►
◄lora_name(none)►
◄lora_strength1.00►
◄attention_backendcomfy kitchen attention►
◄upscale_ontrue►
◄upscale_model(none found in models/latent_upscale_models)►
◄megapixels2.0►
◄temporal_chunkingtrue►
◄force_unloadtrue►
◄upscale_devicecuda►
◄upscale_precisionfp16►
◄seed0►
◄sampler_nameeuler►
◄filename_prefixvideo/MiniMax_H3►
CategoryIronLeo78/MTV
Inputs (20)
| Name | Type | Default | Description |
|---|---|---|---|
| latent | COMBO | The first-pass latent (<clip>_video.latent). Lists ComfyUI/output (newest first) and uploaded files. Use the Upload button to add files from anywhere. | |
| conditioning | COMBO | (match the latent) | '(match the latent)' loads the <clip>.conditioning saved with the selected latent. Pick another file to use a different clip's conditioning. |
| freeze_audio | BOOLEAN | true | ON: the audio latent gets a zero noise mask (as in the MTV), so the refine pass does not re-noise the song. |
| unet_name | COMBO | (none found in models/diffusion_models) | Diffusion model (UNETLoader). |
| vae_video | COMBO | (none found in models/vae) | H3 video VAE. |
| vae_audio | COMBO | (none found in models/vae) | H3 audio VAE. |
| lora_on | BOOLEAN | false | The 'LoRa OFF/ON' switch of the workflow (model-only LoRA, e.g. the 4-step turbo). |
| lora_name | COMBO | (none) | 1 options: (none) |
| lora_strength | FLOAT | 1.00-10–10 | — |
| attention_backend | COMBO | comfy kitchen attention | Model Attention Backend. 'comfy kitchen attention' falls back to PyTorch attention by itself when it is not available. |
| upscale_on | BOOLEAN | true | OFF: refine at the saved size (no Separate / Upscaler / Concat). |
| upscale_model | COMBO | (none found in models/latent_upscale_models) | 1 options: (none found in models/latent_upscale_models) |
| megapixels | FLOAT | 2.00.1–2 | Target size of the upscaled video, in megapixels (max 2.0). |
| temporal_chunking | BOOLEAN | true | — |
| force_unload | BOOLEAN | true | Unload the upscale model to CPU afterwards to free VRAM. |
| upscale_device | COMBO | cuda | 3 options: cuda, rocm, cpu |
| upscale_precision | COMBO | fp16 | 3 options: fp16, bf16, fp32 |
| seed | INT | 00–18446744073709550000 | Use the clip's own seed to repeat it (the refine pass should use the same seed). |
| sampler_name | COMBO | euler | 44 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +38 |
| filename_prefix | STRING | video/MiniMax_H3 | The video is always saved (format, codec and quality: auto). |
Outputs (5)
| Name | Type | Description |
|---|---|---|
| video | VIDEO | — |
| images | IMAGE | — |
| audio | AUDIO | — |
| latent | LATENT | — |
| status | STRING | — |