ComfyUI Node
H3 Video Upsampler
Upscales a finished clip to about 2 MP: H3's 3D latent upscaler enlarges it, then the Regen2K LoRA re-draws fine detail in a few steps. Long clips are cut into overlapping pieces that are re-drawn together, so there are no seams, and it adapts to your GPU's memory by itself.
H3 Video Upsampler
- model
- vae
- images
- conditioning
- audio
- audio_vae
- images
- audio
- report
◄upscale_model▾►
◄megapixels2.00►
◄denoise0.800►
◄loraopen_h3_regenerate_2k_r256pp_bf16.safetensors►
◄tile_tokens0►
◄sampler_nameeuler►
◄redraw_seed42►
◄attentionsparse►
◄chunk_joinstepped►
◄chunkingtrue►
◄tilingtrue►
◄redraw_start0.60►
◄redraw_steps3 (recipe)►
◄fps—►
CategoryPlagueKind/minimax/upscaler
Inputs (20)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | The H3 model (UNETLoader: minimax_h3_fl2va_pruned_int8_convrot). | |
| vae | VAE | The H3 video VAE. Tip: the sol-VAE (minimax_h3_sol_vae_fp16) made a whole run about 20% faster in tests, with hair edges a touch softer. | |
| images | IMAGE | The clip's frames. Any size: it's fitted to a multiple of 32 px. | |
| upscale_model | COMBO | The 3D latent upscaler, from models/latent_upscale_models (minimax_h3_latent_upscaler_3d_bf16.safetensors). | |
| megapixels | FLOAT | 2.000.1–16 | Output size. 2 = about 1920x1080. The shape of the clip is kept. |
| denoise | FLOAT | 0.8000–1 | How strongly the Regen2K LoRA re-draws detail. 0.8 = normal. 0 = no re-draw, the clip is only enlarged. |
| lora | COMBO | open_h3_regenerate_2k_r256pp_bf16.safetensors | The Regen2K LoRA, from models/loras (open_h3_regenerate_2k_r256pp_bf16.safetensors). |
| tile_tokens | INT | 00–1048576 | How much work goes to the GPU at once. 0 = automatic (recommended): it learns what your card can take and remembers it for the next run. A number fixes it: lower uses less VRAM, by cutting the clip into shorter pieces and then tiles. |
| sampler_name | COMBO | euler | Keep euler: Regen2K is tuned for it. |
| redraw_seed | INT | 420–2147483647 | Noise seed for the re-draw. Same seed, same result. |
| attention | COMBO | sparse | sparse (default) = ComfyUI's sparse attention: in side-by-side tests a whole run was 28% faster and looked the same. dense = exact attention, slower. A ComfyUI without sparse attention runs dense. |
| chunk_join | COMBO | stepped | How the overlapping pieces of a long clip are joined. stepped (recommended) = all pieces are re-drawn together and agree at every step, so faces don't change at a join. crossfade = each piece is re-drawn on its own and the overlaps are blended (the old way; a face can shift at a join). Same speed. |
| chunking | BOOLEAN | true | On = a long clip may be cut into overlapping pieces (about 4 s each) when it's too much for the GPU at once. Off = the whole clip in one go: no joins, but VRAM grows with length, so very long clips are cut anyway. |
| tiling | BOOLEAN | true | On = when even a short piece is too much for the GPU, each frame may also be cut into overlapping tiles. Off = never tile; if VRAM runs out, it stops with a message saying what to change. |
| redraw_start | FLOAT | 0.600.3–0.95 | How much the re-draw is allowed to change. 0.6 (default) keeps faces close to your clip. 0.8 = the original Regen2K recipe: sharper, but faces drift more. Lower than 0.6 changes less again and sharpens less. |
| redraw_steps | COMBO | 3 (recipe) | Re-draw steps. 3 = the Regen2K recipe. 2 = a third faster; in side-by-side tests it looked the same. |
| conditioningopt | CONDITIONING | Optional: a text prompt (CLIPTextEncode with the H3 text encoder) describing the clip. Describe the look ("2D anime, flat cel shading, clean line art") more than the objects. Unwired = no prompt, which works well for most clips. | |
| audioopt | AUDIO | The clip's soundtrack. The re-draw listens to it to keep lip sync; the output soundtrack is the original, untouched. | |
| fpsopt | FLOAT | The clip's frame rate (from Video Frames). Needed for lip sync when the clip isn't 24 fps. Unwired = 24. Tip: for 48 fps clips, take the frames at 24 fps and use Save Video's frame_gen x2: half the work, exact sync. | |
| audio_vaeopt | VAE | The H3 audio VAE. Lets the re-draw hear the soundtrack, which keeps mouths in sync. Without it the re-draw hears silence. |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| images | IMAGE | — |
| audio | AUDIO | — |
| report | STRING | — |