ComfyUI Node
MiniMaxH3 Enhanced Upscale
Single-node H3 latent upscale and per-step Tiled Diffusion. Automatic conservative tile/temporal planning; optional Director guide + video VAE re-encode endpoint images. Audio is preserved. No bbaudio dependency; no downstream widget introspection.
MiniMaxH3 Enhanced Upscale
- model
- conditioning
- latent
- director_guide
- vae
- start_image
- negative
- sampler
- noise
- continuity_context
- latent
- plan
◄scale2.0►
◄upscale_model▾►
◄steps1►
◄denoise0.20►
◄seed0►
◄cfg1.0►
◄sampler_nameeuler►
◄schedulersimple►
◄memory_budget_mb0►
◄upscale_precisionauto►
◄continuity_soft_refinefalse►
◄continuity_mask_strength1.00►
◄spatial_tilingtrue►
◄temporal_chunkingtrue►
◄spike_guardfalse►
◄guard_sensitivity4.3►
◄guard_strength1.00►
◄guard_modeburn_direction►
◄enforce_memory_budgettrue►
CategoryDaSiWa/MiniMax H3
Inputs (29)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| conditioning | CONDITIONING | — | |
| latent | LATENT | — | |
| scale | FLOAT | 2.01–8 | — |
| upscale_model | COMBO | 1 options: interpolation | |
| steps | INT | 11–100 | — |
| denoise | FLOAT | 0.200–1 | — |
| seed | INT | 00–18446744073709550000 | — |
| director_guideopt | MINIMAX_H3_DIRECTOR_GUIDE | — | |
| vaeopt | VAE | — | |
| start_imageopt | IMAGE | — | |
| negativeopt | CONDITIONING | — | |
| cfgopt | FLOAT | 1.00–20 | — |
| sampler_nameopt | COMBO | euler | 44 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +38 |
| scheduleropt | COMBO | simple | 9 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +3 |
| memory_budget_mbopt | INT | 00–131072 | 0: automatic device-memory estimate. Nonzero caps the planning budget; not a hard allocator limit. |
| sampleropt | SAMPLER | — | |
| noiseopt | NOISE | — | |
| continuity_contextopt | DF_H3_CONTINUITY_CONTEXT | — | |
| upscale_precisionopt | COMBO | auto | Learned latent-upscaler compute precision only. Auto uses ComfyUI device/backend policy. Does not change the incoming diffusion model or VAE. Interpolation does not use this setting. |
| continuity_soft_refineopt | BOOLEAN | false | Active Continuity only: smoothly introduce video refinement over the source-tail overlap. Not RGB color matching. |
| continuity_mask_strengthopt | FLOAT | 1.000–1 | 0: original hard mask. 1: full smoothstep ramp over the existing Continuity overlap. Ignored unless soft refine is enabled and Continuity is active. |
| spatial_tilingopt | BOOLEAN | true | Allow automatic spatial tiling for diffusion. A single full-canvas tile uses native sampling without a tiled wrapper. Off always uses the full canvas. Independent of memory-budget enforcement. |
| temporal_chunkingopt | BOOLEAN | true | Allow automatic temporal windows for learned upscale and diffusion, up to 30 latent tokens per planned window. Short videos may use one window. Off processes the complete video. Independent of memory-budget enforcement. |
| spike_guardopt | BOOLEAN | false | Experimental Talon V4 burn-signature correction after learned upscale. Calibrated on 3d_conv_v1; may remove real detail. Interpolation is unaffected. |
| guard_sensitivityopt | FLOAT | 4.31.5–12 | — |
| guard_strengthopt | FLOAT | 1.000–1 | Correction strength; 0 leaves the latent unchanged and skips guard allocations. |
| guard_modeopt | COMBO | burn_direction | burn_direction: submitted signed burn signature. both: also detect its opposite; experimental, with more false-positive risk. |
| enforce_memory_budgetopt | BOOLEAN | true | On: plan within the estimated diffusion budget with 25% workspace headroom. Off: allow plans beyond that estimate; actual OOM remains possible. Does not disable spatial/temporal planning, ComfyUI memory management, or host RAM and spike guard checks. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| latent | LATENT | — |
| plan | STRING | — |