Nodes/MiniMax H3 Contex Loop/MiniMax H3 SelfLift Project — Experimental
ComfyUI Node

MiniMax H3 SelfLift Project — Experimental

Experimental SelfLift project switch for the dedicated Chain workflow. Plan width/height are the final size; lowres_scale defaults to half-size spatial latents. Optional pixel/VAE consistency correction is disabled by default (rho=0). Does not patch ComfyUI or change any other workflow.

By ethanfel·Created 2 months ago·Updated about 21 hours ago· 463
MiniMax H3 SelfLift Project — Experimental
  • plan
  • plan
  • status
◄enabledfalse►
◄upscaler_model▾►
◄high_resolution_steps2►
◄cleanup_between_stagesfalse►
◄lowres_scale0.50►
◄rho0.00►
◄w_min0.50►
◄w_max1.00►
Categoryconditioning/minimax/context_loop

Inputs (9)

NameTypeDefaultDescription
planH3_CHAIN_PLAN—
enabledBOOLEANfalseProject-wide switch in the dedicated SelfLift workflow. Off uses ordinary single-stage sampling; saved clips are not regenerated automatically.
upscaler_modelCOMBOSelect an installed LBH H3 3D checkpoint, tridae (FP32 clean-latent 2x; downloads verified weights on first use), or bilinear (spatial-only, no weights). Tr1dae requires final dimensions divisible by 64. Required only when enabled; generic image/LTX upscalers are not compatible.
high_resolution_stepsINT21–10000Final steps at the Plan's full resolution. Must be lower than each generated scene's total steps (e.g. 6 low + 2 high out of 8).
cleanup_between_stagesoptBOOLEANfalseRelease the retired checkpoint's DynamicVRAM buffers before lifting, and unload the learned upscaler after use. Tr1dae's small legacy upscaler is specifically offloaded to CPU; other classic/non-dynamic models are skipped. Preserves shared models and saved takes; can slow next-scene reloads.
lowres_scaleoptFLOAT0.500.25–1First-stage width/height relative to the final Plan size, rounded to H3's even latent grid. 0.5 keeps the existing half-resolution behavior. Tr1dae requires an exact 2x lift: use 0.5 and final dimensions divisible by 64.
rhooptFLOAT0.000–1Fraction of the most inconsistent latent locations to correct toward a pixel/VAE anchor. 0 keeps the existing direct lift. Above 0 (with w_max > 0) adds a full video VAE decode, pixel resize and VAE encode; costs time/memory and can change detail/color. Experimental, not a guaranteed artifact fix.
w_minoptFLOAT0.500–1Minimum pixel/VAE correction strength at selected locations. Inactive when rho=0. Must be <= w_max; 0 keeps the direct lift and 1 uses the pixel/VAE anchor.
w_maxoptFLOAT1.000–1Maximum pixel/VAE correction strength at the most inconsistent locations. Inactive when rho=0. Must be >= w_min; setting both weights to 0 skips the pixel/VAE pass.

Outputs (2)

NameTypeDescription
planH3_CHAIN_PLAN—
statusSTRING—