Nodes/MiniMax H3 Contex Loop/MiniMax H3 SelfLift Seed Hunt — Experimental
ComfyUI Node

MiniMax H3 SelfLift Seed Hunt — Experimental

Experimental SelfLift seed hunt: saves each low-resolution pass before tiny-VAE preview. Supports Euler or experimental RES4LYF Radau IA 2s (eta=0). Mark one or more takes and choose a main; finish their remaining high-resolution steps sequentially. Saved takes survive OOM/restart. Choose early to finish saving the current candidate, skip the rest and upscale your selection. Turn Review gate off to run one take without a pause, retaining middle-pass recovery. Keep seed fixed to resume. Wire selected_state to Segment Save and Review Gate/Loop End. Review previews require KJNodes' TAEH3 decoder and are silent and approximate, not final-quality renders.

By ethanfel·Created 2 months ago·Updated about 21 hours ago· 463
MiniMax H3 SelfLift Seed Hunt — Experimental
  • state
  • model
  • positive
  • vae
  • latent
  • sampler
  • sigmas
  • negative
  • model_hires
  • highres_tiling
  • output
  • status
  • selected_state
◄seed0►
◄cfg1.0►
◄candidate_count4►
◄batch_namehunt_1►
◄tiny_vae▾►
◄auto_remove_saved_takesfalse►
◄review_enabledtrue►
◄run_moderesume►
Categorysampling/minimax/context_loop

Inputs (18)

NameTypeDefaultDescription
stateH3_CHAIN_STATE—
modelMODEL—
positiveCONDITIONING—
vaeVAE—
latentLATENT—
samplerSAMPLER—
sigmasSIGMAS—
seedINT00–18446744073709550000—
cfgFLOAT1.00–100—
candidate_countINT41–100—
batch_nameSTRINGhunt_1Keep this and the seed unchanged to resume saved takes. Change it for a fresh hunt.
tiny_vaeCOMBO1 options: taeh3.safetensors
negativeoptCONDITIONING—
model_hiresoptMODELOptional compatible H3 diffusion checkpoint for the final full-resolution steps. Unconnected: reuse model. Same sampler, sigmas, CFG, VAE and conditioning; connect desired LoRAs to this model separately. Ignored when SelfLift is off.
highres_tilingoptH3_SELFLIFT_TILINGOptional SelfLift Tiling settings for the final high-resolution denoising steps only. Off/unconnected preserves full-frame sampling. Not TST or latent-upscaler tiling.
auto_remove_saved_takesoptBOOLEANfalseDelete this hunt's temporary latents and previews only after Segment Save successfully saves every marked clip/checkpoint, including the main take. Keep off to retain low passes for later choices. Requires selected_state connected to Segment Save. Failed runs keep recovery files.
review_enabledoptBOOLEANtrueOn: generate candidates and wait for a choice. Off: upscale the saved chosen take, or run only the first/input seed automatically (ignores candidate count). Middle passes still save for recovery; no Tiny-VAE preview is required. Applies when queued; does not change an already running hunt or the final Review Gate.
run_modeoptCOMBOresumeResume reuses this scene's latest saved attempt (including finished results). Regenerate starts a fresh attempt each time this node executes, keeping older takes and saved clips. After a failure, switch back to Resume to recover the new attempt. Applies to each scene in the queued range; seed and sampling settings stay unchanged.

Outputs (3)

NameTypeDescription
outputLATENT—
statusSTRING—
selected_stateH3_CHAIN_STATE—