Nodes/MiniMax-H3 Prompt Rewriter/MiniMax-H3 API Universal Writer
ComfyUI Node

MiniMax-H3 API Universal Writer

Writes an H3 prompt through an already running local LLM or a configured remote API. A connection ID and optional model ID are stored in the workflow; API keys are configured locally. The guide or a complete custom system prompt and the raw request reach the selected model. Enabled media are sent as JPEG images, timestamped video frames and WAV audio, with their roles and instructions. Original references reach Reference Slots or Reference Encode in the same numbered order. RefMods require decoded media from their loader, or explicit descriptions mode. No local model is loaded or stopped.

By pytraveler·Created 3 months ago·Updated 4 days ago· 434
MiniMax-H3 API Universal Writer
  • duration
  • ref_0
  • ref_1
  • ref_2
  • ref_3
  • ref_4
  • ref_5
  • ref_6
  • ref_7
  • ref_8
  • ref_9
  • ref_10
  • ref_11
  • options
  • prompt
  • integrated_multimodal_description
  • subject_definitions
  • summary
  • retention_analysis
  • detailed_description
  • overall_soundscape
  • non_diegetic_music
  • reasoning
  • findings
  • report
  • references
◄profile_id▾►
◄taskT2VA►
◄prompt►
◄resolution16:9►
◄max_tokens4096►
◄system_prompt►
◄validationh3►
◄response_formattext►
◄temperature-1.0►
◄seed-1►
◄timeout_seconds180►
◄repair_oncefalse►
◄auto_download_guidetrue►
◄reference_layout{}►
◄reference_instructions{}►
◄image_max_side1024►
◄repeat_lastfalse►
◄library_pick►
◄media_modenative►
◄reference_descriptions{}►
◄video_max_frames8►
◄video_sample_seconds1.00►
◄video_max_seconds15.0►
◄batch_video_fps24►
◄audio_max_seconds30.0►
◄audio_sample_rate16000►
◄bypassfalse►
◄model►
CategoryMiniMax-H3

Inputs (42)

NameTypeDefaultDescription
profile_idCOMBOServer-side connection ID. Use the API providers window beside the model selector, or the local api_admin CLI. Keys are never node inputs.
taskCOMBOT2VAT2VA ignores references. I2VA starts at Picture 1; L2VA ends there; FL2VA starts at Picture 1 and ends at Picture 2. Ref2VA accepts pictures, subjects, clips and sounds.
promptSTRINGThe raw request, sent to the model without trimming or rewriting.
resolutionCOMBO16:98 options: 48:9, 32:9, 21:9, 16:9, 4:3, 1:1, +2
durationFLOAT,INT100.1–600Target clip length in seconds; drives shot count and pacing. Right-click the node for 'duration': the default value back, or a new upper end for the widget. It offers 30 seconds until you change it and the server takes up to 600, because a widget's range is fixed when the node is declared and one number cannot suit every graph -- MiniMax's own guide is written around clips of a few seconds, while the stretched pipelines the community has built run well past that. What you set is remembered with the workflow.
max_tokensINT40961–131072Output budget per request, bounded by the selected profile. Some models count reasoning against this budget too.
system_promptoptSTRINGReplaces the full guide/system message. Leave empty to use the MiniMax-H3 guide.
validationoptCOMBOh3h3 splits and checks the H3 sections. custom returns plain final text without H3 checks.
response_formatoptCOMBOtextMust be supported by the profile. Custom validation requires text.
temperatureoptFLOAT-1.0-1–2-1 omits the parameter and uses the server default.
seedoptINT-1-1–4294967295-1 with fixed control omits seed. Only use a positive or randomized seed when the profile supports it. A cloud seed does not guarantee identical answers.
timeout_secondsoptFLOAT1801–3600Total deadline per HTTP request, including response reading. Cancellation closes our connection; the server may still process the request.
repair_onceoptBOOLEANfalseAllow at most ONE additional, potentially paid request for repairable H3 findings. Off by default; no transport retries.
auto_download_guideoptBOOLEANtrueFetch the MiniMax guide if missing. A custom system prompt skips guide access entirely.
reference_layoutoptSTRING{}Reference strip order, disabled slots and Picture/Subject roles. Stored with the workflow.
reference_instructionsoptSTRING{}JSON object with usage instructions by ref_0...ref_11. Instructions accompany the image; no local captioner runs.
image_max_sideoptINT102464–4096Maximum side of API image copies. Preserves aspect ratio, never enlarges. All batch frames are sent under the same reference label. Original outputs are unchanged.
repeat_lastoptBOOLEANfalseReturn this node's last prompt, or the prompt chosen in its library, without an API request. With nothing kept yet, run once. No server or local model is stopped.
library_pickoptSTRINGWhich saved prompt this node hands on instead of writing one, as JSON written by the library window. It is a widget so the choice is saved with the workflow and reaches an API run: a pick the graph does not carry is a graph that reproduces something else. It applies only while 'repeat_last' is on. That switch is what hands a kept prompt on at all; this says which one, and empty means the node's own last answer.
ref_0optIMAGE,VIDEO,AUDIO,H3_REWRITER_REFMODIMAGE, VIDEO, AUDIO or a decoded RefMod. Images can be Picture, Subject or Video. Click a VIDEO label to include its sound. Originals pass through references.
ref_1optIMAGE,VIDEO,AUDIO,H3_REWRITER_REFMODIMAGE, VIDEO, AUDIO or a decoded RefMod. Images can be Picture, Subject or Video. Click a VIDEO label to include its sound. Originals pass through references.
ref_2optIMAGE,VIDEO,AUDIO,H3_REWRITER_REFMODIMAGE, VIDEO, AUDIO or a decoded RefMod. Images can be Picture, Subject or Video. Click a VIDEO label to include its sound. Originals pass through references.
ref_3optIMAGE,VIDEO,AUDIO,H3_REWRITER_REFMODIMAGE, VIDEO, AUDIO or a decoded RefMod. Images can be Picture, Subject or Video. Click a VIDEO label to include its sound. Originals pass through references.
ref_4optIMAGE,VIDEO,AUDIO,H3_REWRITER_REFMODIMAGE, VIDEO, AUDIO or a decoded RefMod. Images can be Picture, Subject or Video. Click a VIDEO label to include its sound. Originals pass through references.
ref_5optIMAGE,VIDEO,AUDIO,H3_REWRITER_REFMODIMAGE, VIDEO, AUDIO or a decoded RefMod. Images can be Picture, Subject or Video. Click a VIDEO label to include its sound. Originals pass through references.
ref_6optIMAGE,VIDEO,AUDIO,H3_REWRITER_REFMODIMAGE, VIDEO, AUDIO or a decoded RefMod. Images can be Picture, Subject or Video. Click a VIDEO label to include its sound. Originals pass through references.
ref_7optIMAGE,VIDEO,AUDIO,H3_REWRITER_REFMODIMAGE, VIDEO, AUDIO or a decoded RefMod. Images can be Picture, Subject or Video. Click a VIDEO label to include its sound. Originals pass through references.
ref_8optIMAGE,VIDEO,AUDIO,H3_REWRITER_REFMODIMAGE, VIDEO, AUDIO or a decoded RefMod. Images can be Picture, Subject or Video. Click a VIDEO label to include its sound. Originals pass through references.
ref_9optIMAGE,VIDEO,AUDIO,H3_REWRITER_REFMODIMAGE, VIDEO, AUDIO or a decoded RefMod. Images can be Picture, Subject or Video. Click a VIDEO label to include its sound. Originals pass through references.
ref_10optIMAGE,VIDEO,AUDIO,H3_REWRITER_REFMODIMAGE, VIDEO, AUDIO or a decoded RefMod. Images can be Picture, Subject or Video. Click a VIDEO label to include its sound. Originals pass through references.
ref_11optIMAGE,VIDEO,AUDIO,H3_REWRITER_REFMODIMAGE, VIDEO, AUDIO or a decoded RefMod. Images can be Picture, Subject or Video. Click a VIDEO label to include its sound. Originals pass through references.
media_modeoptCOMBOnativenative sends images, timestamped video frames and enabled WAV audio. descriptions sends only explicit text/saved RefMod descriptions; no automatic fallback or local captioner.
reference_descriptionsoptSTRING{}Descriptions by ref_0...ref_11: a string or {text: description, sound: clip sound description}. Required for each enabled component in descriptions mode. Also accepts a wired JSON string.
video_max_framesoptINT81–64Maximum sampled frames per Video reference; all media share a 64-image limit. Sampling spreads across the selected window when this cap is reached.
video_sample_secondsoptFLOAT1.000.05–120Requested video sampling interval. Actual timestamps and the selected count appear in the report.
video_max_secondsoptFLOAT15.00.1–120Read from the clip's active trim start, for at most this limit, the target duration and H3's 15-second reference limit.
batch_video_fpsoptFLOAT241–240Clock for an IMAGE batch assigned the Video role. VIDEO objects use their own timestamps. RefMods use saved frame times.
audio_max_secondsoptFLOAT30.00.1–120Maximum standalone audio/voice window sent to the API. The report shows cropping. Clip sound follows the same window as its video.
audio_sample_rateoptCOMBO16000API copies are mono 16-bit PCM WAV at this rate. Original audio remains unchanged.
bypassoptBOOLEANfalsePass the original prompt and enabled original references through without an API request, guide access or media conversion.
optionsoptH3_API_WRITER_OPTIONSConnect MiniMax-H3 API Writer Options to set generation and media-copy parameters. Connected options override the legacy settings retained by old workflows.
modeloptSTRINGEmpty uses the profile default. An explicit model ID must be allowed locally by this connection profile. Refresh models reads /models; it does not load or switch server processes.

Outputs (12)

NameTypeDescription
promptSTRING—
integrated_multimodal_descriptionSTRING—
subject_definitionsSTRING—
summarySTRING—
retention_analysisSTRING—
detailed_descriptionSTRING—
overall_soundscapeSTRING—
non_diegetic_musicSTRING—
reasoningSTRING—
findingsSTRING—
reportSTRING—
referencesH3_REWRITER_REFERENCES—