Save 4 hours! We auto-setup your workflow! Free!

Drop your workflow.json — we handle every dependency, custom node, and model. Just open the link and run.

Auto-Setup Workflow Json (Free) Now!
ComfyUI > Nodes > MiniMaxH3-Director-Seed-Hunt > Muse Minimax Director

ComfyUI Node: Muse Minimax Director

Class Name

MuseMinimaxDirector

Category
Muse Collective
Author
muse-collective-26 (Account age: 138days)
Extension
MiniMaxH3-Director-Seed-Hunt
Latest Updated
2026-10-24
Github Stars
0.03K

How to Install MiniMaxH3-Director-Seed-Hunt

Install this extension via the ComfyUI Manager by searching for MiniMaxH3-Director-Seed-Hunt
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter MiniMaxH3-Director-Seed-Hunt in the search bar
After installation, click the Restart button to restart ComfyUI. Then, manually refresh your browser to clear the cache and access the updated list of nodes.

Visit ComfyUI Online for ready-to-use ComfyUI environment

  • Free trial available
  • 16GB VRAM to 80GB VRAM GPU machines
  • 400+ preloaded models/nodes
  • Freedom to upload custom models/nodes
  • 200+ ready-to-run workflows
  • 100% private workspace with up to 200GB storage
  • Dedicated Support

Run ComfyUI Online

Muse Minimax Director Description

Generates synchronized AI video and audio from timeline scenes with references and frame guidance.

Muse Minimax Director:

MuseMinimaxDirector creates synchronized AI video and audio from a timeline-based scene description using MiniMax H3 generation workflows. It supports two generation approaches: Reference mode, which uses character, background, video, and audio references to preserve identity, appearance, motion, voices, and environments; and First/Last Frame mode, which uses opening and ending images to guide the video composition. Longer projects are automatically divided into shorter generation chunks and joined into one result, with continuation references or frame anchors helping maintain visual and audio continuity between chunks. The node also assembles structured prompts from your timeline, supports dialogue and sound direction, can resize reference images consistently, and can run additional seed variations for comparison before refinement. Its main purpose is to turn a directed sequence of shots into a coherent video with generated sound while giving you control over style, timing, continuity, reference fidelity, and generation quality.

Muse Minimax Director Input Parameters:

mode

Selects the generation workflow. Reference mode uses character, background, video, and audio references from the timeline to guide identity, motion, sound, and scene continuity. First/Last Frame mode uses the first and second character reference slots as the opening and ending frame images for keyframe-guided generation. The default is Reference. The available options are Reference and First/Last Frame. Reference mode is generally suitable when preserving subjects, environments, voices, or reference-video characteristics is important; First/Last Frame mode is suitable when you mainly need the generated clip to begin and, when applicable, end at specific images.

model

Provides the main MiniMax model used for generation. In Reference mode, this should normally contain the Reference or Omni checkpoint because the node uses it for reference-guided video and audio generation. It is also used as a fallback in First/Last Frame mode when model_fl2va is not connected, although that arrangement may produce degraded results because First/Last Frame generation expects the dedicated First/Last Frame checkpoint. This parameter has no numeric range or default value because it is a required MODEL connection.

clip

Provides the CLIP component used to interpret the assembled prompt and condition the MiniMax video generation process. The CLIP model should be compatible with the connected MiniMax checkpoint and is required for converting the timeline description, style instructions, dialogue, soundscape, and reference instructions into generation guidance. This parameter is a required CLIP connection and has no widget default or numeric range.

vae

Provides the visual VAE used to decode the sampled video latent into image frames. It affects how the generated visual information is converted into viewable images, so it should match the model family and workflow being used. This parameter is a required VAE connection and has no widget default or numeric range.

audio_vae

Provides the audio VAE used to decode the audio portion of the joint video-and-audio latent. It is required in both modes because the node always generates a combined audio-and-video latent internally, even when using First/Last Frame guidance. Without a compatible audio VAE, final audio decoding cannot complete. This parameter is a required VAE connection and has no widget default or numeric range.

aspect_ratio

Selects the target width-to-height proportion of the generated video. The node combines this choice with megapixels and multiple to calculate an exact output resolution before generation, and it fits character, background, and keyframe images to that resolution. The default is the widescreen option represented by AspectRatio.WIDESCREEN_H.value. The available options are the aspect-ratio values provided by the node's resolution configuration. Choosing a landscape ratio is generally appropriate for cinematic or standard video compositions, while another available ratio can be used when your project is designed for a different delivery format.

megapixels

Controls the approximate pixel area of each generated frame. Higher values can provide more visual detail and a larger output, but they increase memory use, sampling time, and the cost of candidate scouting passes. The default is 0.98 megapixels, with a minimum of 0.1 and a maximum of 4.0, using increments of 0.02. The actual width and height are calculated from the selected aspect ratio and rounded to the requested multiple, so the final resolution may be slightly different from the exact megapixel estimate.

multiple

Sets the dimension alignment requirement used when calculating the output width and height. Both calculated dimensions are rounded to a multiple of this value, helping the latent and model operate at compatible sizes. The default is 32, with a minimum of 8, a maximum of 128, and a step of 4. Smaller values allow finer resolution control, while larger values impose stronger alignment and may produce dimensions farther from the requested megapixel area. This is an advanced setting and normally does not need to be changed unless your model or workflow requires a particular alignment.

resize_method

Determines how reference images are fitted to the calculated output resolution when their original aspect ratio does not match. The default is crop, with the available options crop, pad, and stretch. crop scales the image until the target is fully covered and center-crops the excess, preserving proportions but possibly removing edges. pad scales the entire image to fit and adds black bars, preserving all content but making the bars part of the reference. stretch resizes directly to the target dimensions, filling the frame but distorting proportions. This setting is applied to character images, background images, and First/Last Frame images.

duration_seconds

Sets the total duration of the finished video. The default is 10.0 seconds, with a minimum of 1.0, a maximum of 120.0, and a step of 0.5. When the requested duration is longer than chunk_duration_seconds, the node divides the timeline into multiple MiniMax generation calls and joins their frames and audio together. The timeline segments are distributed across this total duration, so the value should match the intended length of your edited sequence.

chunk_duration_seconds

Sets the approximate duration of each individual MiniMax generation call. The default is 10.0 seconds, with a minimum of 3.0, a maximum of 15.0, and a step of 0.5. MiniMax H3 generation is designed for short clips, so longer projects are split into chunks of this length; the final chunk may be shorter because it absorbs the remaining duration. In Reference mode, later chunks can use the previous chunk's final frames and audio as soft continuation references. In First/Last Frame mode, later chunks use the previous chunk's last frame as a continuation anchor. Shorter chunks can make continuity easier to manage but require more separate generation calls.

ref_image_size

Controls the resolution used internally for reference images in Reference mode. The default is match, with the available options match and max. match scales references to the generation's pixel area and is faster and less demanding. max preserves reference images up to a short edge of approximately 2048 pixels, which can improve identity fidelity but requires substantially more reference processing during sampling. This setting affects Reference mode only and has no meaningful effect on First/Last Frame mode.

hybrid_continuation

Enables a hard frame-locked continuation strategy for later chunks when using Reference mode. The default is False. When enabled and model_fl2va is connected, the first chunk is generated normally in Reference mode, while each later chunk uses the exact last frame of the preceding chunk as a First/Last Frame-style opening anchor. This can reduce visible jumps at chunk boundaries, but later chunks no longer receive fresh character or background reference-image reinforcement in the same way as the first chunk. If enabled without model_fl2va, the node logs a warning and falls back to normal Reference-mode continuity for every chunk.

seed

Sets the base random seed for generation. The default is 0, with a minimum of 0 and a maximum of 18446744073709551615 (0xffffffffffffffff). Using the same seed with the same settings generally makes the sampling process repeatable, while changing the seed creates a different visual and audio interpretation of the same prompt. The seed also determines the main result and is the basis for any additional candidate passes.

seed_hunt

This legacy boolean is retained for compatibility with older saved workflows. Its default is False, but the current execution logic does not use it to decide whether extra candidate passes run. Additional candidates are controlled independently by candidate_2, candidate_3, and candidate_4. Older workflows may still contain this parameter, but changing it should not be relied upon as the current way to enable seed variations.

steps

Sets the number of sampling steps used to generate each chunk. The default is 20, with a minimum of 1 and a maximum of 100. More steps can improve prompt adherence or visual refinement in some situations, but they increase generation time and do not guarantee a better result. Fewer steps are useful for quick tests and scouting, while higher values can be reserved for a final selected generation.

sampler_name

Selects the sampling algorithm used to create the video and audio latent. The default is res_multistep. The available options are res_multistep, euler, euler_ancestral, and dpmpp_2m. Different samplers can change motion character, detail, stability, and variation. res_multistep is the default workflow choice; use another option only when you have a specific reason or are comparing generation styles.

scheduler

Controls how sampling strength is distributed across the selected steps. The default is simple. The available options are simple, normal, beta, and sgm_uniform. Scheduler choice can affect denoising progression, detail, motion, and overall result stability. Keep the scheduler consistent when comparing seeds so that differences are primarily caused by the seed rather than by multiple changed settings.

shift_video

Sets the video latent shift used by the MiniMax H3 sigma-shift process. The default is 12.0, with a minimum of 0.01, a maximum of 100.0, and a step of 0.01. It is an advanced control that affects the model's denoising behavior for visual content. The value should normally remain at the model workflow's recommended setting; changing it can alter motion and image characteristics and may reduce compatibility with expected model behavior.

shift_audio

Sets the audio latent shift used by the MiniMax H3 sigma-shift process. The default is 3.0, with a minimum of 0.01, a maximum of 100.0, and a step of 0.01. It affects the denoising behavior of the generated audio portion. This is an advanced control and should normally be left at its default unless you are deliberately tuning the audio generation behavior for a compatible model workflow.

timeline_data

Contains the timeline and directing information used to build the generation prompt. The default is {}. The value is a JSON-formatted string containing timeline segments and may include characters, background information, style instructions, dialogue language, soundscape direction, non-diegetic music, reference videos, and reference audio clips. The node reads this data to determine shot descriptions, timing, speaker associations, reference usage, and continuation behavior. If the value is empty or cannot be parsed, the node logs a warning when applicable and uses an empty timeline structure, which can result in a weak or nearly empty prompt. This parameter is a required STRING and has no numeric range.

candidate_2

Requests an additional full generation pass using the base seed plus 1,000,003. The default is False. When enabled, the node repeats the complete video-and-audio generation process with the same settings but a different seed and places the result in the candidate_2_images and candidate_2_audio outputs. Running this pass increases processing time and memory use, but it gives you another interpretation to compare before refinement.

candidate_3

Requests an additional full generation pass using the base seed plus 2,000,006. The default is False. When enabled, the node fills the candidate_3_images and candidate_3_audio outputs. It operates independently of candidate_2 and candidate_4, so you can select only the number of alternatives that are useful for your project. Each enabled candidate adds approximately one complete generation workload.

candidate_4

Requests an additional full generation pass using the base seed plus 3,000,009. The default is False. When enabled, the node fills the candidate_4_images and candidate_4_audio outputs. It is independent of the other candidate controls and is useful when you want a fourth variation for comparison. Because it performs another complete pass, it can substantially increase total execution time.

model_fl2va

Optionally provides the separate MiniMax First/Last Frame checkpoint. This model is used by First/Last Frame mode and by hybrid continuation chunks in Reference mode. The source code recommends connecting the dedicated First/Last Frame checkpoint rather than reusing the Reference or Omni checkpoint in model. If it is not connected, First/Last Frame mode falls back to the main model input and logs a warning, which may produce degraded results. This optional parameter is a MODEL connection with no widget default or numeric range.

Muse Minimax Director Output Parameters:

images

Returns the final generated video frames as an IMAGE sequence. Frames from all generated chunks are concatenated in order, producing the complete visual result. When no extra candidate passes are enabled, this is the normal primary video output. When any of candidate_2, candidate_3, or candidate_4 is enabled, this output is intentionally blocked with an ExecutionBlocker so that an unselected scouting result is not accidentally treated as the final video; in that situation, use one of the candidate image outputs instead.

audio

Returns the decoded generated soundtrack as an AUDIO value containing the waveform and sample rate. Audio from separate chunks is concatenated in sequence so it matches the full generated video duration. When no extra candidate passes are enabled, this is the normal primary audio output. When any additional candidate is enabled, this output is intentionally blocked until you choose a candidate, preventing an unreviewed scouting result from being used as the final soundtrack.

compiled_prompt

Returns the complete prompt text assembled and used for the generation. For multi-chunk projects, it contains a clearly separated section for each chunk, including the shot descriptions, reference instructions, style direction, soundscape, music direction, keyframe alignment information, and continuation notes appropriate to the selected mode. This output is useful for inspecting how the timeline was interpreted, diagnosing prompt problems, documenting a generation, or comparing the prompt structure with the visual and audio result.

ref_images_used

Returns the reference images actually used by the first Reference-mode chunk as an IMAGE sequence. These normally include the loaded character reference images and the background image when one is available, in the dense order used for the prompt's Picture references. In Reference mode this output can be passed to a refinement workflow so the same identity and background references are reused without uploading them again. In First/Last Frame mode, which does not use the Reference-mode Picture tagging system, the output is an empty image tensor.

candidate_1_images

Returns the main generation pass as an IMAGE sequence. It always contains the result generated with the original seed, including when extra candidates are enabled. When candidate scouting is enabled, this is the first option you should compare against candidate_2_images, candidate_3_images, and candidate_4_images before selecting a result for refinement.

candidate_1_audio

Returns the audio generated during the main seed pass. It corresponds to candidate_1_images and uses the original seed. This output remains available even when the primary images and audio outputs are blocked because extra candidates were requested, allowing you to evaluate the first candidate's soundtrack alongside its frames.

candidate_2_images

Returns the additional video candidate generated with the seed offset of 1,000,003 when candidate_2 is enabled. If the candidate is disabled, the node returns an empty image tensor with zero frames rather than a generated result. Use the output to compare an alternate interpretation of the same timeline and reference setup.

candidate_2_audio

Returns the audio paired with candidate_2_images when candidate_2 is enabled. If the candidate is disabled, the node returns an empty audio value with no samples. The audio should be evaluated together with its matching candidate images because a visually attractive variation may have different dialogue, ambience, music, or timing.

candidate_3_images

Returns the additional video candidate generated with the seed offset of 2,000,006 when candidate_3 is enabled. If the candidate is disabled, the node returns an empty image tensor with zero frames. This output provides a third selectable interpretation of the same directed scene and can be connected to a selection or refinement workflow.

candidate_3_audio

Returns the audio paired with candidate_3_images when candidate_3 is enabled. If the candidate is disabled, the node returns an empty audio value with no samples. Use it together with the corresponding image candidate when checking dialogue delivery, soundscape quality, and synchronization.

candidate_4_images

Returns the additional video candidate generated with the seed offset of 3,000,009 when candidate_4 is enabled. If the candidate is disabled, the node returns an empty image tensor with zero frames. It is the fourth available variation and can be used for final comparison or sent to a refinement node after selection.

candidate_4_audio

Returns the audio paired with candidate_4_images when candidate_4 is enabled. If the candidate is disabled, the node returns an empty audio value with no samples. Evaluate this output with its matching video candidate to avoid combining the audio from one seed with the images from another.

Muse Minimax Director Usage Tips:

  • Use Reference mode when character identity, a specific environment, voice timbre, reference motion, or audio continuity matters; use First/Last Frame mode when the most important requirement is controlling the beginning and ending visual states.
  • Connect the correct checkpoint to each model socket. Put the Reference or Omni model in model, and connect the dedicated First/Last Frame checkpoint to model_fl2va for First/Last Frame generation or hybrid continuation.
  • Keep aspect_ratio, megapixels, and multiple consistent with the intended delivery format. Start around the default 0.98 megapixels for practical generation speed, then increase resolution after the motion, identity, and timing are acceptable.
  • Choose crop when preserving natural proportions is more important than keeping every edge, pad when the entire reference must remain visible, and stretch only when some distortion is acceptable.
  • For long videos, use chunk durations between roughly 8 and 15 seconds and describe each shot clearly in the timeline. Shorter chunks can improve control and make it easier to isolate continuity problems, while longer chunks reduce the number of joins.
  • Enable hybrid_continuation for Reference-mode projects where hard frame continuity is more important than reinforcing character references on every later chunk. Ensure model_fl2va is connected before enabling it.
  • Use overall_soundscape and non_diegetic_music in the timeline to guide ambience and music. Write actual scene actions and sounds clearly in each shot description so the model has grounded direction rather than broad or contradictory instructions.
  • Define speaker associations and include clear quoted dialogue in shot descriptions. In Reference mode, pair reference audio with the corresponding character slot when you want voice-timbre guidance.
  • Use candidate_2, candidate_3, and candidate_4 selectively. For affordable scouting, lower megapixels and possibly steps, compare the candidate outputs, then send the chosen candidate and matching audio to a refinement workflow.
  • When comparing seeds, keep the prompt, references, aspect ratio, duration, sampler, scheduler, and sampling steps unchanged. This makes it easier to identify which candidate is genuinely better rather than merely different because of multiple setting changes.
  • Inspect compiled_prompt whenever the result ignores a character, speaker, shot time, reference, or sound instruction. It shows how the timeline was transformed into the prompt that the model actually received.

Muse Minimax Director Common Errors and Solutions:

Could not parse timeline_data

  • Explanation: The timeline string is not valid JSON, so the node cannot read the shots, characters, references, or sound instructions. It falls back to an empty timeline structure.
  • Solution: Reopen or regenerate the timeline data in the timeline editor and ensure that it is valid JSON. Avoid manually inserting unescaped quotation marks into descriptions, and verify that the timeline field is not truncated.

Reference image not found

  • Explanation: A character or background reference points to a file that is no longer available at the stored path.
  • Solution: Re-upload the image through the timeline interface or restore the file to the expected location. Confirm that the image entry contains a valid file reference and that ComfyUI can access the file.

Could not load reference image

  • Explanation: The referenced image exists but could not be opened or converted into a usable RGB image.
  • Solution: Convert the image to a common format such as PNG or JPEG, remove corruption, and upload it again. Avoid files that are incomplete, inaccessible, or using an unsupported image encoding.

Reference video not found

  • Explanation: A reference video entry points to a missing file, so its visual reference cannot be decoded.
  • Solution: Re-upload the video and confirm that the timeline entry refers to the new file. Check that the file remains available to the ComfyUI process.

Reference video decode error

  • Explanation: The video file could not be decoded, or the selected trim range does not contain readable frames. Unsupported codecs, damaged containers, and invalid time ranges can cause this message.
  • Solution: Re-encode the video to a widely supported container such as MP4 with H.264 video, check that it plays normally, and choose a trim range inside the actual clip duration. Keep reference clips reasonably short because the node limits decoded reference video to a maximum of 200 frames per entry.

Reference audio not found

  • Explanation: A reference audio entry points to a missing file, so the requested voice or sound reference cannot be loaded.
  • Solution: Re-upload the audio file and verify its timeline entry. Confirm that the file path is accessible and that the file contains an audio stream.

Reference audio decode error

  • Explanation: The audio file could not be decoded or converted to the required stereo 44.1 kHz format, or the selected trim range contains no samples.
  • Solution: Convert the audio to a standard WAV, MP3, or AAC file, verify that it plays correctly, and select a valid trim range. Check that the file actually contains an audio track rather than only video.

hybrid_continuation is on but model_fl2va isn't connected

  • Explanation: Hybrid continuation was requested, but the dedicated First/Last Frame model is unavailable. The node cannot use hard-locked continuation and falls back to normal Reference-mode continuation.
  • Solution: Connect the correct First/Last Frame checkpoint to model_fl2va if hard frame locking is required. Otherwise, disable hybrid_continuation and continue with the normal soft-reference behavior.

First/Last Frame mode has no model_fl2va connected

  • Explanation: First/Last Frame mode is running without its recommended checkpoint and is falling back to the main model input, which is normally intended for Reference or Omni generation.
  • Solution: Connect a compatible First/Last Frame checkpoint to model_fl2va. Only use the fallback when you understand that the result may be degraded.

Could not determine how to execute node

  • Explanation: One of the internal MiniMax or ComfyUI nodes could not be invoked through its available execution interface. This usually indicates an incompatible ComfyUI version, missing node implementation, or changed internal node API.
  • Solution: Update the relevant ComfyUI and MiniMax node implementations together, confirm that the required H3 nodes are installed, and avoid mixing incompatible versions of custom node packages.

KeyError for an aspect ratio

  • Explanation: The selected aspect_ratio value is not present in the installed aspect-ratio configuration. This can happen when a saved workflow uses an option removed or renamed by a later node version.
  • Solution: Open the node and select one of the currently available aspect-ratio options. Re-save the workflow after updating the node package.

Out of memory

  • Explanation: The requested resolution, reference size, number of sampling steps, chunk length, or number of candidate passes exceeds available GPU or system memory. max reference sizing and multiple candidates can be especially demanding.
  • Solution: Lower megapixels, choose ref_image_size as match, reduce steps, use shorter chunks, disable unused candidates, and process one candidate at a time. Close other GPU-heavy workflows before retrying.

images or audio is blocked after execution

  • Explanation: At least one of candidate_2, candidate_3, or candidate_4 was enabled. The node deliberately blocks the primary outputs so an unselected scouting result is not mistaken for the final result.
  • Solution: Use candidate_1_images through candidate_4_images and their matching audio outputs to select a result. Disable the extra candidate controls when you want the normal images and audio outputs to be populated directly.

candidate_2_images, candidate_3_images, or candidate_4_images is empty

  • Explanation: The corresponding candidate control was disabled, so the node returned an empty image tensor for that candidate.
  • Solution: Enable the matching candidate checkbox and run the node again. Remember that each enabled candidate performs another complete generation pass.

candidate_2_audio, candidate_3_audio, or candidate_4_audio has no samples

  • Explanation: The matching candidate audio output is empty because that candidate was not enabled.
  • Solution: Enable the corresponding candidate control and execute the node again. Always connect candidate audio to the same candidate number as the selected candidate images.

model, clip, vae, or audio_vae is missing

  • Explanation: A required model connection is empty or incompatible, preventing prompt conditioning, latent sampling, visual decoding, or audio decoding.
  • Solution: Connect compatible model, CLIP, visual VAE, and audio VAE nodes. Check that the audio VAE is connected even when you are primarily interested in the video frames.

No usable frames were decoded

  • Explanation: A reference video or continuation source did not provide readable frames within the requested time range.
  • Solution: Check the source file, choose a trim interval that contains actual video frames, and re-encode the file if necessary. For continuation problems, reduce the complexity or duration of the preceding chunk and retry.

No usable audio samples were decoded

  • Explanation: A reference audio clip or embedded video soundtrack contained no readable samples in the selected trim interval.
  • Solution: Confirm that the source has an audio stream, choose a valid trim interval, and convert the source to a standard audio format before uploading it again.

ExecutionBlocker

  • Explanation: This is the node's intentional blocking behavior for the main image and audio outputs when extra candidates were requested. It is not necessarily a generation failure.
  • Solution: Select and use the matching candidate_1 through candidate_4 outputs, or disable all extra candidate controls if you want the main outputs to carry the generated result directly.

Muse Minimax Director Related Nodes

Go back to the extension to check out more related nodes.
MiniMaxH3-Director-Seed-Hunt
RunComfy
Copyright 2025 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.

Muse Minimax Director