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 > ComfyUI-TaoMateHD > TaoMate HD Progressive Sampler

ComfyUI Node: TaoMate HD Progressive Sampler

Class Name

TaoMateHDProgressiveSampler

Category
model/sampling/taomate
Author
slmonker (Account age: 765days)
Extension
ComfyUI-TaoMateHD
Latest Updated
2026-11-09
Github Stars
0.02K

How to Install ComfyUI-TaoMateHD

Install this extension via the ComfyUI Manager by searching for ComfyUI-TaoMateHD
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI-TaoMateHD 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

TaoMate HD Progressive Sampler Description

Progressively samples MiniMax H3 audio-video from low to high resolution using TaoMate LoRA.

TaoMate HD Progressive Sampler:

TaoMateHDProgressiveSampler provides a specialized progressive sampling process for native MiniMax H3 audio-video generation with a converted TaoMate LoRA. Rather than performing every part of generation at the final video resolution, it works through two lower-resolution stages, lifts the video representation to the target resolution, and completes a final higher-resolution stage. This approach concentrates early generation work on a smaller canvas before refining the larger one. The resolution transition uses coordinated noise at both scales to connect the lower- and higher-resolution stages, while the audio representation continues through the process without that spatial enlargement. The official TaoMate configuration uses three model evaluations, one per stage, with experimental support for additional evaluations and stronger prompt guidance. The node supports MiniMax H3 text-to-audio-video and reference-to-audio-video conditioning, as well as masked generation that preserves designated portions of an existing clean audio-video representation. It is a model-specific sampler, not a general replacement for samplers used with unrelated image or video models.

TaoMate HD Progressive Sampler Input Parameters:

model

A required MODEL connection containing a native MiniMax H3 model patched with a converted TaoMate LoRA. Load the converted LoRA through LoraLoaderModelOnly before connecting the patched model here, as directed by the node description. The node checks that the underlying model is MiniMax H3 and rejects other model types; however, the visible validation does not independently confirm that the required LoRA has been applied. This connection supplies the model that performs all three progressive sampling stages. There is no numeric range or default model, so you must provide a compatible connection.

positive

A required CONDITIONING connection describing what you want MiniMax H3 to generate. The node explicitly supports T2VA, meaning text-to-audio-video, and Ref2VA, meaning reference-to-audio-video, conditioning. Use conditioning prepared for MiniMax H3 rather than assuming conditioning from another model family is interchangeable. The same positive conditioning guides the progressive generation process across the lower- and higher-resolution stages. Its contents are checked before sampling begins, although the supplied excerpt does not show the complete validation rules. There is no numeric range or default conditioning.

seed

A required integer controlling the random noise used to begin generation, including the coupled lower- and higher-resolution video noise and the audio noise. Keeping this value unchanged helps you compare configuration changes against the same noise setup; changing it changes the starting randomness and can produce a different result. The default is 8301, the minimum is 0, and the maximum is 18446744073709551615, equivalent to the unsigned 64-bit maximum. The interface enables an after-generation seed control, so check its selected behavior if you want the seed to remain fixed across repeated runs. A fixed seed does not by itself guarantee identical results after changing other workflow settings or the execution environment.

steps

A required integer specifying the total sampling-step allocation across the three progressive stages. The default and minimum are 3, the maximum is 24, and the increment is 1; this is an advanced setting. At the official value of 3, each stage receives one evaluation. The tooltip identifies 6 as an experimental configuration with a 2/2/2 split, meaning two evaluations in each stage. Increasing this setting subdivides the existing stage noise intervals rather than introducing additional resolution stages. More steps increase the amount of sampling work, but improved quality is not guaranteed because TaoMate is intended for its official short sampling configuration. When classifier-free guidance is enabled above 1.0, additional guidance-related computation can also affect runtime.

cfg

A required floating-point classifier-free guidance scale that controls the use of positive and negative conditioning during sampling. The default and minimum are 1.0, the maximum is 10.0, and the increment is 0.1; this is an advanced setting. TaoMate is distilled for cfg=1.0, making that the intended starting point rather than the larger guidance values commonly used with some other generation models. Values above 1.0 are explicitly experimental and require a connected negative conditioning input. Raising guidance changes the balance between the positive and negative conditioning predictions; it is not a guaranteed improvement in prompt adherence or visual quality and may change the behavior of the distilled sampler.

low_scale

A required floating-point value controlling the spatial scale of the lower-resolution video canvas relative to the final canvas. The official default is 0.625, expressed in the tooltip as 480/768; the minimum is 0.3, the maximum is 0.9, and the increment is 0.001. The first two stages use the derived lower-resolution canvas, and the third stage operates at the target resolution. Smaller values reduce the spatial size processed during the early stages, while larger values keep those stages closer to the final resolution. The resulting dimensions are determined by the canvas-derivation logic, so this should be understood as a resolution scale rather than a direct width or height setting. It changes the early video canvas, not the audio shape or the number of video time positions.

debug

A required BOOLEAN setting with a default of False and the available values True or False; it is marked as advanced. When enabled, it writes diagnostic information to the application logs, including the high- and low-resolution canvas descriptions, noise-level boundaries, total steps, stage-step allocation, guidance scale, normalized mask shapes when masks are present, and the final latent shape. This is useful for checking how your settings are interpreted or gathering information for troubleshooting. It does not add a new output or act as a quality control; its purpose is to make execution details visible.

negative

An optional CONDITIONING connection describing conditioning used as the negative side of classifier-free guidance. It may be omitted at the intended cfg=1.0 setting, but it must be connected when cfg is above 1.0. Use negative conditioning compatible with MiniMax H3 and the rest of your conditioning workflow. When supplied, it is passed to the guidance system alongside the positive conditioning; its practical effect depends on the guidance setting. There is no numeric range or built-in negative prompt. The execution default is None, meaning no negative conditioning is supplied.

latent

A LATENT connection that is marked optional in the interface but is required when the node executes. Its optional interface status exists only to control socket placement, keeping negative below positive; leaving it disconnected causes an error. Provide the native MiniMax H3 nested audio-video latent containing exactly two streams: video and audio. This latent establishes the target video canvas, temporal length, and audio shape. Without a noise_mask, the sampling process starts from newly generated noise rather than using the supplied sample content as a general image-to-image starting point. With a supported noise_mask, the supplied clean video and audio serve as preservation anchors: a mask value of 0 preserves the clean latent, while 1 allows generation. The execution default is None, with no numeric range; ordinary single-stream image latents are not valid substitutes.

TaoMate HD Progressive Sampler Output Parameters:

latent

The single LATENT output contains the completed MiniMax H3 audio-video latent after the two lower-resolution stages, the video resolution lift, and the final target-resolution stage. Its samples entry holds the native nested video and audio streams, not directly viewable frames or playable audio, so continue with decoding and output nodes compatible with MiniMax H3 audio-video latents. The final video uses the target canvas established by the incoming latent, while the audio stream proceeds through sampling without video-style spatial enlargement. When a noise mask is supplied, the node restores the designated preserved portions from the original clean latent before returning the result. It copies the incoming latent dictionary, replaces samples with the generated result, and removes noise_mask; other dictionary entries are retained.

TaoMate HD Progressive Sampler Usage Tips:

  • Start with the intended configuration: steps=3, cfg=1.0, and low_scale=0.625. Establish a reliable baseline before experimenting with settings that differ from the official TaoMate configuration.
  • Load a native MiniMax H3 model, apply the converted TaoMate LoRA using LoraLoaderModelOnly, and connect that patched model to this node. Passing the model-type check alone does not establish that the correct LoRA has been loaded.
  • Always connect latent, even though its socket is marked optional. Use a MiniMax H3 audio-video latent with both native streams and the desired target canvas rather than a standard empty image latent.
  • Keep seed fixed while comparing prompts, step counts, or resolution scales, and verify that the after-generation seed control is not changing it automatically. Change one setting at a time to make comparisons easier to interpret.
  • If you want to explore additional sampling work, try steps=6 as the explicitly documented experimental 2/2/2 configuration. Compare it with the three-step baseline rather than assuming more evaluations produce better results.
  • Keep cfg=1.0 unless you specifically want to test experimental classifier-free guidance. If you increase it, connect compatible negative conditioning before running the workflow.
  • For masked work, provide the clean content you want to preserve in the incoming audio-video latent and use a supported noise_mask. Remember the direction: 0 preserves and 1 generates. An incoming latent without a mask is not treated as a conventional partially denoised starting image.
  • If early-stage spatial processing is too demanding, test a lower low_scale within the allowed range. This reduces the lower-resolution canvas, but the final stage still runs at the target canvas, so it cannot eliminate the memory requirements of that final stage.
  • Enable debug when checking unexpected resolutions, stage allocation, or mask handling, then inspect the ComfyUI logs. These messages can help you confirm the actual derived canvas sizes rather than estimating them from low_scale alone.

TaoMate HD Progressive Sampler Common Errors and Solutions:

TaoMate HD Progressive Sampler requires a native MiniMax H3 model

  • Explanation: The supplied model does not pass the node’s native MiniMax H3 model-type check. A model from another family cannot be used simply because it also generates images or video.
  • Solution: Connect a native MiniMax H3 model supported by your ComfyUI installation, then apply the converted TaoMate LoRA with LoraLoaderModelOnly before sampling.

latent input is required

  • Explanation: The latent socket is disconnected. Although the interface marks it optional for socket ordering, execution explicitly requires it.
  • Solution: Connect a valid native MiniMax H3 audio-video latent containing the desired target canvas and both audio and video streams.

cfg > 1.0 requires the optional negative conditioning input

  • Explanation: Guidance has been raised above 1.0 without supplying negative conditioning. The negative socket is optional only for the intended guidance setting of 1.0.
  • Solution: Return cfg to 1.0 or connect compatible MiniMax H3 negative conditioning. Use 1.0 when you want the official distilled configuration.

TaoMate HD expects the native MiniMax H3 nested AV latent

  • Explanation: The sample representation is not the native nested audio-video structure expected by the sampler. A conventional image latent is not interchangeable with this format.
  • Solution: Use latent preparation and processing nodes designed for native MiniMax H3 audio-video workflows. Check that upstream nodes have not replaced the nested representation with an ordinary tensor.

expected 2 AV streams, got …

  • Explanation: The nested representation contains a number of streams other than two. The node expects one video stream and one audio stream; the actual count replaces the ellipsis in the message.
  • Solution: Supply a correctly constructed MiniMax H3 audio-video latent. Do not remove the audio stream merely because your immediate focus is the video.

Conditioning validation failure — exact message depends on the validator

  • Explanation: The node validates positive conditioning before sampling. The supplied source excerpt does not include the validator’s exact error messages, but missing or incompatible conditioning can prevent execution.
  • Solution: Connect positive conditioning prepared for MiniMax H3 using a compatible T2VA or Ref2VA workflow. If guidance is above 1.0, also check the negative conditioning connection and format.

Latent or noise-mask validation failure — exact message depends on the validator

  • Explanation: The audio-video latent or its optional mask may not satisfy the required structure or shape rules. The provided source does not include the complete latent and mask validators, so their exact messages and accepted mask layouts cannot be listed reliably.
  • Solution: First test a known-compatible MiniMax H3 latent without a mask. If that works, check how the mask is constructed and ensure it is compatible with the native audio-video representation; do not assume every image-mask layout is accepted.

Out-of-memory error — wording varies by device and backend

  • Explanation: The model, audio-video latents, or sampling-stage calculations exceed available memory. Lower-resolution early stages do not remove the need to process the final target-resolution video.
  • Solution: Reduce the target video dimensions or duration in the upstream latent preparation workflow, close other memory-intensive workloads, and use supported ComfyUI memory-management options. Lowering low_scale may help with early stages, but if the failure occurs in the final stage, reduce the target canvas or temporal workload.

TaoMate HD Progressive Sampler Related Nodes

Go back to the extension to check out more related nodes.
ComfyUI-TaoMateHD
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.

TaoMate HD Progressive Sampler