Muse MiniMax H3 Unified Loader (POC):
MuseMiniMaxH3UnifiedLoader prepares the model components used by a MiniMax H3 workflow in one place. It provides preset-based model selection, lets you load only the transformer models your workflow needs, and offers controls for attention and memory use. This can make a workflow easier to configure and help you balance available memory against your preferred model setup. The node also reports which profile was selected, the detected video memory, and a loading status. Some underlying loading behavior comes from an imported loader whose implementation is not included here; where that behavior is not visible, this documentation does not assume how it works.
Muse MiniMax H3 Unified Loader (POC) Input Parameters:
profile
Chooses the loading profile: Automatic, Low VRAM, Balanced, or Maximum Quality. This is the main preset choice for the model setup. Automatic indicates an automatic selection mode, but the supplied source does not show its selection rules. No default is shown in the supplied source.
low_vram_model
Selects the model associated with the Low VRAM profile. It is a file-selection dropdown; its available files and default are not shown. Choose the intended model file for this profile so it is available when that profile is used.
low_vram_clip
Selects the CLIP component associated with the Low VRAM profile. CLIP is the text-processing component supplied to the rest of the workflow. The dropdown contents and default are not shown.
low_vram_mmproj
Selects the mmproj component associated with the Low VRAM profile. The supplied source identifies it as a selectable component but does not show how the imported loader uses it. Available files and the default are not shown.
balanced_model
Selects the model associated with the Balanced profile. Use it to specify the model file intended for that preset. Available dropdown choices and the default are not shown.
balanced_clip
Selects the CLIP component associated with the Balanced profile. Its file choices and default are not shown; select the component intended to accompany your balanced model setup.
balanced_mmproj
Selects the mmproj component associated with the Balanced profile. Its available choices, default, and detailed handling by the imported loader are not shown.
maximum_quality_model
Selects the model associated with the Maximum Quality profile. Its dropdown contents and default are not shown. The profile name describes the preset, but the supplied source does not quantify its quality or memory requirements.
maximum_quality_clip
Selects the CLIP component associated with the Maximum Quality profile. Its available choices and default are not shown.
maximum_quality_mmproj
Selects the mmproj component associated with the Maximum Quality profile. Its available choices and default are not shown; the supplied source does not describe its detailed loading behavior.
low_vram_fl2va
Selects the First/Last model component associated with the Low VRAM profile. The dropdown contents and default are not shown. Whether it is loaded depends in part on the model-loading mode you choose.
balanced_fl2va
Selects the First/Last model component associated with the Balanced profile. Its available choices and default are not shown. It is relevant when your workflow needs the First/Last model.
maximum_quality_fl2va
Selects the First/Last model component associated with the Maximum Quality profile. Its available choices and default are not shown. It is relevant when your workflow needs the First/Last model.
video_vae
Selects the video VAE component to load. A VAE is a model component used by downstream video processing; the supplied source does not show its detailed handling or list its file choices and default.
audio_vae
Selects the audio VAE component to load for downstream audio processing. Its available files and default are not shown, and the supplied source does not describe the imported loader’s detailed handling of it.
lora_1
Selects the first LoRA file. A LoRA is an optional model adjustment; its effect depends on the selected file, strength, and target. Available choices, including any option for no LoRA, and the default are not shown.
lora_1_strength
Sets the strength for the first LoRA. This is a floating-point number, but the supplied source does not show its minimum, maximum, or default. Check the value accepted by the node interface rather than assuming a particular strength range.
lora_1_target
Chooses where the first LoRA is applied: Both models, Reference model, or First/Last model. This lets you direct the adjustment toward the relevant transformer model or both. No default is shown.
lora_2
Selects the second LoRA file, allowing another optional model adjustment. Available dropdown choices and the default are not shown; its strength and target are configured separately.
lora_2_strength
Sets the strength for the second LoRA as a floating-point number. The supplied source does not show a minimum, maximum, or default, so use a value supported by the node interface and the LoRA you selected.
lora_2_target
Chooses where the second LoRA is applied: Both models, Reference model, or First/Last model. No default is shown.
lora_3
Selects the third LoRA file, allowing a further optional model adjustment. Available dropdown choices and the default are not shown.
lora_3_strength
Sets the strength for the third LoRA as a floating-point number. The supplied source does not show a minimum, maximum, or default.
lora_3_target
Chooses where the third LoRA is applied: Both models, Reference model, or First/Last model. No default is shown.
attention_backend
Chooses a per-model attention backend override: Off, comfy kitchen attention, or pytorch attention. Off means this control makes no override; it does not disable SageAttention enabled when ComfyUI launched or the node’s separate Sage toggle. The supplied source does not show a default for this control.
sage_attention
Turns the optional H3 memory-efficient Sage patch on or off. The default is False (OFF). When this and attention_enabled are on, the node checks Sage compatibility before loading models; the patch requires compatible SageAttention, PyTorch/CUDA, and GPU support. Turning this control off does not disable SageAttention enabled when ComfyUI launched.
low_vram_attention
Turns a low-VRAM attention setting on or off. The node accepts a Boolean value, and attention_enabled acts as a master bypass for attention patches. The supplied source does not show this setting’s default or the imported loader’s detailed implementation.
attention_head_chunks
Sets an integer chunk count for an attention-related setting. The supplied source does not show its minimum, maximum, default, or precise effect in the imported loader. Use the node’s displayed limits when adjusting it.
chunk_feed_forward
Turns feed-forward chunking on or off. This is a Boolean setting governed by the attention_enabled master bypass. Its default and detailed behavior in the imported loader are not shown.
feed_forward_chunks
Sets an integer chunk count associated with feed-forward processing. Its minimum, maximum, default, and detailed handling are not shown in the supplied source.
feed_forward_threshold
Sets an integer threshold associated with feed-forward processing. The supplied source does not establish the threshold’s units, minimum, maximum, default, or detailed effect.
activation_reserve_gb
Sets an activation reserve as a floating-point value in gigabytes. The supplied source does not show its minimum, maximum, default, or how the imported loader applies the reserve.
attention_enabled
Acts as the master switch for every attention and feed-forward patch in this loader. The default is True (Attention optimisations ON); setting it to False bypasses these patches, including the node’s high-resolution stability patch. This controls patches made by the loader, not necessarily attention behavior enabled elsewhere in ComfyUI.
models_to_load
Chooses which transformer models the loader should load: Both / Mixed, Reference only, or First/Last only. The default is Both / Mixed. A single-model choice skips loading the unused transformer and its LoRA and attention patches; match this choice to your Director mode, and use Both / Mixed for mixed timelines.
pinned_memory_cap
Chooses a process-wide limit for page-locked system RAM: Default (ComfyUI), 16 GB, 14 GB, 12 GB, 10 GB, 8 GB, 6 GB, or Default (40% of RAM). The default is Default (ComfyUI), which restores the limit captured before Muse changed it; that original limit depends on your operating system and ComfyUI settings. Numeric selections never exceed that original limit or enable pinning if it was disabled. The setting takes effect when this loader runs, is shared by multiple loaders in the same process, and retains the older Default (40% of RAM) label for saved-workflow compatibility.
high_res_stability
Turns on a stability patch that packs H3 Sage attention inputs into contiguous memory before the attention kernel runs, addressing strided-input issues. The default is False (OFF). It is active only when both attention_enabled and sage_attention are on; it may allocate extra VRAM, so leave it off unless you need it.
extra_loras
Provides an optional single-line string for additional LoRA information. Its default is []. The supplied source does not define the accepted syntax beyond that default or show how the imported loader processes the string, so do not assume a particular entry format without checking the loader’s accompanying instructions.
Muse MiniMax H3 Unified Loader (POC) Output Parameters:
model
A MODEL output for the reference-model path. Connect it to a downstream node that expects this model type. The supplied class does not show the imported loader’s exact model-selection and loading steps.
clip
A CLIP output that provides the loaded text-processing component to compatible downstream nodes. Which selected file produces it depends on the loader behavior that is not shown here.
vae
A VAE output for the video VAE component. Connect it where a downstream video workflow expects that component; the supplied class does not describe its internal loading details.
audio_vae
A VAE output for the audio VAE component. Although it shares the VAE socket type with vae, it occupies a separate output position and is intended for the audio component.
model_fl2va
A MODEL output for the First/Last model path. Its role is distinct from model. Because models_to_load can skip an unused transformer, configure that setting to match the path your downstream workflow needs.
selected_profile
A STRING reporting the profile selected during loading. This is especially useful for checking which profile an Automatic setup chose; the supplied source does not specify the format of the reported text.
detected_vram_gb
A FLOAT reporting detected video memory in gigabytes. Use it as a diagnostic value when assessing a profile choice; the supplied source does not show how detection is performed.
status
A STRING containing the loader’s status message. When high_res_stability is on, this message additionally states either High-Res Stability ON (contiguous Sage inputs). or High-Res Stability inactive: enable attention optimisations and Sage Attention. This makes it useful for confirming whether that requested patch is actually active.
Muse MiniMax H3 Unified Loader (POC) Usage Tips:
- Start with
profileset toAutomaticif you want the loader to choose a profile, then checkselected_profileanddetected_vram_gbto see what it reported. If you need a specific preset, select it explicitly. - Match
models_to_loadto your Director mode. ChooseReference onlyorFirst/Last onlywhen the other transformer is unused; retainBoth / Mixedfor mixed timelines. - Keep
attention_enabledon when you want this loader’s attention or feed-forward patches. Turning it off bypasses those patches even if their individual controls remain on. - Enable
sage_attentiononly when the required SageAttention, PyTorch/CUDA, and GPU support are available. Its compatibility check runs before model loading when attention optimisations are enabled. - Leave
high_res_stabilityoff unless you are addressing a strided-input issue with H3 Sage attention. When you enable it, verify thestatusoutput: the patch also needs bothattention_enabledandsage_attention, and may use extra VRAM. - Treat
attention_backendset toOffas “no override from this control,” not as a way to turn off launch-time SageAttention. Likewise, usepinned_memory_capwith the understanding that its limit is process-wide and cannot raise an existing lower limit.
Muse MiniMax H3 Unified Loader (POC) Common Errors and Solutions:
attention_enabled must be an ON/OFF boolean, got ...
- Explanation: The value supplied to
attention_enabledis not a recognized Boolean. The node acceptsTrueorFalse; text such ason,off,enabled,disabled,1, or0; or numeric1or0. - Solution: Set the control through its ON/OFF widget, or correct the value passed by a saved or programmatically built workflow.
sage_attention must be an ON/OFF boolean, got ...
- Explanation: The Sage toggle received an unrecognized value. For example, an arbitrary string will not be interpreted as ON or OFF.
- Solution: Use the node’s Sage ON/OFF widget or supply one of the recognized Boolean values.
high_res_stability must be an ON/OFF boolean, got ...
- Explanation: The high-resolution stability toggle received an unrecognized value instead of a Boolean or supported legacy text toggle.
- Solution: Set it through its ON/OFF widget, or correct the saved workflow value to a recognized ON/OFF value.
low_vram_attention must be an ON/OFF boolean, got ...
- Explanation: The low-VRAM attention setting received an unsupported value.
- Solution: Supply a valid ON/OFF Boolean value and rerun the loader.
chunk_feed_forward must be an ON/OFF boolean, got ...
- Explanation: The feed-forward chunking toggle received an unsupported value.
- Solution: Use its Boolean widget or replace the invalid workflow value with a recognized ON/OFF value.
High-Res Stability inactive: enable attention optimisations and Sage Attention.
- Explanation: You requested high-resolution stability, but the patch is inactive because
attention_enabledorsage_attentionis off. - Solution: Turn on both controls if you need the patch and your system supports SageAttention. Otherwise, turn
high_res_stabilityoff to remove the inactive request.
Sage compatibility check fails before models load
- Explanation: With
attention_enabledandsage_attentionon, the node runs a Sage compatibility check before calling the underlying loader. The exact error text is not present in the supplied source; the tooltip identifies SageAttention, PyTorch/CUDA, and GPU support as requirements. - Solution: Check that those components are compatible in your environment. To proceed without this optional patch, turn
sage_attentionoff and rerun the loader.
A downstream workflow needs a model you skipped
- Explanation:
Reference onlyandFirst/Last onlyare designed to skip the other transformer, including its LoRA and attention patches. A downstream path requiring the skipped model will not have that model loaded as intended. - Solution: Match
models_to_loadto the Director mode you are using, or chooseBoth / Mixedwhen your timeline needs both models.
