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-Drama3 > Load Drama 3

ComfyUI Node: Load Drama 3

Class Name

Drama3Loader

Category
audio/drama3
Author
PedroMarinhoDev (Account age: 50days)
Extension
ComfyUI-Drama3
Latest Updated
2026-11-12
Github Stars
0.17K

How to Install ComfyUI-Drama3

Install this extension via the ComfyUI Manager by searching for ComfyUI-Drama3
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI-Drama3 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

Load Drama 3 Description

Loads Fish Audio Drama 3 checkpoints with flexible RAM, GPU, or dual-GPU memory management.

Load Drama 3:

Drama3Loader, displayed as “Load Drama 3,” prepares the Fish Audio Drama 3 model for use in a ComfyUI audio-generation workflow. It loads a checkpoint converted with this extension’s conversion tools, checks that the file contains the expected model information and weights, and packages the model with the supporting components needed for generation. Its main benefit is flexible memory management for a large model: it can keep most model weights in system RAM, keep them in GPU memory for more direct access, or distribute processing across two GPUs. Drama 3 uses routed experts—specialized groups of model weights that are selected as needed—so keeping these experts in RAM allows only the currently needed groups to be transferred to the GPU, with recently used groups cached for reuse. This makes it possible to adapt the workflow to different hardware configurations, although substantial memory is still required; the node description identifies the converted int8 ConvRot model as approximately 36 GB. This node prepares the generation model rather than producing playable audio itself.

Load Drama 3 Input Parameters:

model_name

Selects the converted Drama 3 checkpoint to load. This is a required dropdown populated from ComfyUI’s registered diffusion_models file list, normally associated with ComfyUI/models/diffusion_models and any additional configured model paths. The available options depend on the files installed on your system; the source does not specify a fixed default. The checkpoint must be a Drama 3 model converted using this extension’s tools/convert.py model conversion, not an arbitrary diffusion checkpoint or an unconverted model download. The loader checks for conversion metadata and verifies that the model weights match the expected structure. Your selection determines which model file is loaded and its memory requirements; the approximately 36 GB figure in the node description specifically refers to the int8 ConvRot conversion, not necessarily every supported converted file.

experts

Controls where the model’s routed experts are stored. The required options are in RAM and in VRAM, with in RAM as the default. RAM is your computer’s main memory, while VRAM is the memory on your graphics card. According to the node’s tooltip, the routed experts account for about 90% of the model, with 256 experts across each of 40 layers and only 8 experts per layer used for each frame. With in RAM, these weights remain in system memory, the required experts are copied to the GPU as needed, and recently used experts are retained in a VRAM cache; this mode needs system RAM approximately equal to the checkpoint’s file size for the expert-storage requirement, with additional memory needed for the rest of your workload. With in VRAM, the expert weights are kept on the GPU, requiring enough VRAM for the whole checkpoint, potentially distributed across two GPUs, plus runtime overhead. This setting primarily changes memory placement and data-transfer behavior rather than acting as a creative control for the generated speech.

expert_cache_gb

Sets the amount of VRAM, in GB, available for caching recently used experts when experts is set to in RAM. This required floating-point parameter has a minimum of 0.0, a maximum of 256.0, a default of 0.0, and an adjustment step of 0.5. A value of 0.0 means automatic sizing based on the memory available when generation starts; it does not mean that caching is disabled. A positive value specifies a cache budget, allowing you to control how much GPU memory is devoted to retaining experts between uses. A larger cache can reduce repeated transfers from system RAM when experts are reused, but leaves less VRAM for other parts of the workflow. This setting is intended for RAM-resident experts and is not a substitute for the full model memory required by in VRAM mode.

second_gpu

Selects an optional second GPU for part of the model’s processing. This required dropdown defaults to none; the other available choices are supplied dynamically by the extension’s GPU-selection helper, so their exact labels depend on your environment. It is marked as an advanced setting. With none, no second device is requested. With a second GPU selected, the last backbone layers, the head, and the tower are assigned to that device, allowing the model workload and memory usage to be distributed across two GPUs. The tooltip states that the int8 model can fit on two 24 GB cards with experts in VRAM, although other loaded models and runtime allocations still affect available memory. The selected second GPU must differ from ComfyUI’s main processing device; selecting the same device causes an explicit error.

split_layer

Controls where the model’s backbone is divided when a second GPU is selected. This required integer parameter has a minimum of 0, a maximum of 39, and a default of 0, and is marked as an advanced setting. A value of 0 requests automatic balancing based on the two cards’ free memory. A nonzero value specifies the first backbone layer assigned to the second GPU, with the later layers also processed there; the head and tower are placed on the second GPU as described by second_gpu. A lower explicit split moves more backbone layers to the second device, while a higher split leaves more layers on the main device. This is a memory-distribution control rather than an audio-quality setting, and it is relevant when a second GPU is enabled.

Load Drama 3 Output Parameters:

DRAMA3_MODEL

Returns a loaded Drama 3 model object of type DRAMA3_MODEL for downstream nodes that accept this model type. It contains the generation model, ComfyUI model-management wrappers used to manage device placement and loading, the tokenizer that converts text into model-readable tokens, and the line-head data used by the model. Depending on your settings, it can also contain a second model-management wrapper and expert-storage support for the selected memory arrangement. Treat this output as the prepared generation engine for the rest of your Drama 3 workflow: it is not a waveform, an audio file, or a set of generated audio codes. The audio codec is loaded separately using “Load Drama 3 Codec” when your workflow needs encoding or decoding.

Load Drama 3 Usage Tips:

  • Start with experts set to in RAM and expert_cache_gb set to 0.0 if the full model does not fit in your GPU memory. Automatic cache sizing uses the memory available when generation begins, but you still need substantial system RAM and enough VRAM for active processing.
  • Check the converted checkpoint’s file size before loading. For RAM-resident experts, plan for approximately that amount of system RAM for expert storage and leave additional headroom for ComfyUI, your operating system, and other workflow components.
  • Use in VRAM when your GPU configuration can accommodate the complete model plus runtime allocations. Keeping experts on the GPU avoids the on-demand system-RAM-to-GPU transfers used by RAM mode, although actual performance depends on your hardware and workflow.
  • If automatic expert caching leaves insufficient room for other GPU operations, try a smaller explicit expert_cache_gb budget. Increasing this value beyond your available VRAM will not create additional memory.
  • For two-GPU operation, select a device different from ComfyUI’s main GPU and begin with split_layer set to 0. This lets the loader balance placement using the cards’ free memory before you try a manual split.
  • If a manual two-GPU split overloads the main GPU, try a lower nonzero split_layer to place more backbone layers on the second GPU. If the second GPU is overloaded, try a higher split instead; remember that the head and tower also occupy memory on the second device.
  • Use a checkpoint produced by this extension’s tools/convert.py model conversion and place it in a registered diffusion_models location. Load the Fish S2-Pro codec separately through “Load Drama 3 Codec”; the generation checkpoint and codec serve different purposes.

Load Drama 3 Common Errors and Solutions:

Drama 3: the second GPU must differ from ComfyUI's main device …

  • Explanation: This explicit validation error occurs when second_gpu resolves to the same device ComfyUI already uses as its main device. Selecting the main GPU again does not create a two-GPU configuration.
  • Solution: Set second_gpu to none for single-GPU operation, or select a genuinely different GPU. Check ComfyUI’s device configuration if you are unsure which GPU is the main device.

… is not a converted Drama 3 file (tools/convert.py model)

  • Explanation: The checkpoint does not contain the required drama3_config metadata. It may be an unconverted model, the wrong kind of checkpoint, or a file saved without the metadata required by this extension.
  • Solution: Convert the Drama 3 model using the extension’s tools/convert.py model conversion and select the resulting checkpoint. Do not select the codec file or an unrelated checkpoint as model_name.

Drama 3 checkpoint keys: unexpected …, missing …

  • Explanation: The model’s weight names do not match the structure expected by the loader. The error reports up to six unexpected names and up to six missing names, helping identify an incompatible or incomplete checkpoint.
  • Solution: Confirm that you selected the correct converted Drama 3 model and that its conversion is compatible with your installed extension. If necessary, repeat the conversion using the tools supplied with that installation and check whether the original files were downloaded completely.

Model file not found or path-resolution error

  • Explanation: The node resolves model_name through ComfyUI’s registered diffusion_models paths before loading. If the selected file was moved, renamed, deleted, or is no longer available through a configured model path, loading cannot proceed; the exact error wording depends on ComfyUI.
  • Solution: Verify that the checkpoint exists in a registered diffusion_models location, refresh the model selection or restart ComfyUI if needed, and select the file again. Check any additional model-path configuration if the file is stored outside the standard models folder.

CUDA out of memory or GPU allocation failure

  • Explanation: This is a possible runtime error rather than a custom message defined by the node. The selected placement may require more VRAM than is available, particularly with experts set to in VRAM, a large explicit expert cache, or an unbalanced two-GPU split.
  • Solution: Try in RAM, reduce an explicit expert_cache_gb budget, and free GPU memory used by other applications or models. With two GPUs, start with split_layer at 0, then adjust the split if one device remains overloaded. Remember that the checkpoint size does not include every runtime memory allocation.

System memory allocation failure or process terminated during loading

  • Explanation: These symptoms can occur when system RAM is insufficient. The in RAM option explicitly requires approximately the checkpoint’s file size in system memory for expert storage, and loading and other applications also need memory. The operating system may terminate the process without a Drama 3-specific error.
  • Solution: Close memory-heavy applications and verify that sufficient system RAM is available. Use in VRAM only if your GPU memory can accommodate the model, or run the workflow on hardware with more system memory. Lowering expert_cache_gb reduces GPU cache allocation, not the underlying RAM requirement for stored experts.

Invalid device or CUDA device-selection error

  • Explanation: A selected second GPU may no longer be available or valid in the current environment, for example after a change in GPU visibility or runtime configuration. Device-validation failures can occur when the selection is converted to a device or when the loader attempts to use it; exact wording depends on PyTorch and the runtime.
  • Solution: Restart ComfyUI after hardware or device-visibility changes, choose an available GPU from the current dropdown, or use none to test single-device loading.

File-loading or checkpoint-reading error

  • Explanation: The model file, or a supporting file used by the loader, may be unreadable, incomplete, or damaged. The loader also reads bundled tokenizer data and uses a separate line-head weights file if those weights are not embedded in the checkpoint. Exact errors depend on which file cannot be read.
  • Solution: Check the file path and read permissions, verify downloads and conversion outputs, and restore missing extension assets from a complete installation. If the error names the tokenizer or line-head file, repair the extension’s supporting files rather than only replacing the model checkpoint.

Load Drama 3 Related Nodes

Go back to the extension to check out more related nodes.
ComfyUI-Drama3
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.

Load Drama 3