ComfyUI Node: ComfyArt Precise UNET Loader

ComfyArt Precise UNET Loader:

ComfyArtPreciseUNETLoader, displayed as “ComfyArt Precise UNET Loader” in the ComfyArt category, loads a diffusion model using manual-cast operations rather than the accelerated FP8 matrix-multiplication path enabled by --fast fp8_matrix_mult. Diffusion models are the part of an image-generation workflow that progressively transforms noise into visual content. This node is specifically intended to avoid compatibility problems with unscaled FP8 checkpoints, such as Qwen Image 2512, that can break when that accelerated path is used. Manual casting lets model operations handle the conversion of numerical data types without relying on that particular fast path. The main benefit is a targeted alternative for loading affected models, not a general promise of higher image quality or faster generation. Despite “UNET” in its name, the implementation uses ComfyUI’s diffusion-model loader, so actual model compatibility depends on the installed ComfyUI version.

ComfyArt Precise UNET Loader Input Parameters:

unet_name

A required dropdown selection, with interface type COMBO[STRING], that identifies the diffusion-model file you want to load. Its options are populated dynamically from ComfyUI’s registered diffusion_models locations, so the available choices depend on the files installed in your setup rather than a fixed list included with the node. The node resolves your selection to a full file path and loads that file using manual-cast operations. Your choice determines which diffusion model is available to the rest of your workflow, including its supported architecture and resource requirements; this node does not convert an incompatible checkpoint into a supported one. For example, qwen_image_2512_fp8_e4m3fn.safetensors appears in the supplied test context, but it is available for selection only if installed in a registered location. No explicit default is defined in the source code, and numerical minimum and maximum values do not apply.

ComfyArt Precise UNET Loader Output Parameters:

model

A MODEL object containing the loaded diffusion model configured to use comfy.ops.manual_cast as its custom operations implementation. This is a model object for downstream processing, not a generated image, a file path, or a measurement of model accuracy. You can connect it to compatible nodes that accept MODEL, such as model-processing nodes or a sampler’s model connection, as appropriate for your chosen architecture. Its importance is that subsequent model execution uses the operations selected during loading, avoiding the particular accelerated FP8 matrix-multiplication path targeted by this node. The node does not separately return a text encoder or VAE, so you must provide those elsewhere when your workflow requires them.

ComfyArt Precise UNET Loader Usage Tips:

• Use this node when working with an affected unscaled FP8 diffusion checkpoint, particularly when the normal loading path encounters problems with --fast fp8_matrix_mult. Its workaround is applied automatically; there is no additional switch to enable.
• Select a model installed in ComfyUI’s registered diffusion_models locations, commonly ComfyUI/models/diffusion_models. If a newly added file does not appear, refresh the interface or restart ComfyUI and check the dropdown again.
• To replace a standard “Load Diffusion Model” node, select the same model file here and reconnect its model output to the appropriate downstream connection. This node has no weight_dtype parameter, so do not carry that input over into an API workflow.
• Keep the text encoder, VAE, conditioning, and sampling setup appropriate for your selected model. Changing the loader does not make components from unrelated model families compatible.
• Do not assume that “Precise” means the checkpoint is converted to full precision or that it will always produce better images. The implementation selects manual-cast operations but does not explicitly set a weight data type.
• If you compare this loader with the standard loader, keep the checkpoint, prompt, seed, and sampling settings unchanged. Compare successful execution, image results, generation time, and memory use rather than assuming the manual-cast path is faster or uses less memory.

ComfyArt Precise UNET Loader Common Errors and Solutions:

Model file cannot be found

• Explanation: The node resolves the selected filename through ComfyUI’s registered diffusion_models locations and raises an error if it cannot find the file. This can happen when a workflow references a model that was moved, renamed, deleted, or never installed. The exact error wording depends on ComfyUI.
• Solution: Install the intended diffusion-model file in a registered model location, refresh the model list, and select the file again. If you run the workflow on another machine, confirm that the model is installed there too.

Value not in list: unet_name

• Explanation: This is a possible ComfyUI validation message when a saved workflow or API request specifies a filename that is not among the node’s current dropdown options. It occurs before the node can load the model.
• Solution: Choose an available entry from unet_name. If the intended file is missing, verify its location and spelling, refresh or restart ComfyUI, and update the workflow to use the listed filename, including any subfolder prefix.

ComfyArt Precise UNET Loader is missing or not registered

• Explanation: If ComfyUI reports an unknown node type or displays a missing custom node, the ComfyArt extension may not be installed or may have failed to import. The exact message varies by interface and ComfyUI version; model-file installation alone does not register this node.
• Solution: Confirm that ComfyArt is installed in the custom-nodes directory of the ComfyUI instance you are actually running, then restart ComfyUI. Check the startup console for import errors and resolve any reported dependency or version problems.

Unsupported or unrecognized diffusion model

• Explanation: The underlying diffusion-model loader may fail if your installed ComfyUI version does not recognize the checkpoint architecture or if the selected file is not a compatible standalone diffusion model. This is a possible loading failure rather than an error message defined by this node.
• Solution: Verify that you selected the intended diffusion-model file rather than a LoRA, VAE, or text encoder. Check the model’s requirements and update ComfyUI if that architecture requires a newer version. Manual-cast operations do not add support for otherwise unsupported architectures.

Error while deserializing header or another checkpoint-reading error

• Explanation: A checkpoint-reading error can indicate an incomplete download, a damaged file, or a file whose contents do not match its expected format. Such errors originate in the underlying file-loading libraries, so their exact wording varies.
• Solution: Compare the downloaded file’s size and checksum with the publisher’s information when available. Download it again from a trusted source if needed, and make sure the download has finished before selecting it.

CUDA out of memory

• Explanation: Loading or subsequently running the selected model may require more GPU memory than is available. This node changes the operations implementation; it does not guarantee reduced memory requirements or prevent memory exhaustion.
• Solution: Close other GPU-heavy applications and unload unused models. If the error occurs during generation rather than loading, try a smaller image size or batch size. If the model itself cannot fit, use a supported smaller model or ComfyUI’s documented memory-management options.

Downstream model or conditioning compatibility error

• Explanation: A model may load successfully but fail later when connected to conditioning, model patches, or sampling components intended for a different architecture. Possible symptoms include tensor-shape or matrix-dimension errors; the exact messages depend on the downstream operation.
• Solution: Start with a workflow documented for the exact model family you selected. Check its text encoders, VAE, conditioning nodes, and any model patches, and remove incompatible additions while troubleshooting. This loader addresses the specified FP8 operations path, not every source of workflow incompatibility.