MiniMax-H3 Apply Text Embeddings:
MinimaxH3ApplyTextEmbeddings adds a saved MiniMax-H3 text embedding template to existing conditioning. It does this by placing the template’s tokens before the tokens already present, so you can combine a reusable embedding with conditioning from another part of your workflow. The template can be scaled to make its contribution stronger, weaker, or negative. When supported token tags are present, the node extends them to stay aligned with the added tokens. The embedding must have the format and width expected by this node; it does not convert an incompatible embedding.
MiniMax-H3 Apply Text Embeddings Input Parameters:
conditioning
The existing CONDITIONING data to modify. For each valid conditioning entry, the node adds embedding tokens before its current tokens and retains the entry’s metadata. It accepts conditioning tensors with either two dimensions or three dimensions; for three-dimensional tensors, it adds the same embedding tokens to each batch item. Entries that do not start with a tensor are passed through unchanged. This input has no minimum, maximum, default, or selectable options specified by the node.
embedding_name
Selects a file from the available embeddings filenames. The node reads that file as a safetensors embedding and expects a two-dimensional tensor named weight with a width of 5,120 values per token. The number of rows in that tensor determines how many tokens are added. Available options depend on the embedding files found in your environment; the source specifies no fixed option list or default. When strength is zero, the node returns the incoming conditioning without resolving or loading this file.
strength
Controls how much the selected embedding contributes by multiplying its values before they are added to the conditioning. The default is 1.0, which uses the saved values as they are; 0.0 bypasses the operation and returns the original conditioning; negative values reverse the sign of the embedding values. The allowed range is -100.0 to 100.0 in steps of 0.01. This setting scales the embedding values, not the number of tokens it adds when nonzero.
MiniMax-H3 Apply Text Embeddings Output Parameters:
conditioning
The resulting CONDITIONING data, ready to connect to a downstream node that accepts conditioning. At nonzero strength, each valid entry contains the scaled embedding tokens before its original tokens; existing metadata is retained, and any present minimax_token_tags or text_token_tags list or tensor receives a 1 for each added token at the front. Entries without a tensor in their first position pass through unchanged. At zero strength, this output is the original conditioning, returned without modification.
MiniMax-H3 Apply Text Embeddings Usage Tips:
- Choose an embedding file made for this node’s expected format: a safetensors file with a
weighttensor that has 5,120 values per token. - Start with the default
strengthof1.0to use the embedding’s saved values. Adjust it gradually to compare results;0.0provides a quick bypass without changing your connections. - Check that the conditioning comes from a compatible workflow. The node can add tokens to two- or three-dimensional tensors, but their channel width must also match the embedding’s 5,120 values for the tokens to be joined.
- If you chain several embedding applications, remember that each nonzero application adds more tokens at the front. The most recently applied embedding’s tokens will appear first.
MiniMax-H3 Apply Text Embeddings Common Errors and Solutions:
Expected key 'weight' in <path>. Found keys: <keys>
- Explanation: The selected safetensors file opened, but it does not contain a tensor named
weight. - Solution: Select an embedding file with a
weighttensor in the format this node expects.
Expected 2D tensor for 'weight' in <path>, got shape <shape>
- Explanation: The file’s
weighttensor does not have the required token-by-channel layout. - Solution: Use a compatible embedding whose
weighttensor has two dimensions.
Expected embedding width 5120 in <path>, got shape <shape>
- Explanation: The embedding has the wrong number of values per token for this node.
- Solution: Select a MiniMax-H3 embedding with a
weighttensor whose second dimension is5120.
Unsupported conditioning tensor rank <rank>. Expected 2D or 3D tensor.
- Explanation: A conditioning entry begins with a tensor that is neither two-dimensional nor three-dimensional, so the node cannot determine where to add the tokens.
- Solution: Supply conditioning in a supported two- or three-dimensional form from a compatible upstream node.
Sizes of tensors must match except in dimension ...
- Explanation: The saved embedding tokens and existing conditioning cannot be joined. One possible cause is a conditioning channel width that differs from the embedding’s width.
- Solution: Check that the upstream conditioning is compatible with 5,120-value embedding tokens; selecting a correctly formatted embedding alone will not fix incompatible conditioning.
Embedding file not found or cannot be read
- Explanation: The selected filename cannot be resolved in the
embeddingslocation, or its contents cannot be loaded as a safetensors file. The exact error text depends on the file and environment. - Solution: Confirm that the file still exists in an available embeddings location, select it again if needed, and use a valid safetensors embedding.
