TextGenerateLTX2Prompt - ComfyUI Built-in Node Documentation

The TextGenerateLTX2Prompt node expands a short user prompt into a detailed, audio-visual description suitable for generating video with the LTX-2 series of video models. It automatically adds task-specific system instructions, sends the formatted prompt to a language model, and returns the enhanced text. When an optional reference image is supplied, the node switches to image-to-video mode and expands the prompt starting from that image’s content.

Inputs

Common Inputs

Parameter	Description	Data Type	Required	Range
`clip`	The CLIP model used for text encoding. The node checks the model’s tokenizer name to select the matching instructions: Gemma 4 based models use the LTX-2.4 format, while other models use the LTX-2 (Gemma 3) format.	CLIP	Yes	-
`prompt`	The raw text input describing the scene or concept to be expanded into a detailed video generation prompt.	STRING	Yes	-
`image`	An optional input image used as the first frame of the video. When provided, the node switches to image-to-video mode and uses a system prompt that expands the user prompt based on the image’s content.	IMAGE	No	-
`video`	An optional video input used as additional context. Passed to the language model as an image batch; assumed to be 24 FPS and subsampled to 1 FPS internally.	IMAGE	No	-
`audio`	An optional audio input that can be used as additional context for generation.	AUDIO	No	-
`max_length`	The maximum number of tokens the language model is allowed to generate (default: 512).	INT	Yes	1 to 32768
`sampling_mode`	Controls whether random sampling is used during text generation. When set to `"on"`, the sampling parameters below become available; with `"off"` the node generates text without random sampling.	DYNAMIC_COMBO	Yes	`"on"` `"off"`
`thinking`	When enabled, the model is instructed to reason before answering. Any reasoning block is removed from the returned output (default: False).	BOOLEAN	No	True/False
`use_default_template`	When enabled, the node uses the default chat template for formatting (default: True). Advanced setting.	BOOLEAN	No	True/False
`mtp`	Speculative decoding with the checkpoint’s multi-token-prediction head. Has no effect without MTP weights. `"auto"` adapts the draft depth, `"2"` to `"5"` pin it. Sampled output stays correctly distributed but differs from non-MTP output for the same seed (default: `"auto"`).	COMBO	No	`"auto"` `"off"` `"2"` `"3"` `"4"` `"5"`

Sampling Parameters (when `sampling_mode` is “on”)

Parameter	Description	Data Type	Required	Range
`temperature`	Controls the randomness of the output. Lower values make the output more predictable, higher values make it more creative (default: 0.7).	FLOAT	Yes	0.01 to 2.0
`top_k`	Limits the sampling pool to the top K most likely next tokens. A value of 0 disables this filter (default: 64).	INT	Yes	0 to 1000
`top_p`	Uses nucleus sampling: it keeps the smallest set of most likely tokens whose cumulative probability reaches this value. (default: 0.95)	FLOAT	Yes	0.0 to 1.0
`min_p`	Sets a minimum probability threshold for tokens to be considered (default: 0.05).	FLOAT	Yes	0.0 to 1.0
`repetition_penalty`	Penalizes tokens that have already been generated to reduce repetition. A value of 1.0 applies no penalty (default: 1.05).	FLOAT	Yes	0.0 to 5.0
`seed`	A number used to initialize the random number generator for reproducible results (default: 0).	INT	Yes	0 to 18446744073709551615
`presence_penalty`	Penalizes new tokens based on whether they have appeared in the text so far, encouraging the model to talk about new topics (default: 0.0).	FLOAT	No	0.0 to 5.0

Note: The sampling parameters above are only active and visible in the node interface when sampling_mode is set to “on”. When it is set to “off”, no sampling parameters are available and the node generates text without random sampling. Note: The behavior of the node changes based on its inputs:

If an image is provided, the generated prompt is formatted for an image-to-video task using a system prompt that describes how to expand the prompt based on the image’s content. If no image is provided, the formatting is for a text-to-video task using a system prompt that expands the prompt into a detailed video generation description.
If the CLIP tokenizer’s name contains “gemma4”, the node uses the LTX-2.4 system prompts and the Gemma 4 chat format. Otherwise, it uses the LTX-2 (Gemma 3) system prompts and chat format.
When thinking is enabled with a Gemma 4 model, the model is opened on its reasoning channel; when disabled, the model is opened directly on the final-answer channel. For non-Gemma 4 models, thinking is passed through to the underlying generation step.
If the language model produces no usable text after removing reasoning blocks, the node returns the original prompt instead.

Outputs

Output Name	Description	Data Type
`generated_text`	The enhanced video-generation prompt produced by the language model, with any reasoning block removed. If the result is empty, the original user prompt is returned.	STRING

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): 1da4a388b7c358e5649b4746b9b8d288977ec6fbed3eedc4c8709187c9f7b943

​Inputs

​Common Inputs

​Sampling Parameters (when sampling_mode is “on”)

​Outputs

Inputs

Common Inputs

Sampling Parameters (when `sampling_mode` is “on”)

Outputs