---

name: prompt-conditioning description: "Configure Gemma/API prompt conditioning, prompt enhancement, saved conditioning artifacts, dynamic conditioning, and multimodal guiders for ComfyUI-LTXVideo workflows." metadata: disco-role: operating repo-skill-id: comfyui-ltxvideo sub-skill-id: prompt-conditioning node-families: - LTXVGemmaCLIPModelLoader - LTXVGemmaEnhancePrompt - GemmaAPITextEncode - LTXVPromptEnhancerLoader - LTXVPromptEnhancer - LTXVLoadConditioning - LTXVSaveConditioning - DynamicConditioning - GuiderParameters - MultimodalGuider disable-model-invocation: true license: NOASSERTION

Prompt Conditioning for ComfyUI-LTXVideo

Use this sub-skill when a task is about producing, enhancing, saving, loading, or tuning prompt conditioning for LTX-2 video/audio workflows in ComfyUI-LTXVideo.

The backend, model-folder, CUDA, and optional dependency baseline is owned by the root model and backend requirements. This sub-skill assumes ComfyUI-LTXVideo is installed and focuses on the conditioning side of the graph.

Route here when

• The workflow uses local Gemma text encoding through LTXVGemmaCLIPModelLoader.
• The workflow uses GemmaAPITextEncode instead of a local CLIP/Gemma encoder.
• The user wants to enhance raw prompts with LTXVGemmaEnhancePrompt or LTXVPromptEnhancer.
• The user wants to save/reuse CONDITIONING as safetensors files with LTXVSaveConditioning and LTXVLoadConditioning.
• The task involves DynamicConditioning, GuiderParameters, or MultimodalGuider setup for text/audio/video guidance.
• The error mentions missing Gemma config.json, processor/tokenizer files, API key/model-id metadata, empty conditioning, missing conditioning_data_*, duplicate guider modality, or prompt-enhancer downloads.

Route elsewhere

• Core model/latent/sampler/decode wiring, frame counts, tiling, looping, or low-VRAM loading: use ../core-generation/SKILL.md.
• IC-LoRA reference images, audio-only latent routing, HDR, masks, inpaint/outpaint, sparse tracks, or DubIt recipes: use ../specialized-workflows/SKILL.md.
• Expert STG/APG presets, Q8 patching, VAE patching, attention-bank/PAG/FETA/flow-edit tricks, or advanced sampler internals: use ../advanced-control/SKILL.md.

Operating graph

1. Choose the encoding path.
   • Local/offline path: LTXVGemmaCLIPModelLoader returns a CLIP object from a Gemma model folder under ComfyUI models/text_encoders and an LTX checkpoint. See Gemma conditioning.
   • Hosted/API path: GemmaAPITextEncode sends the prompt and model id to the LTX Video API and returns CONDITIONING directly. It requires credentials and an LTX checkpoint/diffusion-model file carrying the required metadata. See Gemma conditioning.
2. Optionally enhance prompt text before encoding.
   • Use LTXVGemmaEnhancePrompt when the local Gemma CLIP has processor files and you want Gemma-based T2V/I2V rewriting.
   • Use LTXVPromptEnhancerLoader + LTXVPromptEnhancer for a separate Hugging Face LLM/image-captioner prompt enhancer. It may download/cache models on first use. See prompt enhancement.
3. Create positive and negative conditioning.
   • API workflows usually use separate GemmaAPITextEncode nodes for positive and negative prompts.
   • Local Gemma workflows use the loaded CLIP wherever the graph expects CLIP text encoding or prompt enhancement before encoding.
4. Reuse conditioning when appropriate.
   • LTXVSaveConditioning writes a sanitized safetensors artifact under ComfyUI embeddings.
   • LTXVLoadConditioning reads that artifact back on CPU or GPU. Validate suspicious files with validate_conditioning_safetensors.py. See conditioning artifacts.
5. Patch or tune guidance only after conditioning exists.
   • DynamicConditioning patches the MODEL denoise mask behavior before sampling.
   • GuiderParameters builds one GUIDER_PARAMETERS entry per modality (VIDEO and/or AUDIO), and MultimodalGuider combines MODEL, positive/negative CONDITIONING, parameters, and skip_blocks into a sampler-ready GUIDER. See multimodal guiders.
6. Troubleshoot from the nearest symptom.
   • Use troubleshooting before changing graph families. Many failures are file placement, metadata, or device-selection problems rather than sampler problems.

Quick reference

Need	Primary node(s)	Key inputs	Output
Load local Gemma encoder	LTXVGemmaCLIPModelLoader	gemma_path, ltxv_path, max_length	CLIP
Gemma prompt rewrite	LTXVGemmaEnhancePrompt	clip, prompt, system_prompt, max_tokens, bypass_i2v, optional image, seed	enhanced prompt string
API prompt embedding	GemmaAPITextEncode	api_key, prompt, enhance_prompt, ckpt_name	CONDITIONING
Generic prompt enhancer	LTXVPromptEnhancerLoader, LTXVPromptEnhancer	Hugging Face model names, prompt, optional image_prompt	enhanced prompt string
Save/load conditioning	LTXVSaveConditioning, LTXVLoadConditioning	conditioning, filename, dtype; then file_name, device	safetensors file; CONDITIONING
Dynamic mask power	DynamicConditioning	model, power, only_first_frame	patched MODEL
Multimodal guidance	GuiderParameters, MultimodalGuider	modality CFG/STG/rescale controls, positive/negative conditioning, skip_blocks	GUIDER

Safety and side effects

• Do not run native ComfyUI generation just to answer prompt-conditioning questions; static graph inspection and the bundled validator are enough for file-structure checks.
• Do not put private filesystem paths, API keys, or local environment names in generated workflows or reports.
• Treat prompt enhancer and API nodes as network/credential surfaces. Ask before triggering downloads or sending user prompts to an external API.