Image-to-Video Prompt Expander
Turn a still into a video prompt the model can actually follow
- image
- positive_prompt
- negative_prompt
- image_description
- status
Image-to-video is the workflow most local video people actually run - anchor the first frame to a real image and the consistency problems of pure text-to-video mostly evaporate. The catch is that I2V models still need a prompt that describes what's in the frame and what should move. This node does both halves for you: a vision model reads your starting image, then a second LLM pass turns that description plus your motion direction into a prompt the video model can follow.
It's the pack's dedicated I2V node (Eric's Prompt Enhancers), and it's a genuinely nice fit for the Wan 2.2 and LTX workflows people are running - those models reward detailed, physical descriptions, and punishing short prompts is a known failure mode in the ecosystem.
How it works
Two stages, two LLM calls. First the vision pass: your image gets captioned by a vision-capable model - either a VL model loaded in LM Studio or Ollama (vision_backend), or the pack's local Qwen3-VL backend. Then the expansion pass takes that caption, your motion_description, and the aesthetic dropdowns, and writes a video prompt that keeps the subject recognizable while adding camera work and movement.
The one setting that changes the whole character of the output is reference_mode:
- recreate_exact - keep the exact character, outfit and lighting; just animate this scene. The default, and what you want 80% of the time.
- style_transfer - match lighting and mood only, invent a new pose or scene.
- character_remix - keep the character, change the outfit and setting.
- face_only - preserve the face, ignore everything else.
motion_description is your plain-language direction ("slowly turns head and smiles at camera"), and camera_movement, shot_size, motion_speed are the usual options, all defaulting to auto so you can ignore them.
The outputs
- positive_prompt - wire this into your I2V model's conditioning. It's the whole point.
- negative_prompt - auto-generated, usually worth keeping.
- image_description - the raw caption from the vision pass. Inspect this when results feel off; it's also handy on its own as a detailed prompt for other things.
- status - tells you whether the vision pass actually ran (more on that below).
The gotcha: vision is the fragile part
use_vision_model defaults to true, but the vision pass only works if your backend can actually see. With the default lm_studio backend, the model currently loaded in LM Studio must be a vision model - a plain Llama 3 will caption nothing, and you'll get an empty image_description and a weak prompt. Fix it either by loading a VL model in LM Studio, or switch vision_backend to qwen3_vl. That last one is the pack's self-contained path: it loads Qwen3-VL locally via transformers and needs no API server. Just know that the first run downloads a multi-gigabyte model from Hugging Face, and it needs the optional deps first:
pip install transformers>=4.42.0 accelerate>=0.30.0 huggingface_hub>=0.23.0 bitsandbytes>=0.43.0
The expansion pass can also use its own backend (expansion_backend), so you can pair a small vision model for captioning with a bigger LLM for writing.
Install
Shared with the pack: ComfyUI Manager → search "Local LLM Prompt Enhancer", or clone https://github.com/EricRollei/Local_LLM_Prompt_Enhancer.git into ComfyUI/custom_nodes/, pip install -r requirements.txt, restart. The core requirements file is just requests; the transformers stack above is only needed if you use the qwen3_vl backend.
One workflow note: if you're already on the Advanced video node, you don't need this one - the Advanced node's reference_image input does the same image-to-video pass with auto mode detection and more reference modes. This node is the simpler, more focused version. For most people, that's the right one.
Inputs (27)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | — | |
| motion_description | STRING | slowly turns head and smiles at camera | — |
| preset | COMBO | cinematic | 7 options: custom, cinematic, surreal, action, stylized, noir, +1 |
| expansion_tier | COMBO | enhanced | 5 options: auto, basic, enhanced, advanced, cinematic |
| reference_mode | COMBO | recreate_exact | recreate_exact: Keep exact character, outfit, lighting - animate this style_transfer: Match lighting/mood only, create new pose/scene character_remix: Use character but change outfit/setting face_only: Preserve face identity only, ignore background/lighting |
| use_vision_model | BOOLEAN | true | — |
| vision_backend | COMBO | lm_studio | lm_studio: Uses currently loaded model in LM Studio ollama: Uses currently loaded model in Ollama qwen3_vl: Auto-detects local Qwen3-VL model (no API server needed) |
| vision_endpoint | STRING | http://localhost:1234/v1 | lm_studio/ollama: API endpoint URL qwen3_vl: Leave default, or specify custom model path like 'A:\path\to\model' |
| expansion_backend | COMBO | lm_studio | lm_studio: Uses currently loaded model in LM Studio ollama: Uses currently loaded model in Ollama qwen3_vl: Auto-detects local Qwen3-VL model (no API server needed) |
| expansion_endpoint | STRING | http://localhost:1234/v1 | lm_studio/ollama: API endpoint URL qwen3_vl: Leave default, or specify custom model path like 'A:\path\to\model' |
| temperature | FLOAT | 0.70.1–2 | — |
| camera_movement | COMBO | auto | 15 options: auto, none, static shot, camera pushes in, dolly in, camera pulls back, +9 |
| shot_size | COMBO | auto | 8 options: auto, none, extreme close-up shot, close-up shot, medium close-up shot, medium shot, +2 |
| motion_speed | COMBO | auto | 6 options: auto, very slow, slow, normal, fast, very fast |
| lighting_type | COMBO | auto | 7 options: auto, none, soft lighting, hard lighting, edge lighting, rim lighting, +1 |
| time_of_day | COMBO | maintain current | 7 options: auto, none, maintain current, sunrise, daytime, sunset, +1 |
| positive_keywords | STRING | — | |
| negative_keywords | STRING | — | |
| save_to_file | BOOLEAN | false | — |
| filename_base | STRING | img2vid_prompt | — |
| wildcard_seedopt | INT | -1-1–2147483647 | Seed for wildcards: __name__, {a|b|c}, __objects[tag AND tag]__. -1 = fresh roll every run. 0 or more = reproducible picks. |
| wildcard_llm_blockopt | BOOLEAN | true | Tell the LLM which elements were picked at random so it keeps them: concepts keep their meaning, __=verbatim__ picks stay word for word. Off = the LLM only sees the resolved text. |
| wildcard_instructionsopt | STRING | Connect the llm_instructions output of a Wildcard Processor or Semantic Variation node here when the wildcards were resolved before this node. The LLM is then told which elements were picked at random, and verbatim picks are checked in its answer, exactly as if this node had resolved them itself. | |
| target_modelopt | COMBO | minimax_h3 | The video model the prompt is written for. wan22_video = this node's original behaviour, unchanged: Wan 2.2 tiers and shot structure. Any other model: the prompt is built from that model's profile (profiles/video/<id>.yaml) - its format, length, camera vocabulary, sound and dialogue rules - together with this node's preset, aesthetic controls and creativity settings. The tier then sets how long the prompt is inside the model's own range. MiniMax H3 prompts come out as its three named fields, not as a paragraph. |
| duration_secondsopt | FLOAT | 6.00–60 | Length of the clip you will generate. 0 = not stated. Ignored for wan22_video. Models with timed shots and sound (MiniMax H3, Seedance, Kling, Veo) need it: the timeline and the amount of action and dialogue are fitted to it. Set it to the same value as your sampler. |
| nsfwopt | COMBO | off | How far wildcard picks may go (a word match on each entry and on list names, no model). off = SFW: no suggestive, nude or explicit lists or entries suggestive = also lingerie, bikinis, sexy, seductive ... nudity = also nude, naked, topless ... explicit = everything, including lists under _nsfw/ Fix a wrong rating in _catalog_user.yaml (nsfw: on a list, ratings: on entries). Entries pairing minor-related and sexual terms are removed at every level (config.json "content_filter"). |
| violence_filteropt | COMBO | off | Hold back violent wildcard picks (a word match on each entry and on list names, no model). off = nothing held back no gore = blood and gore, torture, mutilation, corpses no gore or violent acts = also murder, execution, stabbing, strangling, suicide ... no violence or weapons = also guns, swords, knives, bombs, battle, war Only wildcard picks are filtered; text you type is left as is. |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| positive_prompt | STRING | — |
| negative_prompt | STRING | — |
| image_description | STRING | — |
| status | STRING | — |