Hunyuan Video Styler

Here's the first thing to know: the name is a lie. Hunyuan Video Styler never touches the Hunyuan model, doesn't need any Hunyuan weights, and would work just as well wired into a Wan or LTX workflow - or an image model, honestly. It's a prompt-building machine, a stack of hand-written template fragments that turn "a spaceship flying through an asteroid field" into "a spaceship flying through an asteroid field . epic space battle scene, dynamic starship movement, laser fire effects, dramatic space vista, high-speed pursuit, star-filled background . smooth tracking shot, stable camera movement, professional dolly motion, cinematic tracking" with one click. If you've ever spent twenty minutes googling "how do I describe a dolly zoom in a prompt," this is the shortcut.

How it works

No model, no API, no key, no GPU. Open the pack folder and you'll find a data/ directory of JSON files - one per category, each entry a tiny template with a prompt (containing a {prompt} placeholder) and a negative_prompt. The node loads them all at startup and applies whichever you pick, in a fixed order: movie scene, time of day, weather, lighting, shooting, camera, composition, effects, overall video style, cinematic color. Your base text replaces {prompt}, and each style's negative gets comma-joined onto whatever negative you passed in. That last bit matters - it's not just a positive-prompt prettifier, it hands you a matching negative too, which is nice because writing negatives for video is pure guesswork for most people.

The whole thing is pure Python stdlib. No requirements.txt, no model downloads, nothing to babysit. You can read every prompt it will ever emit in those JSON files, which is both the appeal and the ceiling: it's a friendly wrapper around a spreadsheet of cinematography vocabulary.

The inputs that matter

You set text_positive and text_negative (your base prompt - can be empty), then the ten dropdowns. Every one defaults to None, which means "skip," so a bare node just passes your text straight through untouched.

• movie_scenes - the fun one: matrix-lobby, inception-hallway, star-wars-space, blade-runner-2049, 2001-space-odyssey, a couple dozen more. These are fan-style scene descriptions, not LoRAs - they steer via words.
• camera, lighting, composition - your cinematography vocabulary: steady-tracking, low-angle-hero, dolly-zoom, butterfly-lighting, leading-lines, golden-ratio.
• timeofday, weather, effects, shooting, video_styles, cinematic_color - atmosphere, motion, and grade.

You can combine as many as you want; the output is chained in that fixed order. One quirk to know: the weather dropdown also carries the time-of-day list baked in (the pack ships timeofday_styles.json inside the weather folder too), so the same options appear in both menus - don't stack weather and timeofday with the same pick or you'll apply that text twice.

Outputs are two STRING sockets, text_positive and text_negative, which feed straight into a CLIPTextEncode (or the Hunyuan-specific equivalent) that goes into your sampler's conditioning. The pack's example workflow shows exactly this: styler → two CLIPTextEncode nodes → sampling.

Installing it

Search "HunyuanVideoStyler" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ShmuelRonen/ComfyUI-HunyuanVideoStyler

Restart ComfyUI. That's genuinely it - there's nothing else to download. The "Hunyuan Video" dependency everyone assumes is there doesn't exist.

Troubleshooting and honest takes

• The prompts get loaded at startup. Edit a JSON file while ComfyUI is running and it won't show until you restart.
• More dropdowns ≠ better. Pick two or three; stacking all ten produces prompt soup where the movie scene drowns in camera jargon. The README's own advice is right: scene first, one environmental layer, one camera layer.
• It appends, it doesn't rewrite. Your base prompt still does the heavy lifting. Start with a clear subject and action, treat the node as dressing, not the meal.
• debug_prompt (the boolean toggle) prints the composed prompt to your ComfyUI console after each stage - genuinely useful the first time, to see exactly what got appended and in what order.
• Lower your expectations for the model itself. Hunyuan Video is a 13B December-2024 first mover that wants 45–60GB VRAM and has been overtaken by Wan for most people; this node won't fix VRAM or motion quality, it only fixes prompt styling. A styler can't rescue a workflow the model can't run.

It's a niche, near-zero-traction pack (you won't find it talked about anywhere), and you shouldn't expect updates. But for what it does - translating "make it look like that movie" into usable cinematography language - it's fast, free, and inspectable.