Open-Weight Video Models: Prompting Wan, Hunyuan and SkyReels (2026)
Open-weight video generation stopped being a hobbyist compromise. In 2026, running Wan, Hunyuan Video or SkyReels locally means real production leverage: no queue, no per-clip credits, reproducible seeds, LoRAs and ControlNet-style rigs on top, and ComfyUI workflows you can version like code. For teams shipping volume — Q4 stacks Halloween 10⁄31, Black Friday 11⁄27, Christmas 12⁄25 and New Year 1⁄1 — local inference is where batch economics live.
But there’s a catch that trips everyone migrating from hosted models: open-weight models speak different prompting dialects. The prompt that renders beautifully on one family falls flat on another — not because one is worse, but because motion priors and default aesthetics were baked differently. Copy-pasting one prompt across all three is the fastest way to conclude that “open models can’t follow instructions.”
They can — in their own dialects. This guide maps them: how the seven shot elements translate across the three families, which local constraints should shape your prompt text, and three templates ready for a ComfyUI queue.
Three families, three temperaments
A few test prompts reveal the differences faster than model cards:
- Wan rewards explicit, physically-described motion: verbs with direction and speed, clear subject-actor relationships, motion stated as mechanics.
- Hunyuan Video leans cinematic and prose-friendly: scene descriptions with atmosphere, light and mood hold together, and director-style writing (“a slow, mournful dolly past empty chairs”) beats telegraphic fragments.
- SkyReels is built around character performance and drama-style framing: expression beats, eyelines, reaction timing, shot/reverse-shot thinking. Prompts that read like a shooting script get the most out of it.
None of this is religious — test on your own checkpoint.
The seven shot elements across dialects
Same house framework, three translations:
| Element | Wan dialect | Hunyuan dialect | SkyReels dialect |
|---|---|---|---|
| Subject | Actor + physical attributes | Character + atmosphere | Character + emotion + eyeline |
| Framing | Functional: shot size, position in frame | Compositional: layers, depth, negative space | Dramatic: shot size chosen for the beat |
| Camera | Mechanical: “dolly left at 2m/s, tracking” | Cinematic: “a drifting handheld float” | Coverage language: “push in on the reaction” |
| Motion | Physics and mechanics of the action | Motion described with mood and rhythm | Gesture, expression, performance beats |
| Light & color | Practical lighting | Painterly: grade, palette, mood words | Key/fill for faces, emotional temperature |
| Timing | Action durations, phases | Pacing words: languid, sudden, suspended | Beat structure: line, pause, reaction |
| Sound | Not generated — foley in edit | Not generated — score mood in edit | Dialogue/ADR and room tone planned separately |
The row to watch is timing: open-weight models default to clip-length grids (seconds per generation, extendable in chains), so beats written inside the clip — “still one second, then the push begins” — matter more than on hosted models.
Local constraints should shape the prompt text
Running locally, the machine writes part of your prompt for you. Design around it:
- VRAM decides resolution and length. For 480p motion drafts upscaled later, write what survives low res: silhouette, motion, composition. Save texture claims (“engraved filigree”) for the high-res pass.
- Negative prompts are part of the dialect: blurry, warped hands, text artifacts, duplicated limbs earn their keep. Keep negatives short and physical.
- Seeds are a creative tool. Find one with good motion on your base prompt, then vary one element per run — controlled iteration over re-rolls.
- I2V beats T2V for control. Most families accept a start image (some end frames): feed your own key frame — a product shot, a Midjourney V8 or GPT-Image-2 still — and let the model own only the motion.
- Adapters are prompt shortcuts. A style LoRA changes your dialect: trigger words in, aesthetic adjectives out — read its page before writing style description it already handles.
- Chain clips deliberately. End-frame-to-start-frame stitching beats hoping one generation covers 20 seconds; write each clip with its own beats.
Same discipline for sources: reference clips teach camera grammar. Extract the shot structure, rebuild it in your own nouns — logos, likenesses, licensed music and scripts don’t travel.
Copy-paste templates
1. Wan-style mechanical motion (draft pass, 480p → upscale, 16:9)
Medium-wide shot, plain studio backdrop: a stunt rig drops a folded silk flag from the top of frame; the
flag falls under gravity for half a second, then snaps fully open as a crosswind hits from the left,
fabric rippling in three visible waves before settling into a slow hang. Camera dollies right at constant
speed (about 1.5 m/s equivalent), keeping the flag in the center third; a foreground stand casts a soft
moving shadow. Hard key light from the upper left, neutral color, deep background falloff. 5 seconds:
fall 0–0.5s, snap at 0.5s, ripple 0.5–3s, settle 3–5s. Sharp motion edges, no motion blur on the snap
frame. No text, no logos.
2. Hunyuan-style cinematic prose (atmosphere pass, 16:9)
A slow, mournful dolly past a row of empty wooden chairs in a shuttered seaside café at dawn. Cold blue
light seeps through salt-stained windows; dust hangs in the air and drifts through a single warm shaft
from a half-open door. The camera floats past the chairs at hip height, each one passing close enough to
read the grain and the chipped paint, before easing to a stop on a table where a single cup has been left
behind, steam long gone. Painterly realism, desaturated palette with one warm accent, gentle film grain,
soft shadows. 8 seconds, languid pacing, no cuts, no people. Sound plan: distant surf, one creaking chair,
sparse piano note at 6s.
3. SkyReels-style performance beat (character drama, 16:9)
Drama-style medium close-up in a dim hallway at night: a middle-aged stage actor in a charcoal cardigan
(original character, no resemblance to real actors) stands under a dying ceiling lamp, holding a sealed
envelope. Beat one: he looks down at the envelope, jaw tight, eyes glistening. Beat two: a long pause —
he glances toward the dark stairwell at frame left (eyeline shift), then back. Beat three: he exhales,
shoulders dropping, and tears the envelope open as the lamp flickers once. Camera pushes in slowly from
medium shot to close-up across the three beats, staying on his eyeline. Warm practical lamp against cool
window spill, shallow depth of field. 7 seconds, no dialogue — performance is carried by face and hands.
Sound plan: room tone, paper tear at 6s, lamp buzz throughout.
Before you queue the workflow: checklist
- [ ] Prompt drafted in the target model’s dialect (mechanical / cinematic / performance)
- [ ] Beats written inside the clip length your VRAM renders
- [ ] Fine-detail claims deferred to the high-res pass or an adapter
- [ ] Negative prompt short and physical (blur, warped limbs, text artifacts)
- [ ] Seed locked for iteration — vary one element per run
- [ ] Start image supplied where the family supports I2V
- [ ] Adapter trigger words used, redundant style adjectives removed
- [ ] Clip chaining planned: end frames match next starts, each clip has its own beats
Common mistakes and FAQ
“Same prompt, three models, three disasters — which model is broken?” None — you wrote one dialect and asked three. Split the prompt into seven elements and re-skin per family; the physics stays, the phrasing changes.
“My generations are inconsistent run to run.” Fix the seed, change one variable. If it’s still inconsistent, your prompt holds two competing motions — open models blend them into morphing.
“Can I just generate longer clips instead of chaining?” If your VRAM allows — but long generations drift. 4–8 second clips with clean beats, chained by matched frames, are more controllable than one 20-second drift and cheaper to re-run.
“How do I use reference videos legally here?” Same rule as anywhere: shot grammar is learnable, content isn’t. Extract a clip’s camera and timing structure with a video-to-prompt pass, rebuild it around your own subject — and if a contest demands originality (like Kling 4.0’s challenge), extraction notes document your process.
Open weights reward writing for the machine in front of you. Learn each dialect, keep your notes, and local video becomes the cheapest iteration loop you’ve ever had — start by pulling a reference clip apart with SocialToPrompt and rebuilding its beats in all three dialects.
Share Article