Wan is the open-source path. Wan 2.2 is the latest version with genuinely downloadable weights — a Mixture-of-Experts video model that runs on a single 24GB GPU — while "Wan 2.5/2.6/2.7" are closed, hosted-API versions on commercial platforms. If you want unlimited free generation and full control, Wan 2.2 in ComfyUI is the only real answer in 2026.
How to prompt Wan with JSON
Keep it leaner than Veo or Kling: lead with subject and action, keep the strongest three style modifiers, and drop the audio block entirely — open-weights Wan renders silent video. Our Wan tab trims the JSON accordingly. Structured prompts still outperform prose here, but Wan rewards brevity more than the hosted models do.
Open weights vs. hosted Wan
Hosted Wan 2.6 (Higgsfield, WaveSpeed) adds multi-shot, audio sync and reference-based character consistency — the fuller JSON from other tabs works there. Self-hosted 2.2 trades those features for zero marginal cost and no content-policy gatekeeping, which is why it dominates certain workflows.
Setup
Grab the weights from the official Wan-Video GitHub or the Wan-AI HuggingFace org and run through ComfyUI. Budget an afternoon for setup and expect slower iteration than any hosted model — the trade is cost, not convenience.
Prompting rules for Wan 2.2 / 2.6
Wan 2.2 is the latest version with genuinely open, downloadable weights. "Wan 2.5/2.6/2.7" are hosted API-only versions on commercial platforms.
Open-weights Wan 2.2 has no audio track — drop the audio field or move sound design to post.
Keep prompts shorter than for Veo or Kling: lead with subject and action, trim stylistic modifiers to the strongest three.
Wan 2.6 (hosted) adds multi-shot and audio sync — the fuller JSON works there.
Example template, Wan 2.2 / 2.6-shaped
The Wan 2.2 / 2.6 tab shows this site's block order for the model. Browse all 422 templates.
Google · duration 8s · 720p, 1080p, or 4K · audio native · verified 2026-07-10 · click any value to edit
{
"scene": "Downtown street at golden hour, soft haze",
"subject": "Young woman with brown hair and denim jacket",
"action": "Walks toward camera through light traffic",
"camera": {
"lens_mm": 35,
"movement": "dolly_in",
"movement_speed": "slow"
},
"style": "Cinematic, natural light, gentle contrast",
"audio": "Soft indie guitar",
"technical": {
"aspect_ratio": "16:9",
"seed": 137421,
"lock_identity": true,
"lock_style": true,
"negative_prompt": "extra people, text overlays, hands",
"duration_seconds": 8,
"fps": 24,
"creativity": 0.4
}
}
Veo 3.1 follows structured prompts unusually well — lock camera, lighting and audio as separate JSON fields and it respects each one.
Native audio: describe dialogue, ambient sound and music directly in the audio field.
Kuaishou · duration provider-dependent, commonly up to 15s · 1080p HD; native 4K modes advertised by Kling · audio native sound generation · verified 2026-07-10 · click any value to edit
{
"subject": "Young woman with brown hair and denim jacket",
"action": "Walks toward camera through light traffic",
"scene": "Downtown street at golden hour, soft haze",
"camera": {
"lens_mm": 35,
"movement": "dolly_in",
"movement_speed": "slow"
},
"style": "Cinematic, natural light, gentle contrast",
"audio": "Soft indie guitar",
"technical": {
"aspect_ratio": "16:9",
"seed": 137421,
"lock_identity": true,
"lock_style": true,
"negative_prompt": "extra people, text overlays, hands",
"duration_seconds": 8,
"fps": 24,
"creativity": 0.4
}
}
Kling 3.0 reasons about the scene like a director — write the action field as a shot note, not a keyword list.
Multi-shot is the signature feature: split the action into "Shot 1: … Shot 2: …" inside the action field for cuts within one generation.
ByteDance · duration provider-dependent · 720p/1080p; some hosts expose 4K tiers · audio native synchronized audio on supported hosts · verified 2026-07-10 · click any value to edit
{
"camera": {
"lens_mm": 35,
"movement": "dolly_in",
"movement_speed": "slow"
},
"subject": "Young woman with brown hair and denim jacket",
"action": "Walks toward camera through light traffic",
"scene": "Downtown street at golden hour, soft haze",
"style": "Cinematic, natural light, gentle contrast",
"audio": "Soft indie guitar",
"technical": {
"aspect_ratio": "16:9",
"seed": 137421,
"lock_identity": true,
"lock_style": true,
"negative_prompt": "extra people, text overlays, hands",
"duration_seconds": 8,
"fps": 24,
"creativity": 0.4
}
}
Seedance responds best to a strict block order: CAMERA → SUBJECT → ACTION → ENVIRONMENT → LIGHTING → STYLE. The JSON below is ordered to match.
Keep the camera block to a shot type plus one movement — stacking movements degrades output.
Alibaba · duration workflow-dependent · 720p open-weights TI2V · audio none in base text/image-to-video pipeline · verified 2026-07-10 · click any value to edit
{
"subject": "Young woman with brown hair and denim jacket",
"action": "Walks toward camera through light traffic",
"scene": "Downtown street at golden hour, soft haze",
"camera": {
"lens_mm": 35,
"movement": "dolly_in",
"movement_speed": "slow"
},
"style": "Cinematic, natural light, gentle contrast",
"technical": {
"aspect_ratio": "16:9",
"seed": 137421,
"lock_identity": true,
"lock_style": true,
"negative_prompt": "extra people, text overlays, hands",
"duration_seconds": 8,
"fps": 24,
"creativity": 0.4
}
}
Wan 2.2 is the latest version with genuinely open, downloadable weights. "Wan 2.5/2.6/2.7" are hosted API-only versions on commercial platforms.
Open-weights Wan 2.2 has no audio track — drop the audio field or move sound design to post.
Runway · duration 2-10s · 720p; aspect-ratio-specific pixel sizes · audio none in video generation · verified 2026-07-10 · click any value to edit
{
"subject": "Young woman with brown hair and denim jacket",
"action": "Walks toward camera through light traffic",
"camera": {
"lens_mm": 35,
"movement": "dolly_in",
"movement_speed": "slow"
},
"scene": "Downtown street at golden hour, soft haze",
"style": "Cinematic, natural light, gentle contrast",
"technical": {
"aspect_ratio": "16:9",
"seed": 137421,
"lock_identity": true,
"lock_style": true,
"negative_prompt": "extra people, text overlays, hands",
"duration_seconds": 8,
"fps": 24,
"creativity": 0.4
}
}
Runway exposes camera controls in the UI — use the camera block here as your settings checklist rather than prompt text.
Gen-4.5 excels with reference images: the subject field doubles as your reference-tagging description.