SkyReels-V2 (Diffusion Forcing / T2V / I2V)
An open video model family whose Diffusion Forcing (DF) variants generate effectively unbounded-length video by autoregressive extension rather than producing a single fixed clip.
- Resolution
- 720x1280
- Weights
- Downloadable
- API
- No
- Commercial use
- Vendor states commercial use is permitted
Sizes are 1.3B and 14B, at 540P (544x960, 97 frames) and 720P (720x1280, 121 frames) per generation window, at 24 fps. Separate T2V and I2V checkpoints are published.
Skywork Community License (Skywork 模型社区许可协议), not an OSI license.
Specification
- What it takes in
- text, image
- What it makes
- video
- Maximum duration
- not established
- Maximum resolution
- 720x1280
- Aspect ratios
- not established
- Controls
- text prompt, reference image, diffusion-forcing autoregressive extension, resolution selection (540P/720P), overlap history frames, AR step size, causal block size, prompt enhancer, seed
- Weights
- Downloadable
- API
- No
- Pricing
- Free open weights under the Skywork Community License, self-hosting only with no first-party API. The 1.3B variants target consumer GPUs while the 14B checkpoints need datacenter-class hardware; the repository does not publish a clean VRAM table, so no figure is asserted.
- Released or updated
- 2025-04-21
Maximum duration and supported aspect ratios are not established here. That means no page we could retrieve stated them, not that the model lacks them.
Strengths
- Diffusion Forcing enables genuinely long-form video rather than 5-second clips
- 24 fps native, smoother than the 16 fps Wan 2.1 lineage
- 1.3B options for lighter hardware
- Community license permits commercial use without an approval gate
- Both T2V and I2V plus a film-oriented prompt taxonomy
Weaknesses
- Non-OSI custom license with Chinese-language governing text
- Quality drift accumulates over long autoregressive generations
- VRAM requirements not clearly documented
- No first-party API
- Less community tooling than Wan or Hunyuan
Best for
- Long-form open-weights video (minutes rather than seconds) where a permissive-but-custom license is acceptable.
Where the sources disagree
Recorded rather than resolved. Picking the tidier number would hide the fact that the vendor's own pages do not agree.
- A 5B variant is listed as 'coming soon' in the repository and may or may not have shipped. Demonstrated long-form durations (10-60s) come from repository examples, not a stated specification.
What to understand before using it
Explainers naming the mechanisms this model's own documentation mentions.
See it from the vendor
Shotcraft hosts no sample output. Judging a model means watching its own showreel with your own eyes, and a clip re-hosted here would be a clip we chose.
Nearby in the directory
Adobe Firefly Video Model
The Adobe Firefly Video Model generates video from a text prompt, optionally guided by an image used as the first or final frame. Adobe's Firefly Serv
Video generation
Allegro
A small text-to-video model comprising a 2.8B VideoDiT and a 175M VideoVAE, released under Apache-2.0 in October 2024. It generates 88-frame clips at
Video generation
Amazon Nova Reel 1.1
Amazon Nova Reel generates video from text prompts or from text plus an image, and is available only through Amazon Bedrock. AWS documents output at 1
Video generation