Wan 2.7
Alibaba Tongyi Lab's open-weight video suite, the control-freak's and self-hoster's choice. T2V, I2V, R2V, and instruction-based editing share one 27B Mixture-of-Experts backbone, with first/last-frame control, multi-reference inputs, and native audio. Because the weights are open, the Wan family is the basis for uncensored community workflows.
The lineup
| Model | Job |
|---|---|
Wan2.7-t2v | Text-to-video with "Thinking Mode" prompt planning. |
Wan2.7-i2v | Image-to-video; first-frame, first+last-frame, audio-driven. |
Wan2.7-r2v | Reference-to-video: up to 5 real-person image inputs + vocal-timbre reference, multi-character. |
Wan2.7-videoedit | Instruction-based editing & style transfer on existing clips. |
Specs at a glance
Open weights = total control
The video, image, I2V, image-edit, and video-edit checkpoints are downloadable (HuggingFace Wan-AI, ModelScope) and run in ComfyUI. Inference code is on GitHub; training code/data are not released.
| GPU | T2V / I2V |
|---|---|
| RTX 4090 (24GB) | Yes, ~3–5 min/clip |
| RTX 5090 (32GB) | Yes, faster (~2–4 min) |
| RTX 4070 (12GB) | GGUF quant only |
| A100 (80GB) | Yes, fast |
Per Alibaba's pattern, open weights for a .7 release land 4–8 weeks after cloud launch. Until then, use cloud APIs (DashScope, WaveSpeed, Together, fal). For a no-hardware uncensored route today, Venice hosts a filter-free build as Wan 2.7 Uncensored (below).
Wan 2.7 Uncensored
Base Wan ships with safety filters. The open weights let those filters come off. Two practical routes:
1. Wan 2.7 Uncensored (hosted on Venice)
Venice hosts a filter-free build based on Wan 2.7 as Wan 2.7 Uncensored. It's tuned for anatomical accuracy, skin texture and lighting, and natural intimate motion. No GPU, no setup, no LoRA wrangling. Use it through your Venice API key like any other model.
- Your prompt gets rewritten for you. Behind the scenes Venice runs a private pre-processing pass that reads your input image and writes a prompt hyper-tuned to what this model handles best (the rewrite isn't stored). The upshot: hyper-specific prompt engineering is mostly redundant here. Put your effort into the input image, not into elaborate wording.
- The input image carries the result. The model can't easily add NSFW elements that aren't already present in the source. For mature output, start from a source image that already shows what you want, framed and lit the way you want it. This is the single biggest quality lever.
- It's weak at fluids of any kind. Don't try to prompt your way to convincing fluid dynamics; it won't get there. Choose subjects and motion that avoid it.
- Keep it simple in-frame. More than two people degrades fast, and animals are a known failure case. One or two subjects, clean composition.
What about text-to-video? On the Uncensored build, T2V is really I2V with one extra step: Venice generates a likely-good source image first, then runs it through the same image-driven pipeline. It's a convenience over base Wan 2.7 (which struggles with mature content), but you get more control by supplying your own source image and using I2V directly.
2. Local ComfyUI + NSFW LoRAs
Run Wan open weights locally and stack community NSFW LoRAs (e.g. all-in-one "rapid" merges with an optional unlocker LoRA, disabled by default). This gives maximum privacy and control. Nothing leaves your machine. The cost is setup and a capable GPU.
For most people the hosted Wan 2.7 Uncensored on Venice is the fastest path. Self-host only when full privacy or custom LoRAs matter.
What Wan 2.7 is best for
Optimal prompt pattern
Same director formula as the rest. Wan adds Thinking Mode (chain-of-thought planning of composition, placement, lighting before generation). Give it spatial detail to chew on, and it returns coherent multi-object scenes. Negative prompts are supported.
For mature content via Wan 2.7 Uncensored or local LoRAs, the tasteful technique that improves realism (soft directional lighting, slow deliberate motion, anatomy-aware phrasing, high-quality reference images) is covered in the NSFW best practices guide.
Content & licensing
- License: Apache 2.0. Commercial use permitted for the open weights.
- Filters: present in the base model; removed in Venice's Wan 2.7 Uncensored build and in self-hosted configurations.
- Responsibility: open weights mean no platform safety net. The consent rules are on you, and just as binding.