Wan 2.7 Uncensored: Why negative_prompt Still Matters

Wan 3.0 shipped without a negative_prompt field, on every variant — plain and Spicy alike. Migrate your NSFW pipeline to it and you lose a tool you probably used on every generation: a dedicated place to tell the model what not to produce. Alibaba's older Wan 2.7 Uncensored family — alibaba/wan-2.7-t2v-uncensored, -i2v-uncensored, -ref2v-uncensored, all available through Siray's API — never dropped it. All three endpoints still take a real negative_prompt field.

Key Takeaways

• All three Wan 2.7 Uncensored endpoints (t2v, i2v, ref2v) support negative_prompt; Wan 3.0's full lineup does not.
• negative_prompt works best as a short, targeted exclusion list for one specific defect — not a catch-all blacklist.
• aspect_ratio only exists on the t2v endpoint, and it has no adaptive option — unlike Wan 3.0 Spicy's six-option enum.

What negative_prompt Still Does on These Three Endpoints

A negative prompt is an exclusion instruction, not a reliable blocklist — it lowers the odds of an unwanted element, it doesn't guarantee it stays out, and it works best as a short list targeted at one shot's actual risk rather than a sprawling catalog (Artlist). Where an API has no dedicated field, creators typically fall back to typing exclusions into the main prompt, which is weaker than a structured negative_prompt input (Filmora). Wan 2.7 Uncensored caps size at 720p or 1080p, duration at 2–15 whole seconds, and seed at -1 to 2147483647 — a narrower spec than Wan 3.0's 30-second ceiling, but with the field Wan 3.0 lost.

Excluding Unwanted Anatomy Artifacts

Diffusion video models still generate the occasional malformed hand, warped joint, or extra limb, especially in motion-heavy NSFW scenes. A negative_prompt targeted at the specific artifact you're seeing — not a generic "bad anatomy" catch-all — cuts its recurrence without dragging down other detail. Try negative_prompt: "extra fingers, fused fingers, warped hand, malformed limb" and iterate from what you actually saw.

Keeping ref2v References From Bleeding Into the Wrong Style

The ref2v endpoint accepts up to 3 reference images and 3 reference videos — fewer than Wan 3.0's ref2v (10 images, 5 videos). With a tighter reference set, an unwanted style or wardrobe element bleeding in from one reference is more noticeable, not less. "cartoon style, illustration, wrong outfit color, mismatched lighting" keeps the generation anchored to the intended reference rather than an average of all of them.

Blocking Text and Watermark-Like Overlays

Text-like artifacts — garbled captions, logo-shaped smudges, watermark ghosts — show up more often than most creators expect, particularly on i2v generations sourced from a still with a baked-in overlay. negative_prompt: "text, watermark, logo, caption, subtitle" is worth keeping in nearly every generation on this family.

Stopping the Duplicate-Subject Problem

A second, fainter copy of the main subject at the frame's edge is a known failure mode across video diffusion models, and it's one of the cases negative prompts handle reasonably well (CapCut). negative_prompt: "duplicate subject, cloned figure, extra person, mirrored double" addresses it directly — though it won't fix a duplicate caused by a bad reference image, since that's a source-material problem.

A Full Prompt + negative_prompt Pair for i2v

Here's a complete pair for the i2v endpoint, which also accepts image, end_image, and audio inputs:

Prompt: "A woman in a silk robe walks slowly toward the camera in a dimly lit bedroom, warm lamp light, natural movement, cinematic framing."

negative_prompt: "extra fingers, warped hand, text, watermark, duplicate subject, cartoon style, low resolution."Everything in the list maps to a real failure mode covered above — not padding. That's the difference between a negative prompt that earns its keep and one that just adds noise.

Developer Notes

Two gaps worth flagging first. aspect_ratio exists only on t2v with no adaptive option — pick a ratio explicitly if you're used to Wan 3.0 Spicy's default. And ref2v's 3/3 reference cap is real and smaller than Wan 3.0's; a scene needing more than 3 references isn't a fit. Pricing isn't independently verified against Siray's console for this family — check the model detail page's Pricing block before budgeting.

If your pipeline needs a real exclusion field and Wan 3.0 doesn't fit your migration, Wan 2.7 Uncensored is worth keeping for that reason alone. For a side-by-side on migrating anyway, see Wan 3.0 Spicy vs Wan 2.7 i2v Spicy: Should You Migrate; for the i2v-only 1080p workflow, see Wan 2.7 i2v Spicy: Uncensored 1080p Video From a Still Image.

Ready to test negative_prompt yourself? Create your free Siray account and try Wan 2.7 Uncensored alongside 300+ other models on one API.

A Note on Compliance

Siray's own NSFW and uncensored models generate legal, consensual adult content only — CSAM and any illegal content are refused with zero tolerance. "Uncensored" means no moderation beyond what the law requires, not that anything goes. Wan 2.7 Uncensored is a third-party Alibaba model available through Siray's API, not a Siray-hosted model, and the same standard applies to everything generated with it.