Seedance 2.5 Spicy vs Wan 3.0 Spicy: Multi-Character Fix

A single reference image works fine for one person. Ask the same setup to hold two consistent faces in the same uncensored clip, and it usually falls apart — one face blends into the other, an outfit jumps between characters, or the second person just gets rendered as a copy of the first.

Key Takeaways

• One reference image anchors one identity. A two-person prompt gives the model nothing to bind the second face to, so it drifts or duplicates the first.
• Holding two faces steady requires multi-image reference endpoints — ref2v variants that accept a separate image per character, not a shared one.
• On Siray, bytedance/seedance-2.5-ref2v-spicy and alibaba/wan-3.0-ref2v-spicy both do this, and both now cap out at the same 1080p ceiling — this isn't a "pick the higher-spec one" decision anymore.
• Tag each reference image explicitly in the prompt ("the woman from image 1," "the man from image 2") so identity binds to position instead of blending.

Why Single-Face Methods Break

Most uncensored i2v workflows pass one image field as the reference and one prompt describing the scene. That single image is the only identity anchor the model has. Feed it a prompt describing two people, and there's no second anchor to hold — the model either reuses the first face for both characters or blends features from the prompt description into something that resembles neither reference.

This isn't a bug so much as a structural limit: one image field can only carry one identity. Two-character continuity needs a second slot, and single-image i2v doesn't have one.

What Actually Holds Two Faces

The fix is a ref2v endpoint that takes more than one reference image, with one image assigned per character. On Siray, two uncensored options do this:

Both endpoints tie at a 1080p ceiling — that's worth stating plainly because it's a recent change. Seedance 2.5 Spicy's ref2v variant also carries an audio reference slot (1-10 clips) that its non-Spicy counterpart doesn't have, which matters if the scene needs a locked voice alongside a locked face.

The technique itself is simple: give each character's reference image its own tag in the prompt instead of describing "two people" generically.

image1: reference photo of character A
image2 and 3: reference photo of background
video1: reference video
prompt: "
The man from @image1 is walking naturally in the scene.  
Use @image2 and @image3 as the background environment.  
Match the motion, camera style, and overall feel from the video reference @video1."

Naming the images by position — image1, image2 — gives the model something concrete to bind each face to, instead of leaving it to infer who's who from adjectives alone.

Developer Note

Both endpoints sit behind the same request shape as any other Siray video call — swap the model string between bytedance/seedance-2.5-ref2v-spicy and alibaba/wan-3.0-ref2v-spicy and the images array works the same way. One API key, no separate integration for either.

Create your free Siray account and start testing multi-character consistency with ref2v before you burn generations on a single-image approach that can't hold the second face.

If a scene only needs one consistent character, our character-consistency guide for Seedance 2.5 covers the simpler single-face case. For locking a character across multiple separate clips rather than within one scene, see our cross-endpoint character workflow.

Compliance: Siray is committed to responsible AI use. All content generated through the platform must comply with applicable laws. Siray maintains zero tolerance for CSAM and any illegal content. Uncensored generation capabilities are provided exclusively for lawful adult content creation by and for consenting adults aged 18 and over.