Short answer: artifacts are not random. Each one is the model failing at a specific thing for a specific reason: hands fail because they are small, complex and underrepresented; limbs merge because two bodies confuse the model’s sense of who owns what; faces drift in video because identity is re-inferred every frame; skin goes plastic because the model averages toward smoothness. Because the causes are specific, so are the fixes. Diagnose from the table, then read the entry.

The quick table

Hands and fingers

The classic. Hands are the most complex articulated object the model ever draws, they occupy a few hundred pixels, and photo datasets show them in thousands of half-occluded positions. The model has strong opinions about skin and weak ones about finger count.

Fixes, in order of effort: keep hands away from the camera (composition, not surgery); put `extra fingers, fused fingers, deformed hands` in the negative prompt on SDXL-family models; on Flux and Illustrious, which have the best hands, often just regenerate; inpaint the hand alone at higher resolution — the single most reliable repair; in video, a clean hand in the reference frame survives far better than one the model invents mid-clip.

Extra and merged limbs

One body is easy; two bodies in contact is where the model must decide which arm belongs to whom, with no skeleton to consult. Every contact-scene failure — the third arm, the leg from nowhere, the merged torso — is that decision going wrong.

Fixes: prefer models with strong contact physics (this is Wan 2.7‘s specialty); one act LoRA at a time — two act adapters is the most reliable way to grow a limb (Wan 2.2 mods); simplify the pose in the reference frame, letting the prompt add complexity; on Pony-family checkpoints, precise position tags beat prose descriptions (Pony guide).

Faces: sameface and video drift

Two distinct problems. Sameface — every generation looks like cousins — is a checkpoint trained on too narrow a face set; switch checkpoints or add specific facial tags. Video drift — she is someone slightly else by second nine — is the i2v model re-inferring identity each frame from limited evidence; the smaller and softer the face in the reference, the faster it drifts.

Fixes for drift: larger, sharper face in the frame (enhance first — Krea workflow); 5–8 second clips, extended rather than generated long; medium shots over extreme close-ups on Seedance; for series work, reference conditioning or a character LoRA so every frame starts from the same identity anchor.

Skin

Plastic skin is averaging: thousands of retouched photos teach the model that skin is smooth, and explicit checkpoints over-trained on a narrow aesthetic make it worse. The fix stack: a realism merge rather than a base model; prompt for texture (`detailed skin, skin pores, natural skin`); CFG a notch lower than you think; enhancement at low creativity — high creativity is how you get the porcelain look back. Video inherits whatever the reference frame’s skin looks like: fix it before animating, not after.

Text, seams and the rest

Garbled text is structural — diffusion models paint letter-like shapes. Crop text out of reference frames (it also becomes animated garbage in video); add real text in post. Inpainting halos come from the blend zone: mask generously, feather, and run a light second pass over the seam. Uncanny lighting is two light logics colliding; name one source in the prompt (“single window light from the left”) and use reference frames with a clear shadow direction — the reference frame guide covers the five properties that matter.

A diagnostic habit

When an output is wrong, name which artifact class you are looking at before regenerating. Rerolling the same prompt against a structural cause wastes generations; each class above has a lever, and pulling the right one usually fixes it in one attempt. That habit — cause, lever, retry — is most of what separates creators who ship consistent sets from people posting hand-horror to Reddit.

Related in the Lab

• NSFW image-to-video: preparing reference frames

• Stable Diffusion for NSFW: SDXL vs Pony vs Illustrious

• Wan 2.7 NSFW: open weights, cost, uncensored use

Key takeaways

• Every artifact has a specific cause: identify the class first, then pull its lever.

• Hands: composition, negatives, inpaint. Limbs: one LoRA, simpler pose, contact-strong models.

• Video drift: sharper faces in the frame, shorter clips, extend don’t lengthen.

• Skin: realism merges, texture prompts, low-creativity enhancement.

• Text never works; crop it and add it in post.

FAQ

Why does AI porn give people extra fingers?

Hands are small, complex and inconsistently represented in training data, so the model’s confidence about finger count is genuinely low. Composition, negative prompts and inpainting are the fixes.

Why does the face change during AI videos?

Image-to-video models re-infer identity every frame; a small or soft face in the reference frame gives too little signal, and errors accumulate with duration.

How do I make AI skin look real?

Use a realism checkpoint, prompt for skin texture, lower the guidance a notch, and enhance at low creativity.

What is the best model for two people in contact?

Wan 2.7 leads on contact physics among the models available uncensored; use one act LoRA at most.

Last updated: August 2026.