The Real-Skin Problem: Why AI Faces Drift Toward CGI
Every generation hop is a re-interpretation of the last one -- and photoreal skin texture is the first casualty of the chain.
The chain that quietly renders itself
A common way to build a hero shot looks reasonable on paper: generate a stylized reference sheet for the character, generate a specific key visual from that reference sheet, then animate the key visual into video. Three steps, each one feeding the next. The final clip comes back and something is off -- the skin looks rendered. Not obviously fake, not cartoonish, just... plastic. CGI-adjacent in a way that's hard to name but easy to see.
The cause isn't any single bad step. It's the chain itself.
Why re-interpretation compounds
Every generation step in that chain is not looking at a photograph. It's looking at the OUTPUT of the previous step -- itself already a generation, already an interpretation of whatever came before it. The reference sheet interprets a real photo (if one existed at all) into a stylized illustration. The key visual interprets that stylized illustration into a specific pose and composition. The animation interprets the key visual into motion. By the time video comes out the other end, the skin has been re-interpreted twice before the video model even sees a frame of it.
Fine texture -- pores, subsurface light scattering, the thousand tiny imperfections that read as "real skin" to a human eye instead of "rendered skin" -- is exactly the kind of detail that degrades first under repeated re-interpretation. Each hop smooths a little more of it away, because each generation step is fundamentally a re-guess of structure, and texture is the highest-frequency, lowest-priority information in that guess. Three hops in, the guess has compounded into something that reads as CGI even though every individual step looked fine in isolation.
The fix: get one hop from a real photograph
The contrast case is a generation that's only one hop removed from an actual photograph -- anchored directly on a trained identity model (see "Soul Cinema: Training an Identity Model for Exact Lookalikes") rather than routed through a stylized bible or key-visual chain first. That single hop holds real skin texture dramatically better than the same generation arrived at through three stylized intermediate steps, because the trained identity is built directly from real reference photos, not from a re-interpretation of a re-interpretation.
The practical rule: when a shot's hero frame absolutely needs to hold photoreal skin, anchor it directly on the trained identity. Save the stylized bible and key-visual workflow for shots where a more illustrated, storyboard-driven look is actually the goal -- not for every hero frame by default.
Three caveats to run every time
Anchoring directly on a trained identity is not a free lunch. It comes with three caveats worth running as a checklist before every hero-frame generation.
Mask secondary faces. An identity model maps the trained face onto every human face it detects in a frame -- not just the character you intended. Any secondary character sharing a shot with the hero must be masked, kept faceless, or turned away from camera. Skip this and the extra in the background comes out wearing the hero's face too.
Use image-references sparingly. Passing a composition-guide image alongside the identity to control layout can visibly distort held props -- a weapon or object can warp, because the model is reconciling two competing sources of structural information at once. Prefer clean prompt-only composition guidance for structural elements like held objects, and reach for an image-reference only when it's genuinely necessary.
Reinforce real skin explicitly in the prompt. Real working phrasing that measurably helps: "real human skin, pores, natural texture, not CGI, not rendered, not plastic, shot on a real camera." This is not decorative language -- adding it to a hero-frame prompt anchored on a trained identity noticeably improves how much texture survives the generation.
Spotting the drift before it ships
The failure is easiest to catch early, before it's baked into an expensive animation pass. Pull the still frames at each hop of a chain -- the bible, the key visual, the first frame of the animated clip -- and look at the same patch of skin (a cheekbone, a jawline) across all three. If it visibly smooths, flattens, or gains a waxy highlight from hop to hop, the chain is already drifting and the video model will only inherit and extend that drift, not fix it. Catching it at the key-visual stage costs you a re-generation of one still. Catching it after the animation pass costs you the whole render.
The same test applies in reverse when you anchor directly on a trained identity: the hero frame should visibly hold more texture than anything produced through the stylized chain, because it's a genuinely shorter path back to a real photograph. If it doesn't, check the two caveats first -- a stray image-reference fighting the identity, or prompt language that dropped the explicit real-skin reinforcement -- before assuming the identity itself is the problem.
What this buys you
None of this is about avoiding stylized reference sheets and key visuals entirely -- they're genuinely useful for locking composition, wardrobe, and blocking before committing to an expensive animation pass. It's about knowing which shots need to hold photoreal skin and routing those specific hero frames through the shortest possible path back to a real photograph, while masking anyone who shouldn't inherit the trained face and reinforcing the texture language explicitly in every prompt that matters.