Which AI image model produces the most realistic skin texture with pores and imperfections for film character portraits?

For film character portraits with visible pores, fine lines, and stubble, Recraft V4 is the model to use inside the invideo agent — it generates faces with the micro-imperfections that read as real skin. Nano Banana Pro handles the full character sheet around that portrait. Run both in parallel through a casting sub-agent, lock four options, then feed into Seedance 2.0 for video.

Use Recraft V4 for the hero face at 4K. It renders the small stuff — pores, lines, redness, stubble — that other models smooth away, which is why AI portraits usually read plastic. Hridaye, invideo's creative director, put it directly: "ReCraft actually gives you those imperfections like pores, lines, stubble, like all the little stuff that makes a face look like an actual face." That texture is what survives the jump from still portrait to video and makes the character read as a person rather than a render.

Run a parallel model casting comparison through a casting agent. Spin up a casting sub-agent inside the invideo agent and instruct it to run the same character prompt through Recraft V4 and Nano Banana Pro simultaneously. Recraft wins on skin authenticity; Nano Banana Pro has stronger prompt adherence and is the right call for the 360-degree turnaround sheet (front, side, profile, back, plus face and mid-angle closeups, 4 angles at 4K). You pick the aesthetic per asset rather than forcing one model to do both jobs.

Write the prompt for skin, not for a face. Specify visible pores, fine lines, skin texture, no smoothing, plus a real camera and lens descriptor (e.g. 85mm, shallow depth of field) and the practical lighting source. Generic "photorealistic portrait" prompts get you the over-smoothed default; explicit micro-texture language is what pulls Recraft toward the pore-level detail.

Lock four options, then stop generating. Generate four portrait variations per character, pick the strongest, and log it as the canonical reference in the invideo agent's context. This is the asset-lock step that prevents the face from drifting once you move to character sheets and Seedance 2.0 video. In one 70-second short, the team locked characters this way with no LoRA needed; in a longer production, locking one character took roughly five generations at about $9.78 in credits.

Add a post-generation pass to kill the plastic feel. Even with Recraft, AI faces benefit from a finishing pass — light grain, a touch of blur, and a grade that breaks the synthetic sharpness. Hridaye describes the same logic for Seedance 2.0 video output ("there tends to be this ultra-sharpness, there's this very plasticky feeling on the skin"), and the fix carries over to stills. Dedicated skin-texture tools like Enhancor, Claid, and Upsampler exist for this exact step if the portrait still feels too clean.

Why this matters for the film, not just the image. The portrait isn't the deliverable — the character sheet is, and the character sheet anchors every Seedance 2.0 clip downstream. A face with real pore detail upstream gives every generated shot something specific to hold onto; a smoothed face upstream propagates that plastic look across every frame of the film. The invideo agent holds all the models — Recraft, Nano Banana Pro, Seedance 2.0, Veo, Kling — so the routing happens inside one workflow without switching platforms.

These are some of the ways to solve skin realism for a film portrait — what works depends on the character, the lighting, and how the face will move in the final shot.

Watch some of these to see what works for you:

ReCraft actually gives you those imperfections like pores, lines, stubble, like all the little stuff that makes a face look like an actual face.

— Hridaye, invideo's creative director