How I Fixed My AI Ghost Portraits After 60 Bad Tries

Cinematic close-up portrait of Simon Ghost Riley from Call of Duty Modern Warfare II, knitted cotton balaclava substrate named first, weathered white dried acrylic paint layer cracked and lifting at the weave, black vertical paint stripes bleeding down from eye sockets and jawline into individual fabric fibers, amber-brown eyes with visible sclera and iris plus one hard catchlight through the mask openings, black tactical headset with boom microphone, plate carrier vest with MOLLE webbing and radio handset at right shoulder, dark green volumetric smoke wrapping neck and shoulders as a rim-lit modifier rather than ambient fog, Rembrandt lighting with warm 3200K key at 45 degrees upper left and cool 5600K green edge light from frame left, shallow depth of field at f/1.8, photorealistic skin texture at eye corners, visible fabric weave in balaclava shadow areas, shot on 35mm anamorphic lens with oval bokeh isolated to smoke highlights, teal-amber color grade with crushed blacks, MWII promotional key art style --ar 2:3 --style raw --s 250

The failure is easy to recognize once you have seen it twice. The balaclava comes back as a smooth, evenly lit shell with a skull graphic stamped flat across it. The crackle pattern resolves into uniform procedural noise rather than paint that has dried, shrunk, and split along the fabric grain. The vertical stripes read as clean painted lines instead of pigment that has bled into individual knit fibers. Worst of all, the eye openings become solid black voids with no sclera, no iris, and no catchlight, so the face reads as a prop rather than a person.

That single symptom — graphic instead of material — accounts for most of the wasted renders on this subject. The cause is rarely the model version. It is phrase ordering. A prompt that opens with "weathered white skull-patterned balaclava" hands the tokenizer a noun phrase where "patterned" statistically dominates as a print. Diffusion models resolve ambiguity toward the most common reading. Naming the substrate first, then the paint layer, then the cracking, forces the material branch instead of the graphic branch.

What the Skull Element Actually Is to the Model

The mask is not one object. It is a stack of three surfaces, and each one needs its own texture vocabulary and its own lighting response. The base is knitted cotton, which shows weave only at grazing angles and in shadow areas. On top sits dried acrylic paint, which micro-cracks, lifts slightly at the edges, and holds a sharper specular than the fabric beneath it. Behind the openings are the only true specular surfaces in the frame: wet eyes. Treating the mask as a single "textured" material is what collapses all three into one flat value.

The amber-brown eye requirement is doing more work than it appears. Eyes are the only place in this composition where the renderer can produce a believable catchlight, and that catchlight is what convinces a viewer there is a living person under the cloth. Without an explicit iris color and the words "visible sclera," the eye opening gets filled with the same dark value as the surrounding paint. The face flattens immediately, no matter how good the crackle detail is. This is the same trap that sinks masked character portraits generally, from the Joker portrait workflow to the Deadpool mask build: the mask fails the moment it is described as a graphic rather than a worn object.

Ordering inside the prompt matters as much as vocabulary. Substrate, then coating, then damage, then the body underneath. Reversing that order — starting from the dramatic cracked look and working backward to the fabric — reliably produces the sticker effect, because the model commits to the surface finish before it knows what the finish is sitting on.

How Lighting Cues Decide Whether the Paint Reads as Cracked

The Rembrandt setup with a warm 3200K key at 45 degrees upper left and a cool 5600K green edge light from frame left is not decoration. The temperature split is the mechanism. Cracked paint holds a slightly higher specular response than matte knit, so a warm key striking at a grazing angle catches the raised crack edges while the weave absorbs and scatters the same light. Without that contrast pair, both surfaces return the same mid-tone and the crackle pattern becomes invisible at normal viewing size.

Volumetric smoke behaves as a light modifier, not an atmosphere preset. Described as wrapping the neck and shoulders, dark green smoke picks up the 5600K edge light and creates rim separation between the mask and the plate carrier. Described as ambient fog, it lowers global contrast and kills the crushed blacks the grade depends on. The wording "wrapping around neck and shoulders" is doing technical work: it constrains the volume to a band where it can catch light instead of veiling the whole frame.

Shallow depth of field at f/1.8 carries a trade-off worth naming. The far shoulder, the radio handset, and most of the MOLLE webbing fall out of focus. That is correct for promotional key art, where the eyes must be the sharpest element, but it means the gear detail you specified will not be legible. Either accept soft gear with sharp eyes, or stop down to f/2.8 and lose the anamorphic bokeh character that gives the frame its cinematic signature.

Edge Cases: Anamorphic Bokeh, Headset Geometry, and Model Drift

Oval bokeh from a 35mm anamorphic lens only appears where there are out-of-focus highlights. If the prompt requests anamorphic rendering without specifying that the bokeh should land in the smoke highlights and the defocused radio, the model applies the effect to nothing visible and the lens signature disappears. Anchoring the bokeh to a named out-of-focus element is what makes the anamorphic read survive.

Boom microphone placement is a known mirroring problem. The prompt asks for the mic left of the mouth; models frequently flip it, because left-right orientation is weakly encoded in text-to-image training. If the composition depends on that side, state it as "boom microphone on the subject's right side of frame" and verify in the render. Otherwise accept the mirror rather than burning generations on it.

Stylization interacts with this specific failure mode. At --s 250 the model keeps photographic structure while still pushing contrast and color. Above roughly 400, the aesthetic prior reasserts itself and the graphic-mask failure returns even with correct phrase ordering, because stylization amplifies the dominant interpretation rather than the literal one. The Midjourney documentation covers parameter ranges, but the interaction between stylization and material ambiguity is something you have to observe per version.

That last point is the honest limitation. Parameter behavior drifts between model versions, and a prompt tuned on one release can shift the balance of paint versus print on the next. The same text sent to a different engine will not honor --ar or --style raw at all, so the layer-ordering logic transfers but the parameter values do not.

Checklist Before You Accept the Render

One: zoom to the eye openings. You need visible sclera, a defined iris in amber-brown, and at least one hard catchlight. Solid black voids mean the mask has already failed regardless of how good the crackle looks.

Two: check the crack edges against the fabric grain. Real dried paint splits along knit direction and lifts at the borders. Uniform, evenly distributed noise across the whole mask is a procedural tell, not texture.

Three: confirm the vertical stripes bleed into fibers rather than stopping at a clean painted edge. Sharp-edged stripes read as printed decal; feathered, uneven ones read as pigment absorbed into cotton.

Four: verify the smoke is rim-lit and localized to the neck and shoulders. If it has flattened the blacks across the frame, the volume was read as fog and the contrast grade is gone.

Five: check where the bokeh actually falls. If nothing in the frame is softly oval, the anamorphic cue was dropped and the lens character needs to be re-anchored to a specific defocused element.

Six: confirm the boom microphone is on the intended side and that the headset sits symmetrically on the skull. Asymmetric ear cups are the most common geometry artifact on this subject.

Once those six hold, the frame is usable as key art. The next step is pushing the same layer-ordering logic onto other masked and costumed subjects, where the same graphic-versus-material failure repeats with different surface vocabulary.