Detailed directorial prompts vs reference image prompts for AI fashion generation — which works better?

Neither alone wins — a hybrid wins. Detailed directorial prompts control specifics (fabric behavior, light direction, pose at micro level); reference images lock look and identity (face, garment, palette, set). For editorial-grade AI fashion, use references as anchors and layer craft-specific directorial text on top — and decompose your references by attribute instead of stacking them as composites.

Use references to LOCK identity and look; use directorial prose to DIRECT behavior, light, and pose. Reference images carry what is hard to describe — a specific face, the exact weave of a fabric, the cut of a garment, the color palette of a location. Directorial text carries what a reference cannot encode: how the fabric should move in this shot, where the key light comes from, how the model's weight is distributed, what the gaze angle is. Each does work the other can't, and skipping either degrades the output.

The invideo agent is built around this split — it holds your reference sheets in project context and routes your directorial language to the right image and video model (GPT-Image-2, Nano Banana, Recraft, Seedance 2.0) per shot, so you write direction once and the references travel with it.

Where references win outright Face, garment, fabric weave, scale, and location consistency. For fashion specifically, upload close-up, front, side, back, and worn-on-person shots of each fabric — that gives the model enough to render the cloth's behavior accurately across environments. Cast faces without costumes first, lock them, then let the agent auto-combine those locked faces with lookbook wardrobe. For intricate categories like jewelry, references are non-negotiable: build the aesthetic with GPT-Image-2, then run Nano Banana on top to lock the exact piece into the frame.

Where directorial prompts win outright Fabric behavior, lighting direction, depth, and pose. "Single hard warm key light, side-raked, skin rim-lit, environment falling into shadow" produces editorial separation that no reference image alone will reliably reproduce. Pose direction must go micro: weight distribution, hand position, jaw tension, gaze angle in degrees. For materials, write a texture-language line per fabric — how it feels, its temperature, reflectivity, organic quality — and add a per-shot direction note describing how the fabric moves and interacts with the environment. Once that is in context, fabric behavior holds across shots without re-prompting it each time.

The hybrid that actually works: decomposed references + craft-specific text Stop stacking references as composites — the model blends them into mush, picks one wholesale, or ignores the rest. Instead, pull ONE attribute from each reference, conversationally: "pose from this one, lighting from this one, framing from this one, depth structure from this one." Then layer the directorial text on top — fabric behavior note, light direction, pose micro-direction, and the four-plane depth structure (sharp textured foreground 0–3ft, subject plane 3–8ft, prop midground 8–15ft, soft painted backdrop). The set does the depth-of-field work, not the lens.

Validate before you scale Before batching a campaign, generate one probe shot containing character, garment, set, skin, and pose all in frame. If it holds, generate the rest. If it doesn't, fix the brief — don't propagate the failure across 40 stills. Validate product consistency at three focal distances (close, mid, wide) before committing the run.

A worked example of the hybrid at scale One documented editorial campaign produced 40 stills and 30 motion clips across 2 models and 5 locations in 3–4 hours for ~$150 (630 credits, including every rejected generation). The spine was decomposed references + per-shot directorial prose with global rules baked in (skin realism, four-plane depth, lighting direction, posing). On a separate fashion film, framing required the most iteration; fabric behavior and texture were largely correct from initial generations once the texture-language and direction notes were locked in context — proof that directorial text earns its keep on the behavior layer while references hold the look.

Hridaye, invideo's creative director, puts the principle plainly: "Just the pose. Just the framing. Just the lighting. Or a combination from all your references. Entirely conversationally: 'I want the depth structure from this one.'"

Rule of thumb Start with references for look-and-feel (face, garment, palette, set). Layer directorial text for everything the reference can't encode (fabric behavior, light direction, micro-pose). Decompose references by attribute, never as composites. Probe one full-frame shot before batching.

Watch some of these to see what works for you:

Just the pose. Just the framing. Just the lighting. Or a combination from all your references. Entirely conversationally: 'I want the depth structure from this one.'

— Hridaye, invideo's creative director