How do you maintain consistent fabric rendering across multiple fabric types — silk, denim, and wool — in the same AI video?

Hold each fabric consistent across shots by writing a Texture Language entry per material (silk, denim, wool — feel, sheen, weight, drape), loading it into the invideo agent's context once alongside a lookbook with close-up + front/side/back/on-body references per garment, then adding a fabric-movement direction note to every shot in your shooting script. The agent then renders each material correctly across every interaction without per-shot prompting.

invideo is an agentic video tool that holds project context across every generation and routes shots to whichever video model fits best — so the fabric rules you set once apply to every clip you make in that project.

1. Write a Texture Language entry per fabric. For silk, denim, and wool separately, describe in plain words how each one behaves: silk as liquid, anisotropic sheen with soft drape; raw denim as matte, stiff, visible warp threads, structured fold; wool as heavy, fuzzy surface, subsurface softness, slow swing. Hridaye, invideo's creative director, puts it directly: "the texture language is basically for every material in the film, I'm describing how it feels in words. Doing this allows the agent to generate the fabric in the way that you actually want it to behave." Paste all three entries into the invideo agent's context tab — that tab is the project brain and the descriptions persist across every image and clip.

2. Upload a lookbook with five reference angles per garment. For each fabric type, give the agent a close-up of the weave, a front shot, a side shot, a back shot, and one shot worn on a person. The close-up teaches weave and reflectivity, the on-body shot teaches drape against a real frame. Variety matters more than volume — silk needs the close-up to capture specular behavior; denim needs the side angle to show structured fold; wool needs the on-body shot to show weight.

3. Add a per-shot fabric-movement note in the shooting script. In the shooting script PDF you upload alongside the lookbook, every shot gets one line on how the fabric moves and interacts with the environment — "silk blouse catches sidelight as she turns", "denim holds its fold as he sits", "wool coat swings heavy in the wind". This is the line that prevents per-shot re-prompting: the agent reads the note, knows which material is in frame from the lookbook, and renders the interaction correctly. "The most important for this PDF was the direction note per shot, which describes how the fabric is moving and how it's interacting with the environment."

4. Lock framing on a still before spending video credits. For any shot where multiple fabrics share the frame, generate the keyframe image first, iterate cheaply until the silk highlight, denim weave, and wool surface all read correctly side by side, then animate that locked frame in Seedance 2.0. Spending video credits only on locked frames is what kept one documented fashion production at ~$600 total for two ads with fabric consistent across every shot. In a different production, framing took the most iteration — fabric behavior itself held from early generations once the texture language and lookbook were loaded.

5. Evaluate every generated clip on three fabric checks. Before locking any clip, run it against: weave accuracy (do silk's specular highlights, denim's warp, wool's fuzz read correctly at this distance?), color accuracy (does each material hold its tone under the shot's lighting?), and garment-body-environment interaction (does the silk move like silk against skin, the denim hold its structure on the leg, the wool swing with its real weight?). Reject anything that fails any of the three — across documented productions, ~85% of generated clips get rejected, and the cost numbers above already include those rejects.

6. Route models by what each fabric needs. Different shots benefit from different generators: Seedance 2.0 reference-to-video carries the locked keyframe forward and is strong for sustained drape and motion across a single shot; Kling 3.0 handles multi-shot sequences natively where you need the same wool coat across cuts. The invideo agent routes each shot to the right model automatically based on the keyframe and the shot note — you don't pick a platform per material because every roster model is available inside invideo.

7. Add a Standing Don'ts section to your context. List the failure modes that wreck fabric specifically: "no plastic-looking silk", "no painted-on denim with missing seam structure", "no flat wool that reads like felt", "no fabric melting between shots". This stays in context and acts as a guardrail on every generation.

Done once at setup, this workflow holds silk, denim, and wool consistent across every shot in the same video regardless of sunlight, wind, touch, or motion — without prompting fabric behavior per clip.

Watch some of these to see what works for you:

The texture language is basically for every material in the film, I'm describing how it feels in words. Doing this allows the agent to generate the fabric in the way that you actually want it to behave.

— Hridaye, invideo's creative director