realistic-image-generationlisted

# Realistic Image Generation (keyframe + portrait craft) This is the **image-craft** layer. The Storyboard owns the *workflow* (a card's `keyframe` tier → its `clip` tier); `ai-asset-generation` owns the *mechanics* (call the model, save the file). This skill owns the *craft of making the image good* — because the image is the foundation of the whole video (FLF and i2v only animate the still you give them), so a weak or wrong keyframe poisons everything downstream. Produce ONE image well; the board decides when and how it becomes a clip. Use it whenever the goal is "looks like a real phone photo, not AI" — a fictional creator portrait (mandatory for `ugc-product-video` Stage 1), a character/product reference, or an FLF start/end keyframe. Phase 4 round 1 shipped a flux/dev portrait with a generic prompt and the user flagged it as "doesn't look real" — the fix is both the model and the template. ## Model picker (image gen for realism) **`openai/gpt-image-2` is the hardened default for any realism image — do NOT let `recommend_model` OR `search_models` downgrade it.** Image models are the foundation of the whole video (FLF and i2v only interpolate/animate the still you give them), so a weak or wrong keyframe poisons everything downstream. gpt-image-2 is the strongest at realism, prompt-adherence, and — critically — **correct anatomy** (hands, fingers). A weaker model botched the anatomy in a real run (a nail file rendered passing *through* a finger), which is exactly why th