video-workbench/investigations/2026-06-24-s14-reference-capable-generation/reviews/summary-and-rule-implications.md

Summary And Rule Implications

Status

• Document status: investigation conclusion with user decision recorded.
• Rule changes: promoted to AGENTS.md and VIDEO_WORKBENCH.md.
• User decision date: 2026-06-24.

User Decision Conclusion

This investigation established the working Video Workbench success channel for image-reference generation:

1. Generate the prompt/spec with the gpt-image-2 Skill Advisor workflow.
2. Load the real reference image into the conversation context before generation.
3. Use host image_gen with the loaded image assigned an explicit role such as Image 1.

This is the default channel for current and future Video Workbench image generation. API/Garden image generation is not the normal path.

Both reference strategies tested here worked:

• the full character anchor board preserved strong character consistency;
• the side crop also preserved strong character consistency;
• near / identity-heavy shots should prioritize crop references;
• distant / landscape-led shots should prioritize the accepted anchor board or project-approved distant reference.

For 在路上, most shots are landscape, aerial, or distant. The project does not require high-fidelity portrait likeness in most frames; it primarily needs character consistency. The current crop set, which was derived from the accepted character anchor board and made transparent, is sufficient by default.

For future projects with higher identity requirements and many close shots, create higher-resolution crop references through image-to-image from the accepted anchor board before production.

What This Investigation Proves

The host built-in image generation path can use an image made visible in the conversation context as a reference input for identity-sensitive generation.

This was not a pure text-to-image run:

• Arm A loaded the master anchor board before generation.
• Arm B loaded the side identity crop before generation.
• Both prompts explicitly assigned the loaded image as Image 1.
• Both outputs preserved concrete visual traits that were present in the loaded reference images: black rectangular glasses, dark messy hair, fuller beard / stubble, dark long coat, and backpack.

Boundary

This investigation does not establish an API/Garden workflow, because:

• gpt-image-2 Garden mode is not enabled;
• no OPENAI_API_KEY is present;
• the built-in image_gen tool schema exposes only a text prompt parameter.

The validated route should be named precisely:

host built-in reference-by-visible-context

Do not describe the default workflow as:

Garden image-to-image
explicit file-path upload
API multipart image edit

Operating Rule

For Video Workbench use, the built-in route is the default for reference-capable generation if the prompt workflow records the evidence:

1. load the exact reference image into the conversation context before generation;
2. state the image's role in the prompt as Image 1, Image 2, etc.;
3. record the input file path, output file path, and evidence caveat in generation-log.md or per-shot prompt/review;
4. do not count a run as reference-capable if the prompt merely names a local path without loading the image.

Reference selection defaults:

Shot type	Recommended primary reference
Near shot, face, half-profile, expression, identity-heavy frame	Matching crop reference first
Distant shot, landscape-led frame, aerial / wide scene, low identity burden	Accepted character anchor board or project-approved distant reference first
Medium action, back view, walking pose, prop continuity	Matching crop, transparent slice, or anchor board according to visible direction
High identity fidelity project with many close shots	Generate higher-resolution crop references through image-to-image from the accepted anchor board before production

Rule Promotion

Promoted to:

• AGENTS.md
• VIDEO_WORKBENCH.md

Project-specific registries may add narrower reference selection tables when a project needs them, but the global path and near/distant reference-choice defaults now live in the workspace rules.