Capstone: a computational photography pipeline and provenance policy

This capstone brings the course together in two linked deliverables. First, you design a computational photography pipeline for a defined use, specifying capture strategy, alignment and merging, demosaicing, tone mapping, any AI enhancement, and output, with justified parameters, a compute budget and a test plan. Second, you write an AI-use and provenance policy for a photography practice that would use such tools, covering which interventions are allowed for which kinds of work, disclosure wording, content credentials, client data and likeness, and how to respond to questions about authenticity. The two deliverables connect because every technical choice in a pipeline, from merging frames to AI upscaling, has consequences for what you can honestly claim about the resulting images. This lesson provides a worked example, a runnable miniature pipeline, templates, a checklist and an assessment rubric.

The brief

Choose one of the following scenarios, or propose your own with your instructor's approval.

• Scenario A, night events: an event photography studio wants a phone-based or camera-based workflow for hand-held candid images at evening receptions, with natural-looking colour and no visible ghosting of dancing guests.
• Scenario B, product e-commerce: a small studio photographs hundreds of products weekly and wants consistent, clean images with backgrounds extended for different web layouts, delivered quickly.
• Scenario C, documentary field work: a journalist working in low light needs maximum image quality while preserving a defensible chain of provenance and avoiding invented detail.
• Scenario D, landscape fine art: a photographer wants high-resolution prints from hand-held captures with wide dynamic range and is open to creative use of AI.

Each scenario pulls the design in different directions. Night events need robust merging and motion-aware frame lengths (lessons 3 and 6). Product work benefits from consistency, generative expand and careful disclosure (lessons 9 and 10). Documentary work must stay at the measurement end of the spectrum and prioritize provenance (lessons 1 and 11). Fine art can embrace super-resolution, HDR and even generative techniques, honestly described (lessons 4, 7 and 9).

Part 1: the pipeline specification

Your pipeline specification should cover the following sections, each with a short justification referring to the course.

• Use case and constraints: subjects, light levels, motion, output sizes, delivery times and device.
• Capture: single frame, burst or brackets; frame count and exposure per frame as a function of light and motion; raw or processed.
• Alignment and merging: global or tile-based, robust weighting, handling of moving subjects.
• Raw processing: noise model, demosaicing method, white balance strategy, colour correction.
• Tone mapping: global or local, HDR output with gain maps or SDR only, target look.
• AI enhancement: denoising, sharpening, upscaling or generative steps, with settings and where each sits on the intervention ladder.
• Output and provenance: formats, colour spaces, content credentials, metadata and archiving.
• Compute and storage budget: data volumes and processing time per image.
• Test plan: scenes, metrics and acceptance criteria.

A worked example: Scenario A, night events

The following condensed specification shows the expected level of detail.

Use case: hand-held candids at receptions lit by warm tungsten and LED uplighting, about EV 3 to 5, guests dancing; deliver 300 to 600 edited images within a week, printed up to 8 by 10 inches and shared online. Device: a full-frame mirrorless camera with a fast 35 mm lens, plus a recent phone for backup and social clips.

Capture: with the camera, shoot single frames at 1/125 s to freeze dancing, f/1.8, ISO 6,400 to 12,800, raw. A burst would freeze motion only if frames were very short, and robust merging would reject moving dancers anyway, so for this subject a single short frame plus strong AI denoising is the better design. For static details such as table settings and the venue, use a 6 to 8 frame hand-held burst at 1/30 s and merge, or the camera's multi-shot noise reduction mode.

Raw processing and AI: apply a raw-domain learned denoiser at moderate strength, checking faces and fabric at 100 per cent for invented texture; set white balance per scene with a reference shot of a grey card under the dominant light, keeping some warmth so the reception looks like evening; use AI subject masks to lift faces modestly. These are levels 2 and 3 on the ladder. No generative additions; minor cleanup such as exit signs only with the client's agreement in the contract.

Output: sRGB JPEG for online galleries, high-quality JPEG or TIFF for print, with copyright and contact metadata; content credentials enabled in the editing software's export; raw originals archived with two copies, one off site.

A miniature pipeline you can run

The following program strings together simplified versions of the techniques from lessons 2 to 4 and 8 on synthetic data: it captures a noisy burst of a scene with a moving object, aligns nothing (the camera is assumed steady), merges robustly, tone maps with the Reinhard operator, and reports noise and ghosting at each stage. Use it as a starting point for experimenting with parameters in your specification.

import numpy as np
rng = np.random.default_rng(420)
h, w, n = 60, 90, 8
radiance = np.full((h, w), 0.05); radiance[:, 60:] = 2.0        # dark room, bright window (40:1)
gain, read = 200.0, 2.0                                          # electrons per unit, read noise
def frame(k):
    r = radiance.copy()
    r[20:40, 10 + 4 * k:16 + 4 * k] = 0.6                        # a moving person
    e = rng.poisson(r * gain) + rng.normal(0, read, r.shape)
    return np.clip(e / gain, 0, 1.2)                             # sensor clips at 1.2 units
burst = np.array([frame(k) for k in range(n)])
ref = burst[0]
sigma = np.sqrt(np.maximum(ref, 0.01) / gain + (read / gain) ** 2)
w8 = np.clip(1 - (np.abs(burst - ref) - 2 * sigma) / (2 * sigma), 0, 1)
merged = (w8 * burst).sum(0) / w8.sum(0)
def reinhard(x, key=0.18):
    lw = np.exp(np.mean(np.log(x + 1e-4))); L = key * x / lw
    return L * (1 + L / L.max() ** 2) / (1 + L)
room = (slice(45, 60), slice(0, 50))
print(f"room noise: single frame {ref[room].std():.4f}, merged {merged[room].std():.4f}")
path = (slice(20, 40), slice(20, 44))
print(f"ghosting on the person's path: plain mean {burst.mean(0)[path].mean() - 0.05:+.3f}, robust {merged[path].mean() - 0.05:+.3f}")
out = reinhard(merged)
print(f"window clipped in capture: {np.mean(ref[:, 60:] >= 1.2) * 100:.0f} % of window pixels")
print(f"display values: room {out[room].mean():.3f}, window {out[:, 65:].mean():.3f}, max {out.max():.3f}")

room noise: single frame 0.0181, merged 0.0076
ghosting on the person's path: plain mean +0.098, robust -0.001
window clipped in capture: 100 % of window pixels
display values: room 0.058, window 1.000, max 1.000

The output shows the merged room is more than twice as clean as a single frame and robust merging suppresses the moving person's ghost. Tone mapping, however, leaves the room very dark and pushes the window to pure white. That reveals the limitation: the window is clipped in every frame, because all frames share one exposure. A real design would underexpose to protect the window and lift the room in tone mapping, or add a short frame, which is exactly the kind of reasoning your specification should document. Try changing the gain, the number of frames or the clipping level and observe the effects.

Test plan and acceptance criteria

A pipeline is only as good as its testing. Your test plan should include standard targets and real scenes, quantitative metrics and visual checks, and explicit pass criteria.

Part 2: the AI-use and provenance policy

Your policy should be written for clients and the public, in plain language, and should be specific enough to guide decisions. A strong policy includes the following sections.

• Scope: which services and image types it covers.
• Our approach: a short statement of values, for example that photographs are presented as truthful records of real events unless clearly labelled otherwise.
• What we do: the ladder levels used by service, such as AI culling, denoising and masking for all work; cleanup for portraits on request; generative expand for product layouts with approval.
• What we do not do: for example, no generative alteration of documentary images; no changes to people's bodies or faces beyond agreed retouching; no fully synthetic images presented as photographs; no deepfakes of real people.
• Disclosure: how generative changes are labelled in captions, delivery notes and metadata, and that content credentials are attached where supported.
• Client images and data: which AI services are used, whether uploads may be used for training, storage locations, biometric features and consent, and deletion on request.
• Likeness and consent: releases for people in images, and no training of models on clients or subjects without written consent.
• Authenticity questions: how the studio responds if an image's authenticity is questioned, including providing originals and provenance.
• Review: how often the policy is updated as tools and laws change.

Presenting and defending the work

Present your pipeline specification and policy as a short report of about 2,000 to 3,000 words with diagrams, plus your test results. Be ready to answer questions such as: where on the measurement-interpretation-invention spectrum does each stage sit; what would change if the light fell by two stops; how would you respond if a client asked you to remove a person from a documentary image; what happens to your content credentials when the client posts the image to social media; and which parts of your images might not be protected by copyright. Strong answers connect technical and ethical reasoning rather than treating them separately.

Looking ahead

Computational photography and generative AI will keep changing faster than any course can. New sensors will capture more information per exposure; pipelines will increasingly be single neural networks from raw to finished image; generative models will become part of capture itself, filling gaps in real time; and provenance will either become routine or remain patchy, depending on choices made by manufacturers, platforms and lawmakers. The concepts in this course, photon budgets, alignment, robust merging, tone mapping, depth, sampling, the perception-distortion trade-off, diffusion and provenance, will remain the tools for understanding whatever comes next. The habit that matters most is honesty about what a photograph measured, what it interpreted and what it invented.