Work in progress · image engineering

PhotoCraft: rules, orchestration and result validation

An image can look convincing and be wrong in its details. PhotoCraft develops execution rules, orchestration and checks that reveal those differences before work continues.

Illustration of a software orchestration workstation seen from behind: validation table, execution log and Git history on two monitors.
Rules, validation status and Git history: orchestration in focus. Illustration.
Share by email

The application: from a photograph to new lighting

We run photographs, Python scripts, geometry and Blender lighting across three local machines. PhotoCraft frames that chain: retain references, identify versions and let checks reject a result. This article presents one iteration of that work.

A photograph already contains reflections and shadows. Projecting it onto geometry and adding light can combine two lighting conditions. A partly hidden face can also receive a texture that appears twice when the camera turns. A better prompt cannot resolve this alone.

Concept diagram: checks and returning to a stage govern production.
Concept diagram: checks and returning to a stage govern production.

The harness: a verifiable execution framework

For this chain, the harness needs to set the agent’s tools, instructions and execution conditions. It also identifies the inputs, scripts, models and parameters of each run. Without these, an image can be produced; why it differs from the previous one becomes harder to explain.

A run identifier links source files, their hashes, settings and results. A script copied to another machine must be checked before comparing renders. Otherwise, a version difference may be mistaken for an algorithmic improvement. This is the reproducibility framework we are building around the work.

CPU judges — and tests for the judges

Here, CPU judge means a calculated, reproducible check separate from a language model’s opinion. In this iteration it measures pixel preservation at the reference angle, duplicates, unknown areas and text distortion. These are different observations and should not disappear into a single rating.

A model can suggest a defect or correction. To test the judge, we also give it a deliberately degraded mesh. If it barely distinguishes that case from the corrected mesh, the check is insufficient. Images, measurements and control cases establish the difference.

Measured progress, with a trade-off

In the same test at +10°, the internal duplicate indicator falls from 5.06% to 1.72%. Meanwhile, the area classified as unknown increases from 11.8% to 13.2%. Identical pixels at 0° remain at 99.71%. These figures concern this one test and its internal measurements; they are not a general benchmark.

The correction avoids projecting texture onto certain unseen surfaces. It reduces repetition but leaves more areas without information. That trade-off needs to remain visible rather than disappearing inside an overall score.

Yet the degraded-mesh control produces 1.79% duplicates, close to the corrected result’s 1.72%. The conclusion for this iteration is that this indicator alone is insufficient to authorise the next stage.

Internal measureBeforeAfter
Duplicates at +10°5.06%1.72%
Unknown area at +10°11.8%13.2%
Identical pixels at 0°99.71%99.71%

UVs, normals and relighting solve different problems

UV mapping establishes a correspondence between a surface and its texture. It does not reveal parts missing from the photograph. A normal map changes the response to light; it does not by itself create actual volume or a reliable silhouette from a new angle.

The ongoing work also examines depth, normal and appearance prediction using Diffusers’ Marigold pipelines. Available outputs depend on the chosen model. They remain estimates: a predicted material or suggested relief is not a product measurement.

The next step is to separate the object, its sensitive details and its lighting environment as far as possible. Lettering and functional features need separate checks. A generated background can support presentation; it must not be confused with a documented reconstruction of the object.

Three machines, one comparable experiment

Python and Blender execute the processing; local models can help with analysis and control. Across several machines, jobs need distributing without saturation and results need collecting under comparable execution conditions.

The pipeline needs a job queue, CPU, RAM and VRAM budgets per machine, time limits and controlled retries. Three machines’ memory does not automatically become a single memory pool. Low process priority alone does not cap CPU consumption.

Raw renders, lit renders and final results must remain separately comparable. Geometry, material and lighting should not all change at once when trying to isolate a defect’s cause.

ECC, Hermes, MCP and llama.cpp have distinct roles

ECC provides resources for structuring agent work, including skills, hooks and rules. Hermes Agent can connect to tool servers through MCP. MCP describes interaction with tools; it replaces neither render scheduling nor validation criteria.

llama.cpp supports local model execution, with CPU and GPU options depending on configuration. In the intended architecture, the model proposes actions, tools execute and checks determine whether a stage can continue. These components are integration reference points; listing them does not mean they are all already connected to PhotoCraft.

All local: a boundary to verify

The aim is to keep images and processing on local machines. This requires checking configured model providers, initial downloads, telemetry, remote tools and any fallback providers. An MCP server or an installed model alone does not guarantee the absence of external traffic.

We follow emerging tools and try them on this project. A new component needs to improve execution, comparison or control of a run. That is what we seek with PhotoCraft: being able to explain an image’s transformations and recognise when the result no longer supports the next step.

Technical references