Character reference sheets

Pipeline: a nude body render comes out of Blender, the model dresses that body, each garment is cut out, it goes to an image-to-3D generator, and the mesh comes back into Blender.

The goal of the generation is not a beautiful image. It is the same body as the render, now clothed, with every piece legible and cuttable. Everything below exists to protect that.

Cost rule, before anything else

Never call create_asset without asking first. Every generation spends the user's API credit. Write the full prompt, surface the decisions that change the image, and wait for an explicit go-ahead. This applies to each retry after a bad result too — do not silently fix and re-run.

Workflow

1. Measure the input renders with scripts/measure.py. If any criterion in section 8 of references/resolutions-and-proportions.md fails, say so and ask for a re-export from Blender. Cheaper than discovering the problem after the image is paid for.
2. Flatten alpha onto white if the PNG is RGBA. Alpha often becomes black on upload and contaminates the result.
3. Build the prompt using the architecture in references/prompt-template.md.
4. Ask for permission to generate.
5. Generate, one view per call.
6. Measure the output and compare against the input (thresholds below).
7. Copy from IMAGE_OUTPUT_DIR into the character project's own output folder, with a descriptive versioned name: character_outfit_front_v2.png.

Measurement dependency

scripts/measure.py requires Python and Pillow. Before first use, ask before installing the dependency, then run:

python -m pip install -r /path/to/character-reference-sheets/requirements.txt

If Pillow is unavailable and cannot be installed, do not generate: report that the mandatory pre-flight measurement could not be completed.

The rule that matters most

Never describe the body in the prompt. Not the build, not shoulder width, not head size, not height. Describing anatomy makes the model draw a new body from the text instead of preserving the one in the image.

Recorded case: a prompt opening with "heavyset bara-build, very broad shoulders, thick muscular arms, wide barrel chest, heavy thighs, small head" collapsed the character's arm span from 88.9% to 74.3% of the frame width — 14.6 points. Same character, same render, same model, rebuilt with the correct architecture and zero body description: 89.2%, a 0.3 point deviation.

The prompt says what to add and what to preserve. Never what the body is.

Parameters

Model	gpt-image-2
Aspect ratio	2:3 for a standing character · 1:1 for an isolated garment or prop
Quality	Provider default — the current create_asset client does not expose quality
input_fidelity	Provider default — the current client does not expose this option. OpenAI documents GPT Image 2 as supporting high-fidelity image inputs, but this workflow does not explicitly request high
References	2 to 4, base image first
Turnarounds	never in a single image — one view per call

Verification after generating

Run scripts/measure.py on the input and the output and compare horizontal occupancy (ocup_h):

ocup_h deviation	Reading
within ±2 points	normal, accept
beyond ±2 points	the prompt let the model redraw the body — regenerate

Empirical basis: crocodile character +1.6 / −1.4 across three generations; tiger v2 +0.3; tiger v1 (wrong prompt) −14.6.

Vertical occupancy usually rises 2 to 4 points even when everything is right. On its own it is not a failure signal.

Known residual: limbs get thinner

In every recorded generation so far — three of one character, two of another — arms and thighs came out with less volume than the base render. The PRESERVE block fixes framing and scale; it does not fix muscle mass.

Do not treat this as a new bug on each character. Practical consequences:

• Leg and foot pieces (pants, boots, sneakers) — safe to cut out, limb volume does not drive the garment's shape.
• Torso pieces (shirt, vest, jacket) — they were modelled over a narrower torso than the original mesh. The generated 3D tends to come out tight. Warn the user before they send it to an image-to-3D generator.

Where output lands

The server writes to IMAGE_OUTPUT_DIR (an MCP environment variable) and rejects any outputPath outside it. Pass a bare filename and copy the result into the character's folder afterwards.

References

• references/prompt-template.md — the prompt architecture, with a real example
• references/resolutions-and-proportions.md — aspect ratios, per-edge margins, Blender export settings, the mandatory pre-flight check, and what to feed an image-to-3D generator
• scripts/measure.py — bounding box, occupancy and margins