Original forced perspective and PTD helper studies

Native geometry
Blender 5.2.3 LTS Release Candidate build fab76e95306d
Cycles 32 samples with denoising Standard view transform Y world up
Two authored spheres and two cameras with a 35 mm lens and 36 mm sensor width
Camera depth of field disabled No captured photograph or AI-generated scene
blender --background --factory-startup --python-exit-code 1 --python build-forced-perspective-scene.py -- /absolute/new-output
The builder checks both objects remain inside both camera frames before rendering
It saves beauty images ID emission passes native geometry and projected vertex records
A different renderer build or platform can change pixels

Numerical and artwork assets
Place perspective-ptd-study.ts beside build-perspective-ptd-assets.ts
Install sharp@0.34.5 in that isolated folder
Use Bun 1.3.11
bun build-perspective-ptd-assets.ts /absolute/new-output
The output folder must already contain the two native beauty PNGs for the cover artwork
uv run --no-project --with numpy --with pillow python verify-perspective-native.py /absolute/new-output
The mask verifier checks saved silhouettes against native projected vertex bounds
Center-plane pinhole heights are a separate planning quantity from curved sphere bounds

PTD source helpers on CPU
Pinned small source file
https://huggingface.co/taobao-content-ai/MiniMax-H3-PTD-LoRAs/resolve/32d5dfb2b2cc4fc3f5dfef541bdcb23e30f9e2ec/minimax_h3_ptd.py
SHA256 96c52989cf81e52146beb604078fa01806f1a076771b848a37e57679623e4142
Place ptd-source-audit.json in the output folder
uv run --no-project --python 3.12 --with torch==2.9.1 --with numpy==2.3.5 --with safetensors==0.6.2 python verify-ptd-source.py /absolute/minimax_h3_ptd.py /absolute/new-output
The harness verifies the source hash before importing it
It executes pdd_time_grid pdd_sampling_plan and PerHopLoRALinear on original tiny CPU tensors
Four schedule cases compare against independent NumPy and the saved browser helper records
Twelve authored per-hop linear cases compare against independent arithmetic
Unloaded adapter behavior and hop clamping are also checked
These CPU library versions describe the helper test not a recommended full GPU installation

Boundaries
No full checkpoint weights downloaded No apply_ptd_lora call on a real model
No Diffusers generation pipeline ComfyUI queue or GPU inference run
No model quality speed peak memory or frame rate benchmark
Metadata examples are authored release-shaped JSON not extracted tensor headers
The planner restricts its scope to the documented four and eight hop release profiles
The stored block field can be stale so it is reported while a block is derived from intervals divided by hops