conjure3d
/conjure3d
Tworzy teksturowany model 3D w skali rzeczywistej z jednego zdjęcia lub opisu tekstowego na komputerze z kartą NVIDIA.
- Autor
- memgrafter ↗
- Repozytorium
- conjure3d ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 03.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 73
- Skrypty
- 39
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 03.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 39 skryptów w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: conjure3d.py, deploy/modal/app.py, deploy/vast/onstart.sh, imagegen/gen.py, imagegen/setup.sh, mac/convert_weights.py, mac/run_mac.py, mac/setup_mac.sh
Co robi i kiedy
Tworzy teksturowany model 3D w skali rzeczywistej z jednego zdjęcia lub opisu tekstowego na komputerze z kartą NVIDIA.
Wyzwalacze z opisu
Oryginalny opis (EN)
Make a textured, real-scale 3D object (GLB) from one reference image, or from a text prompt through the optional image stage, on a Linux machine with an NVIDIA GPU where this repository is already set up. Use it when the task is "make a 3D model / GLB / prop of X at N metres" or "re-make asset X and check it is identical".
Jak działa
- When to use
- Steps: "make a GLB of a brass lantern, 0.35 m"
- Check after
- Re-make or verify an asset
- Report to the user
Podgląd SKILL.md
---
name: conjure3d
description: Make a textured, real-scale 3D object (GLB) from one reference image, or from a text prompt through the optional image stage, on a Linux machine with an NVIDIA GPU where this repository is already set up. Use it when the task is "make a 3D model / GLB / prop of X at N metres" or "re-make asset X and check it is identical".
---
# conjure3d in a session
Setup, flags, formats and env: [AGENTS.md](AGENTS.md). This page is the workflow.
## When to use
- One object per image. Not for scenes, characters with rigs, or animation.
- You know the object's real size (longest side, metres).
- The machine has the venv and weights from `trellis2/setup.sh all`. If `$T2_VENV/bin/python` or `$T2_WEIGHTS` is
missing, stop and say that setup is needed (it builds CUDA code and downloads large weights; ask before you start it).
## Steps: "make a GLB of a brass lantern, 0.35 m"
1. `source config.env`. Check the GPU is free: `nvidia-smi`. If others share it, wrap GPU commands in
`bash tools/gpu_guard.sh <gpu> "<what>" ...`.
2. Get a reference image: one object, whole, on a plain background, nothing cut off at the frame edge.
- The user gave one: copy it to `images/brass-lantern.png`.
- No image: write a job for the image stage (format in `imagegen/gen.py`), e.g. prompt "a brass oil lantern,
whole object, plain grey background, studio light", 2 to 4 seeds, `w` 1024, `h` 1024. Run
`HF_HUB_OFFLINE=1 $IMAGEGEN_VENV/bin/python -I imagegen/gen.py serve` with a `STOP` file in the queue. Look at
the images and pick one; show them to the user if the choice matters.
3. Write a props file, e.g. `work/props.json`:
`{"props": {"brass-lantern": {"gold": "brass-lantern.png", "size_m": 0.35}}}`.
Lying objects (a watch, a knife): add `"rest_up": [0, 0, 1]` so the image's front faces up.
4. Make the 3D object. Several seeds give very different shapes; for a pick, run 2 to 3 seeds into separate folders:
`HF_HUB_OFFLINE=1 ATTN_BACKEND=xformers PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True $T2_VENV/bin/python -I
trellis2/run.py --job work/props.json:images:work/out-s42 --seed 42`
5. `$T2_VENV/bin/python -I post/post_step.py --props work/props.json --dir work/out-s42`
6. Optional web files: `web/export_web.sh` (see AGENTS.md).
## Check after
- `work/out-s42/brass-lantern/cut.png`: the whole object is in the cut-out, with no background left. If not, use
another image.
- `asset.json`: `checks.passed` is true; `bbox_m_gltf_xyz` has the expected longest side; `pose.chosen` and
`pose.why` make sense (a tall lantern should stand: `trellis-up`).
- No `oom-*.json` or `cutout-error.json` in the object folder. On an out-of-memory file, see docs/LOW_VRAM.md.
- `record.json` `export_note` is "as configured". Any other value means the export retry ran and the bytes cannot be
made again from the inputs alone.
- Look at the result: `blender -b --factory-startup --python render/render_views.py -- work/views work/out-s42/brass-lantern/raw.glb`
(fixed cameras for the [-0.5, 0.5] box of raw.glb), then view the PNGs. Compare with the reference image; report what differs.
## Re-make or verify an asset
Same GPU model only: `bash tools/check_work_identical.sh PROPS GOLD_ROOT REF_OUT NEW_DIR`. Exit 0 = same bytes.
## Report to the user
The GLB path and its `object.json`, seed, GPU model, `size_m`, the pose chosen, anything from the checks above, and the licence note when
the use is commercial: the DINOv3 licence conditions (THIRD_PARTY_NOTICES.md).
Pliki w paczce (73)
- .dockerignore 1 KB
- .gitignore 1 KB
- AGENTS.md 7 KB
- CONTRIBUTING.md 2 KB
- Dockerfile 3 KB
- LICENSE 1 KB
- README.md 13 KB
- SECURITY.md 1 KB
- SECURITY_SCAN.md 5 KB
- SKILL.md 3 KB
- THIRD_PARTY_NOTICES.md 15 KB
- config.example.env 3 KB
- conjure3d.py 14 KB
- deploy/modal/README.md 3 KB
- deploy/modal/app.py 4 KB
- deploy/vast/README.md 5 KB
- deploy/vast/onstart.sh 2 KB
- docs/BUILD_NETWORK.md 4 KB
- docs/CONTAINER.md 10 KB
- docs/LOW_VRAM.md 10 KB
- docs/MAC.md 8 KB
- docs/PATCHES.md 5 KB
- docs/WORK_IDENTICAL.md 8 KB
- examples/props.example.json 1 KB
- imagegen/flux2-klein-4b.sha256 2 KB
- imagegen/gen.py 8 KB
- imagegen/requirements-pypi.lock 76 KB
- imagegen/requirements-pytorch.lock 5 KB
- imagegen/requirements.txt 2 KB
- imagegen/setup.sh 3 KB
- mac/convert_weights.py 9 KB
- mac/requirements.txt 1 KB
- mac/run_mac.py 12 KB
- mac/setup_mac.sh 5 KB
- post/post_step.py 23 KB
- render/render_views.py 2 KB
- render/turntable.py 5 KB
- render/turntable_video.sh 2 KB
- tests/conftest.py 2 KB
- tests/test_compare_outputs.py 2 KB
- tests/test_conjure3d.py 7 KB
- tests/test_flexgemm_patch.py 8 KB
- tests/test_mac.py 11 KB
- tests/test_netscan_report.py 3 KB
- tests/test_pins.py 5 KB
- tests/test_post_step.py 9 KB
- tests/test_preflight.py 4 KB
- tests/test_run_config.py 8 KB
- tests/fixtures/flexgemm_autotuner_6dd94a85.py 15 KB
- tools/build_netscan.sh 3 KB
- tools/check_work_identical.sh 4 KB
- tools/compare_outputs.py 5 KB
- tools/gpu_guard.sh 2 KB
- tools/lock_requirements.sh 2 KB
- tools/netscan_report.py 10 KB
- tools/preflight.py 28 KB
- tools/ram_guard.sh 1 KB
- tools/scan.sh 4 KB
- trellis2/cumesh_lean_remesh.py 15 KB
- trellis2/requirements-pypi.lock 100 KB
- trellis2/requirements-pytorch.lock 6 KB
- trellis2/requirements.txt 2 KB
- trellis2/run.py 83 KB
- trellis2/setup.sh 21 KB
- trellis2/patches/cumesh-deterministic.patch 5 KB
- trellis2/patches/cumesh-no-cubvh.patch 2 KB
- trellis2/patches/flexgemm-autotuner-threadsafe.patch 2 KB
- web/bake_proxy.py 6 KB
- web/bun.lock 12 KB
- web/export_web.sh 5 KB
- web/package.json 1 KB
- web/verify_export.mjs 4 KB
- web/webvariant.mjs 19 KB
Wersje
W bazie nie ma innych skilli o nazwie conjure3d.
Migawka bazy z 13.11.2026 · wersja paczki 6f1100fdd1ed (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100free-text-to-3d
Tworzy teksturowany model 3D GLB z opisu tekstowego przez darmowe narzędzie MCP i zwraca link do modelu oraz podglądu. - 90forge-3d
Generuje z opisu tekstowego model 3D w formacie GLB z teksturą, bez płatności i klucza API, do prototypów, rekwizytów lub postaci. - 100model-maker
Generuje model 3D w formacie GLB z opisu tekstowego lub zdjęcia przez Tripo, z teksturami, limitem wielokątów i materiałami PBR dla Blendera, Unity czy Unreal. - 100modeler
Tworzy model 3D w formacie GLB do gry z opisu, zdjęcia, szkicu lub kilku widoków, dobierając odpowiednie narzędzie Sloyd. - 100game-dev-character-model
Tworzy realistyczny model postaci 3D z opisu lub referencji jako teksturowany, zrigowany plik GLB w podglądzie three.js. - 90pb-image-to-3d
Zamienia jeden obraz referencyjny w oczyszczony, przeskalowany zasób 3D: workflow ComfyUI tworzy siatkę, a Blender redukuje poligony, poprawia skalę, origin i materiał oraz eksportuje plik .glb.