CXP Character Forge
Use scripts/tripo_proxy.py in this skill directory. On first install, run python scripts/bootstrap.py to get fixed version dependencies; read the README for configuration instructions. Only send keys to the proxy address configured by the user; do not call the upstream official CLI's configure.
Tasks and References
Respect the user-confirmed design, proportions, final version, and static/animation use cases. First check existing assets and task records to avoid regenerating. When creating a new character, first ensure Charakuru source code is installed locally (download from Releases and extract; Node 22+, Blender 5.2, VRM Add-on), as a method library for front/side/back views, repaint groups, UV restoration, weight transfer, expressions, and swing physics; see Charakuru Methods. Recommended pipeline: T-pose front and back concept art → generate the whole model at once (only call 1 task at a time, no parallel tasks, no card draws; very expensive on credits) → check gray model and mesh, on-the-spot test separation of head, hair/bangs, accessories, skirt (if they can't be separated, modify the art and regenerate) → quick bind and run a simple MMD motion to verify the skeleton, then have the AI fix bone nodes and weights → Tripo retexture and convert back to UV (most time-consuming) → expressions → physics last; during production, 2D characters use Genshin-style fast NPR materials with gray model checking by default (fast rendering, convenient iteration), and only at the last step use genshin-blender-render for final rendering; eyes, mouth, and expressions are mandatory—do not skip any of the three: the Charakuru methodology, keep the original face shape of the target, and ensure the full expression network and textures (including oral cavity structure) of the reference character; see Expression Prerequisites, Eye-Brow Workflow; do not generate the head and body separately (neck seam, proportions, skin color are hard to match; this differs from the original Charakuru). Steps, experience, and MMD details are in Character Workflow; the component splitting checklist and whether each component should use Tripo / Codex projection / geometric generation / flat shading are in Component Texturing, with scripts in scripts/component_tex/. For eye refinement when placing alongside MMD / Genshin PMX characters (eyeball upper lid shadow, layered pupils, glow specks), see Eye Refinement, with scripts in scripts/eye_refine/. For Houdini UI-less cloth and hair simulation, see Houdini Cloth Hair; for full-song segmented simulation, approved segments only simulating later parts with concatenation, real multi-layer skirt simulation (layer stitching, crinoline, gray model acceptance), remote Linux installation, per-machine authorization rules, and chunk organization, see Houdini Full Song Remote, with scripts in scripts/houdini_song/.
First verify A/T pose, head-to-body ratio, fingers, front and back silhouette, and occlusions. Multi-view must be the same character and same pose; unreliable side/back views should not be forced into generation. Hidden surfaces are assumed to be filled in by the model.
Generation and Recovery
doctoroffline checks configuration;doctor --onlineread-only verification;models,balancequery models and balance. Configuration service compatibility conventions are in Proxy API.- Get the current model value, supported parameters, and prices; do not treat historical models or proxy parameters as official generic specs. Generation may be executed within existing production authorization; connection tests only—do not submit paid tasks.
- Prefer generation with native textures; inspect the true textured model as early as possible. Gray models are for diagnostic assistance only, they cannot replace textured inspection.
upload --file reference.png --record uploads/front.jsonuploads a reference; fill the returned token and current model value into a copy of examples/image-to-model.json.submit --request request.json --job jobs/v1 --dry-runvalidates offline, then use the same command with--executeto submit.wait --job jobs/v1 --timeout 45; exit code 3 means continue waiting; upon success,download --job jobs/v1 --extract.- Reuse the same request's job; use a new directory for changed parameters. When the POST result is unknown, first reconcile before recovering; do not blindly resubmit. Distinguish between native downloads and paid format conversions.
Blender Acceptance
Keep the original generated file; modify a copy. Render the real model from front, side, back, and rotated views; do not treat 2D generated images as model results. Check fingers, face, texture seams, hair roots, head decorations, tail volume and connections, wing roots, and arm intersections.
Preserve the user-approved native textures and facial proportions; NPR cannot fix erroneous geometry or baked-in texture artifacts. Local replacements need 3D fitting and side-view acceptance; do not rely solely on front projection. Tail/hair and similar components must have continuous roots and adequate thickness.
Static delivery does not require forced rigging. When animation is needed, execute prereigcheck → rig → flags/animations → retarget, and verify real deformation. MMD expressions, bone mapping, and clothing/hair physics must be created and validated individually; two independently generated heads cannot be treated as continuous blink shape keys.
Delivery
Organize final output directories by character and version. Blender textures may be embedded or use complete relative paths. Provide GLB, realistic renders, and rotation videos on request. Report the completion stage, angles checked, and remaining issues. When the user requests only the model package, do not include videos/process files; report only animation capabilities that have actually been verified.
Offline regression: python scripts/test_proxy.py. Script checks interface behavior; it does not validate generation quality, skinning, or physics.
Native mouth shape transfer, VMD expression track integration, and mouth area fragmentation/skin color issues, see Mouth Transfer and Repair.