Art Pipeline

First-pass concept art (Midjourney) and the first 3D model (Meshy). Everything under Art/ sits outside Assets/, so Unity never imports it into a build. The game-ready golem lives in Assets/Art/Models/TrainingGolem/.

Art direction (style, palettes, shape language) is written up in the "Art Direction" tab of the Mobile RPG Codex: https://claude.ai/artifact/VF3CNzQbCREegLwy2zAdBS

Midjourney setup

• Account folder: group Velan › folder Concepts (all Velan generations, including the ones not picked here).
• Model: Midjourney V8.2, Standard (not Raw), stylize at the account default, Fast mode.
• Style lock: every subject prompt ends with --sref https://cdn.midjourney.com/e1cbf51f-e14d-47a5-878b-299fa466e41a/0_1.png, the style anchor below. Facets-heavy prompts use --sw 50 so the anchor's ivory/sky palette doesn't swamp the prismatic colours; the overworld re-roll and golem model reference use --sw 40.
• Each image is listed as Midjourney job id and grid index (0 top-left, 1 top-right, 2 bottom-left, 3 bottom-right); the full image is https://cdn.midjourney.com/<job>/0_<index>.png while signed in.

Style exploration

Three treatments of the same subject. The cel-shaded 3D look won: it reads as a playable 3D game at phone size and is the closest to what Meshy and URP can actually produce. The gouache is the prettiest but can't be built in 3D.

Subjects

All prompts below end with --sref https://cdn.midjourney.com/e1cbf51f-e14d-47a5-878b-299fa466e41a/0_1.png (written as --sref <anchor>).

The two sides

Characters

Training Golem

Places

The first overworld attempt (job dbbc8014-e630-487d-b002-45148aa095f0, "tilted high three-quarter top-down diorama view…") came out as a perspective landscape, not a top-down map, so it was re-rolled with an explicit steep camera angle and lower style weight.

Spells

November concept push

A second Midjourney pass (VEL-48 to VEL-51) covered the rest of the game. Each folder has its own README with every prompt and job id, a contact sheet and a list of Meshy candidates:

• Concepts/world/: every region and key location, Prologue to Act 4 (VEL-48).
• Concepts/buildings/: architecture and modular kits per culture (VEL-49).
• Concepts/characters/: the protagonists, Aske and the cast (VEL-50).
• Concepts/creatures-items/: enemies, items and spell effects (VEL-51).

Meshy: Training Golem

Built through the Meshy REST API (the key lives in ~/.secrets/velan_meshy_api_key, never in the repo). Total cost: 68 credits.

1. Image to 3D (30 credits)

Input: Concepts/training-golem/golem-front-meshy-input.png, sent as a base64 data URI.

POST https://api.meshy.ai/openapi/v1/image-to-3d
{
  "image_url": "data:image/png;base64,…",
  "ai_model": "latest",
  "should_remesh": true,
  "topology": "triangle",
  "target_polycount": 15000,
  "should_texture": true,
  "enable_pbr": true,
  "texture_resolution": "2k",
  "target_formats": ["glb", "fbx"]
}

Task 01a19b3f-5bf3-75a4-91c1-bf35dafcb1f2. Result: one mesh, 15,435 triangles, 30,744 vertices, one material with a 2K PBR set (base colour, metallic, roughness, normal; Meshy's current model has no emission map, so the chest glyph's glow has to come from an emissive mask or a light in Unity).

2. Rigging (5 credits)

POST https://api.meshy.ai/openapi/v1/rigging
{ "input_task_id": "01a19b3f-5bf3-75a4-91c1-bf35dafcb1f2", "height_meters": 2.4 }

Task 01a19b47-70e7-7708-8cc7-7f982c6a4151. A 24-bone humanoid skeleton (Hips, Spine/Spine01/Spine02, neck, Head, shoulders, arms, forearms, hands, legs, feet, toes). Mixamo's auto-rigger was tried first and failed on this body: the head sits inside the shoulders with no neck, which its humanoid fitter can't place.

3. Animations (3 credits each)

POST https://api.meshy.ai/openapi/v1/animations
{ "rig_task_id": "01a19b47-70e7-7708-8cc7-7f982c6a4151", "action_id": <id>,
  "post_process": { "operation_type": "extract_armature" } }

extract_armature returns skeleton and motion only (~200 KB) instead of re-embedding the 14 MB mesh in every clip. The files follow Unity's Model@Clip.fbx naming.

The preview above is a Blender render of each clip applied to the rigged mesh (a mid-clip frame each), as a check that the skin deforms cleanly.

Files

• Assets/Art/Models/TrainingGolem/TrainingGolem.fbx: the rigged character (mesh, skeleton, embedded textures).
• Assets/Art/Models/TrainingGolem/Animations/TrainingGolem@<Clip>.fbx: the six clips, skeleton only.
• Assets/Art/Models/TrainingGolem/Textures/: the 2K PBR maps as separate files. URP Lit wants smoothness, not roughness, so invert TrainingGolem_Roughness.png (or pack it into the metallic map's alpha) when building the material.
• Art/Models/TrainingGolem/TrainingGolem_static.glb: the unrigged Image to 3D result, kept as the source mesh.

Not done here (left for the Unity side): .meta files and import settings (Unity creates them on first import; set the rig to Generic or Humanoid and point every clip at TrainingGolem.fbx's avatar), the URP material, and the emissive chest glyph.

Meshy: protagonists

The male and female protagonist (VEL-36): Assets/Art/Models/ProtagonistMale/ and ProtagonistFemale/, each with a README listing its clips, height and Meshy task ids. 56 credits each.

What changed from the golem:

• Multi-image to 3D (POST /openapi/v1/multi-image-to-3d) with the four VEL-64 model-sheet views as image_urls, front first, plus pose_mode: "a-pose"; otherwise the golem's settings (15k triangles, 2K PBR). Same 30 credits.
• Rigged at height_meters: 1.7. Seven clips, one animation task each with extract_armature: Idle (0), Walk (30 Casual Walk), Run (14 Run 02), Cast (130 Mage Spell Cast 1, the two-hand gather and sweep), Hit (178), Victory (59 Victory Cheer), KnockedDown (187 Knock Down). GET /openapi/v1/animations/library lists every action with a preview GIF.
• Humanoid import. The model gets its own avatar. The clip files can't copy it: their root is the armature itself, so Unity reports "Copied Avatar Rig Configuration mis-match". Each clip file creates its own Humanoid avatar instead, which works because Humanoid clips are muscle-space. Unmap Jaw: the automatic mapping picks headfront, which skins the whole face.

Meshy: Aske

Aske Tarran, the ferry keeper and mentor (VEL-37), built the same way as the golem through the Meshy REST API. Total cost: 82 credits. The game-ready files, their Humanoid import settings, the clips and every task id are in Assets/Art/Models/Aske/README.md; VEL-70's Import Character builds the prefab.

1. Multi-image to 3D (30 credits)

Input: the four VEL-64 views in Concepts/model-sheets/aske/ (front.png first, then left.png, back.png, three-quarter.png), as base64 data URIs.

POST https://api.meshy.ai/openapi/v1/multi-image-to-3d
{
  "image_urls": ["data:image/png;base64,…", "…", "…", "…"],
  "ai_model": "latest",
  "pose_mode": "a-pose",
  "should_remesh": true,
  "topology": "triangle",
  "target_polycount": 15000,
  "should_texture": true,
  "enable_pbr": true,
  "texture_resolution": "2k",
  "target_formats": ["glb", "fbx"]
}

Task 01a1a026-3287-735f-9408-37448528ab36. Result: one mesh, 15,535 triangles, a 2K PBR set (base colour, metallic, roughness, normal; no emission map). The face, earrings, turtleneck, rolled sleeves, fingerless glove and silver bands all came through from the sheet.

2. Rigging (5 credits)

POST https://api.meshy.ai/openapi/v1/rigging
{ "input_task_id": "01a1a026-3287-735f-9408-37448528ab36", "height_meters": 1.65 }

Task 01a1a029-51e3-73ce-93b2-438bbd0b65f1. The same 24-bone humanoid skeleton as the golem.

3. Text to Motion (10 credits each)

The library has no point and nothing like the tracing lesson, so those two clips were generated from text, then applied to the rig like a library action ("motion_task_id" in place of "action_id", 3 credits each).

POST https://api.meshy.ai/openapi/v1/text-to-motion
{ "mode": "prime", "duration": 4, "prompt": "An older woman standing still raises her right arm and points firmly ahead and slightly to her right across a river, holds the point for a moment while looking that way, then lowers her arm back to her side and relaxes." }

Task 01a1a026-ec8d-7098-a972-60138dc0a1ed (Point). Trace, task 01a1a026-f042-74f4-b41e-908d0e930f70, 6 s: "An older woman stands behind a student and reaches both arms forward at chest height, hands together as if guiding the student's hands, and slowly traces a large curved symbol in the air in front of her: a smooth S-shaped stroke from upper left down to lower right, then returns her hands to the start and lowers them."

4. Animations (3 credits each)

POST https://api.meshy.ai/openapi/v1/animations
{ "rig_task_id": "01a1a029-51e3-73ce-93b2-438bbd0b65f1", "action_id": <id>,
  "post_process": { "operation_type": "extract_armature" } }

Walk and Cast were picked from the library's preview GIFs: Casual Walk is unhurried without being a shuffle, and Mage Spell Cast stays standing (the other casts crouch or lunge).

The preview is a Blender render of Meshy's skinned version of each clip, one frame each. Applying the extract_armature clips to the rigged model in Blender's FBX importer mangles the arms (its bone orientation differs between the two files). Unity's Humanoid import reads them correctly.

Files

• Assets/Art/Models/Aske/: Aske.fbx, Animations/Aske@<Clip>.fbx and Textures/, imported as Humanoid.
• Art/Models/Aske/Aske_static.glb: the unrigged Multi-image to 3D result, kept as the source mesh.

Meshy: Glint

The glint (VEL-38) is a static Meshy mesh animated in code, so it has no rig and no clips. It was built from Concepts/creatures-items/enemies/glint-ref-newt-chunky-a.png with Image to 3D, mesh only (20 credits, task 01a1a026-eda1-7219-804f-c1159290711b). The mesh is kept as Models/Glint/Glint_meshy.glb; Glint_meshy_sheet.glb is an unused alternative from the VEL-66 sheet (20 credits; it has a face). Models/Glint/build_glint.py (Blender) cuts off the shards Meshy fused onto the tail and adds three separate ones, then writes Assets/Art/Models/Glint/Glint.fbx; Velan > Art > Set Up Glint builds the prefab. The rest is in Assets/Art/Models/Glint/README.md.

Characters in the game (VEL-70)

Velan > Art > Import Character (select model folders) or Import All Characters turns every Humanoid Meshy model under Assets/Art/Models/<Name>/ into a game-ready character. Batchmode: -executeMethod Velan.Editor.CharacterImport.ImportAll. Re-running it is safe: it only changes what isn't right yet. Creatures (Generic or no rig) are left alone. Source: Assets/Editor/CharacterImport.cs.

For each model it:

1. Fixes the import. The model and every clip file are Humanoid with an avatar of their own (a Meshy clip file's root is the armature, so copying the model's avatar fails; Humanoid clips are muscle-space, so they play on any avatar of the same skeleton). Jaw is unmapped (Unity maps it to headfront, which skins the face). Clips are named after their file (Aske@Talk.fbx → Talk) or keep their take names inside a merged FBX (Rook, Tamsin, the faction NPCs); Idle, Walk, Run, Talk and Sit loop; root rotation, height and position are baked into the pose. Clip settings go in a second reimport, because Unity drops them from the .meta when the avatar changes in the same one.

2. Builds the material. Assets/Art/Characters/Materials/<id>.mat, Velan/Toon, from the base colour, normal, and metallic with inverted roughness packed into Assets/Art/Characters/Textures/<id>_MetallicSmoothness.png.

3. Fills the shared controller. Assets/Art/Characters/Character.controller holds empty placeholder clips; <id>.overrideController puts the character's own clip in each slot and its Idle where it has none.

   Parameter	What it does
   Speed (float, m/s)	Idle, Walk (1.4) and Run (4.0) blended. CharacterModel sets it from how fast the root moves (a jump faster than 12 m/s is a warp, not a run).
   Talking (bool)	Loops Talk while standing. CharacterModel sets it while the line on screen is the character's.
   Wave, Cast, Hit, Cheer, Point, Explain, Trace, Alarm, LookAround (triggers)	Play once, back to locomotion.
   Defeat, Recover (triggers)	Fall and stay down; get up.
   Sitting (bool)	The Sit layer sits down and stands up.

   Clip aliases: KnockedDown fills Defeat and Victory fills Cheer.

4. Builds the prefab. Assets/Art/Characters/Resources/Characters/<id>.prefab: the model with its avatar and override controller, feet at the pivot, facing +z, no root motion, one LOD that culls below 2% of the screen height. The game loads it on demand with CharacterPrefabs.Load("Characters/<id>") (Assets/Scripts/World/), the one place to change for Addressables.

5. Writes the data if it has none: Assets/StreamingAssets/Data/Characters/<id>.json (id, name, prefab, scale, portrait, tint) and index.json. portrait is the name the concept portraits share (aske for Art/Concepts/portraits/aske-neutral.png), or null when there are none yet. The tint (the base colour's average) colours the placeholder if the prefab is missing. Edit names and scales by hand; the tool keeps them.

In the game, the player wears protagonist_male or protagonist_female for the save's body type, and a location actor with "character": "<id>" wears that character (Sallowbank's Aske does; VEL-71 points the rest). Without one, or if the prefab is missing, the placeholder capsule stands in.

Rook's one-arm fix stays an import postprocessor (Assets/Art/Models/Rook/Editor/RookOneArm.cs), because it has to run on every import of Rook.fbx, not only when the tool runs.

Sallowbank environment kit

The props for Sallowbank, the Prologue village (VEL-39): houses with the door-mark, Aske's cottage and barn, fences and gate, field tiles, farm clutter, the well, the rope ferry, river banks, trees, bushes, reeds and rocks. No generation service: every prop is built by Python scripts in Blender (Art/Environment/Sallowbank/), sized to the greybox, coloured from one 256 x 256 palette texture and shared by one Velan/Toon material.

• Source scripts, conventions and the door-mark emission setup: Art/Environment/Sallowbank/README.md.
• Rebuild: run build_kit.py in Blender (--background --python), then Velan > Art > Build Sallowbank Kit in Unity, which makes the prefabs and Assets/Art/Environment/Sallowbank/KitPreview.unity.
• Game-ready output: Assets/Art/Environment/Sallowbank/ (Models/, Textures/, Materials/, Prefabs/).

From Art/README.md