Game Asset Pipeline — Local AI
Machine
- Windows 10, RTX 3090 24 GB, 28 GB RAM (→40 GB), Xeon E5-2696 v4
- ComfyUI 0.44.0 at
D:\AI\ComfyUI, models atD:\AI\models\ - Docs:
D:\AI\SETUP_DOCUMENTATION.md,D:\AI\HANDOFF.md,D:\AI\PROJECTS.md
Task → Model mapping
| Task | Model | Workflow | Notes |
|---|---|---|---|
| Fast general images | Z-Image Turbo (int8/FP8) | image_z_image_turbo_int8.json | 8-9 steps, CFG 1.0, 1024-1344px |
| Anime illustration | Nova Anime XL v19 / Prefect Pony XL v6 | nova_anime_xl_v19.json / prefect_pony_xl_v6.json | 28 steps, CFG 5-6, CLIP skip 2 |
| Structure-guided (pose/depth/canny) | Z-Image Turbo + ControlNet Union | image_z_image_turbo_fun_union_controlnet_int8.json | Input image drives structure |
| Sprites (pixel art) | SDXL + Pixel-Art-XL LoRA (not installed yet) | To be created | Downscale nearest-neighbor, palette-limit |
| Tileable textures | Z-Image Turbo or SDXL + tileable LoRA | To be created | + TextureAlchemy PBR extraction |
| 3D props | Pixal3D / TRELLIS.2 (native) | 3d_pixal3d_trellis2_image_to_model.json | GLB + PBR, 1024-1536px |
| Video clips | MiniMax H3 (Ref2VA) | video_minimax_h3_r2v_safe.json | 5s 864x480, 4-step turbo; 15s ~23 min |
| Concept art / skyboxes | Z-Image Turbo | t2i workflow | Equirectangular LoRA for 360 skies |
Sprite pipeline
- Generate at 1024 with SDXL + Pixel-Art-XL LoRA (or Z-Image + pixel style).
- Downscale with nearest-neighbor to 32/48/64/96/128 px.
- Palette-limit to 16-32 colors (Aseprite / Pixelorama).
- Manual cleanup (fix stray pixels, align silhouette).
- Consistency: lock seed, ControlNet OpenPose for poses, or train a character LoRA.
- Animation: generate key poses with ControlNet; hand-clean; pack sprite sheet.
Honest limits: AI makes "pixel-art-like" images, not true pixel grids. Walk cycles need hand-cleaning. Cloud tools (PixelLab, Retro Diffusion, Sprite AI) are more turnkey for animation, but local wins on privacy/cost/style consistency.
Texture pipeline
- Tileable albedo: SDXL + tileable LoRA or Z-Image Turbo, 1024/2048, top-down prompt.
- Seamless fix: TextureAlchemy "Seamless Tiling Maker" + inpaint seam.
- PBR extraction: TextureAlchemy (Marigold/Lotus) → normal, roughness, metallic, AO, height, curvature.
- Alternative: CHORD (Ubisoft, research-only — not for commercial).
- Post: Texture Equalizer (remove baked lighting), channel-pack ORM for Unity/Unreal.
- Upscale to 2K/4K with SeedVR2 (conservative, open source) or ESRGAN (fast).
Honest limits: organic/stylized materials are great; precise parametric architecture (brick patterns) still belongs in Substance Designer. CHORD is research-only.
3D pipeline
- Concept image (Z-Image Turbo) — clean background, front view. Optional multi-view.
- Image-to-3D: ComfyUI template "Pixal3D & TRELLIS.2" → GLB + PBR.
- Boolean off = Pixal3D (better input fidelity), on = TRELLIS.2 (complex topology).
- Post-process in ComfyUI (remesh, decimate, fill holes, smooth normals) or Blender.
- Retopo/UV/LOD/collision in Blender; game-ready pass.
- Rig: SkinTokens (MIT, auto-rig, ~6 GB GGML) or Mesh2Motion (free, local, Mixamo-like).
- Export FBX/GLB to engine.
Honest limits: static props are usable after cleanup; characters need manual rigging/weight painting; animation is the hard part. Closed-source cloud (Meshy, Tripo, Rodin) still better for characters.
Consistency and automation
- Workflow JSONs live in the project repo. Commit them; same workflow + seed = reproducible output.
- Staging folder discipline: output →
staging/, human review, then move to engine project. - ComfyUI API: POST workflow JSON to
/prompt; poll/historyor use websocket. - LoRA training: Kohya sd-scripts. SDXL 12-16 GB, FLUX 16-24 GB. 10-30 images, 1024px, 1000-2500 steps. This is the key to a coherent art style.
Licensing (personal use = no restrictions; for commercial:)
| Category | Models | Commercial? |
|---|---|---|
| Apache 2.0 / MIT | Z-Image/Turbo, FLUX.2 klein 4B, FLUX.1 schnell, Qwen-Image/2512, ERNIE, HiDream (MIT), GLM (MIT), LongCat, Pixal3D, TRELLIS.2 (MIT), SkinTokens (MIT), Mesh2Motion | Yes |
| Revenue-capped | FLUX 3 Klein (<$5M + filters/review), SD 3.5 (<$1M), Krea 2 (<$1M), Stable Fast 3D (<$1M) | Under threshold |
| Non-commercial | FLUX.1 dev, FLUX.2 dev/klein 9B, Qwen-Image-2.1, Ideogram 4, HunyuanImage-3.0 (territorial) | No |
| Research-only | CHORD (Ubisoft PBR), Marigold/Lotus (check) | Prototype only |
Also: Steam requires AI disclosure for shipped games. Civitai LoRAs have individual licenses. US copyright requires human authorship.
What's installed vs. not
Installed: Z-Image Turbo, Nova Anime XL, Prefect Pony XL, Pixal3D, TRELLIS.2, MiniMax H3, ControlNet Union, BiRefNet, MoGe, DINOv3.
NOT installed (available on request): SDXL base, FLUX, pixel-art LoRAs, SDXL ControlNet, H3 FL2VA, upscale models (ESRGAN/SeedVR2), texture LoRAs, Kohya training tools, Krita AI Diffusion, Aseprite/Pixelorama.
References
D:\AI\SETUP_DOCUMENTATION.md— full 13-section referenceD:\AI\HANDOFF.md— agent quick reference (install procedures)D:\AI\PROJECTS.md— project trackerD:\AI\README.txt— user quick reference- Prompt repos: booru-prompt-gallery, sd-akashic, SupaGruen cheat sheet, Dalabad templates, roblaughter style-reference, zsky prompt library