game-visual-forge

Game Visual Forge

English | 简体中文

Game Visual Forge brings five Codex Skills into one creative entry point. Describe what you need in Codex, and it can use configured local models and tools to create, check, and deliver 2D maps, character sprites, video-derived animation frames, and game audio at low cost. Models and tools are set up locally as needed.

Skills

What it provides

• Natural-language intake with explicit choices for style, layout, runtime, and delivery format.
• Built-in image generation when a new visual asset is needed, plus support for user-provided source media.
• Deterministic local processing, validation reports, and reproducible output manifests.
• Playable map handoff with walkability, collision, objects, entrances, and Unity Tilemap support.
• Native map atlases are normalized to the declared grid, tile size, margins, and spacing before validation. This standardizes atlas geometry only; it does not repair artwork, seams, or map topology.
• Provider, cost, and submission confirmation gates for paid or external work.
• forge-video-to-sprite processes existing video locally with FFmpeg/FFprobe, timestamp sampling, rembg/Chroma cleanup, stable alignment, strips, sheets, GIF previews, and motion-quality evidence.
• Generated video routes support explicit existing, MiniMax Hailuo, Jimeng, or optional ComfyUI MiniMax H3 selection; tools and credentials are configured manually and never switched automatically.
• forge-text-audio uses the official local Stable Audio 3 small-sfx model through the isolated stable-audio-3 runtime. It supports text-to-audio, redraw, inpaint, and continue modes, with WAV-only delivery and a final listening review.
• forge-game-music uses local Comfy Desktop MiniMax Music 3 by default; Stable Audio 3 is a documented cloud fallback and Suno remains gated until its official API schema and game terms are verified.

Showcase

HD cleanup uses Pillow for image conversion, NumPy/SciPy for masks, rembg with the birefnet-general model for semantic separation, known-magenta spill reconstruction, and optional PyMatting refinement. If CUDA fails, the processor tries CPU and reports a deterministic Chroma fallback when needed.

Install optional cleanup tools manually:

python -m pip install -e ".[image]"
python -m pip install -e ".[background]"
python -m pip install "rembg[cpu]"
python -m pip install "rembg[gpu]"
python -m pip install -e ".[matting]"
python -c "from rembg import new_session; new_session('birefnet-general')"

Use U2NET_HOME for a shared model directory. CPU is the compatibility default; GPU requires a verified CUDA environment. PyMatting is optional and slower.

Video to sprite: character animations

Original video → v12 walk sprites

The Soldier walk source was generated with a local ComfyUI MiniMax H3 FL2VA workflow: 640 × 640, 24 FPS, 124 frames (about 5.17 seconds), Turbo/LowVRAM 8 steps, fixed seed 20260917. The H3 local open weights are available free under the Community License. This run used local compute, with no paid API call or cloud generation credits. Current local hardware: Intel Core i7-13700KF, 32 GB RAM, and NVIDIA GeForce RTX 4070 Ti with 12 GB VRAM. From newer local tasks, nine completed generations were selected; submission to completion averaged 67.6 seconds (about 1 minute 8 seconds, range 64–72 seconds). This showcase source is a longer 124-frame run: it was submitted at 04:48:23 and completed at 04:50:19 (116 seconds); the ComfyUI log reports 114.58 seconds of prompt execution. Samples and timing sources

The left panel shows seconds 2–4 of the original video. The right panel shows a newly created empty character prefab playing 48 transparent 128 × 128 frames re-extracted from that same source at 24 FPS, looping every 2 seconds. The original twelve-frame Soldier_L1_Walk_PseudoTopDown_128_from2s_v12 asset remains unchanged; the new preview adds intermediate poses at the source speed.

Full original video · MP4 comparison

Soldier_L1_View

Soldier animation MP4

The Soldier Idle, Walk, and Attack sprites were extracted from locally generated H3 videos. The preview retains the prefab's visual hierarchy, anchors, shadows, and Sprite animation references. Attack retains its clip timing and repeats after a short Idle pause. The current Walk uses the first eight v16 frames, separate from the v12 example above. Source videos and animation timing: Idle 39 frames / 1.625 seconds → 12 frames / 12 FPS; Walk 124 frames / about 5.17 seconds → 8 frames / 8 FPS; Attack 107 frames / about 4.46 seconds → 12 frames / 12 FPS.

Animation in action

35-second animation capture MP4

The capture shows Soldier_L1_View, Enemy_Basic_View, and Enemy_Fast_View playing their animations. Soldier's three actions and Enemy_Basic's Walk were made from locally generated H3 videos. The capture itself records actual Tuanjie Editor playback; only the viewport and time interval were trimmed, preserving the original 12 FPS. Timing: the GIF uses seconds 6–18 (12 seconds) of the recording; the MP4 uses seconds 0–35 (35 seconds). Asset provenance and parameters

Stable Audio 3 examples

Generated with forge-text-audio and the official stabilityai/stable-audio-3-small-sfx model.

Wooden UI click

Prompt:

Dry wooden UI click, short transient, no music, no voice

Listen to the wooden UI click

Blacksmith hammer sound

Prompt:

TrackType: SFX, a clean professional studio Foley recording of one natural strike of a small steel blacksmith hammer against a red-hot iron billet on a solid anvil, an isolated metallic impact with a fast attack and a short clean natural decay, recorded with a dry close microphone in a quiet room.

• Listen to blacksmith hammer candidate 1
• Listen to blacksmith hammer candidate 2
• Listen to blacksmith hammer candidate 3

Implementation and validation: generated locally with Stable Audio 3 Small-SFX, exported through no-boost WAV processing, and passed format, clipping, and sustained-noise checks.

Local MiniMax Music 3 BGM test

The local Comfy Desktop audio_minimax_music_3 workflow produced this cheerful instrumental game-BGM candidate. The checked-in WAV is 60 seconds, 44,100 Hz, 16-bit stereo, with a -1.57 dBFS peak; native Comfy output was MP3.

Prompt focus: continuous medium-energy 4/4 groove, no intro or low-energy section, mid-register marimba and muted pluck, strictly instrumental, clean resolved ending.

• Listen to the 60-second WAV
• Listen to the 60-second MP3 · Listen to the final 5 seconds
• Waveform · Provenance

Install

• Unified installation guide
• 统一安装指南（中文）
• Stable Audio 3 setup
• Comfy MCP official installation
• MiniMax H3 Prompt Writing Skill official installation

The unified guide contains both a copyable Agent request and the complete manual installation flow. It installs the five core Forge Skills only. Optional workflows are opt-in. The Agent asks whether to enable them, inspects first, then asks again before installing missing components.

Provider setup stays separate from core Skill installation. Core installation does not auto-install providers, FFmpeg, credentials, model weights, or optional workflow dependencies.

To delegate the optional audio runtime setup to an Agent, copy this request:

Install and configure the official stable-audio-3 runtime for forge-text-audio: ask me to choose the installation directory, keep the isolated Python environment, model weights, and all caches under that directory, require me to accept all licenses personally, never use a hosted API, run the repository's provider configure command without changing user environment variables or PATH, and finish by running and reporting the local offline preflight.

Detailed instructions are in Stable Audio 3 setup. The ignored game-visual-forge.local.json stores only local paths; provider configure creates it. License acceptance remains a user action.

Example requests

Use forge-2d-map to create a top-down village with a north-south creek,
one walkable wooden bridge, building entrances, collision, and a Unity Tilemap.

Use forge-2d-sprite to create a modern pixel-art player walk sheet with
transparent output and a preview GIF.

Use forge-video-to-sprite with my existing MP4 and export a feet-aligned
24-frame sprite strip plus a preview GIF.

Use forge-video-to-sprite with local ComfyUI MiniMax H3, optimize an I2VA
prompt from my first frame, then export aligned 16- and 32-frame sprite sheets.

Use forge-text-audio to create three dry iron-sword impacts against a steel
shield, no music or voice, review the candidates, and prepare a Unity AudioClip.

For local video processing, install FFmpeg and FFprobe yourself. MiniMax Hailuo API uses MINIMAX_API_KEY; Jimeng API uses JIMENG_ACCESS_KEY and JIMENG_SECRET_KEY. The official mmx and dreamina CLIs are optional.

Unity integration

The Tilemap package lives at integrations/unity/com.game-visual-forge.tilemap and imports validated Tilemap bundles into textures, Tiles, Palettes, and Tilemap Prefabs. The independent audio package lives at integrations/unity/com.game-visual-forge.audio and imports reviewed WAV bundles as configured AudioClip assets without changing scenes. Scene placement is performed through Unity MCP only after an explicit request.

Repository layout

skills/       Codex Skill instructions and launchers
src/          Shared contracts, routing, processing, and reports
integrations/ Unity Tilemap and audio packages plus tests
assets/       Small checked-in examples and showcase artifacts
install/      Unified installation and independent runtime guides

Development check

Run the test suite from the repository root (the directory that contains tests/):

Set-Location "game-visual-forge项目路径"
python -m unittest discover -s tests -q

License

MIT