Asset Generator

Generate PNG images (Gemini or xAI Grok) and GLB 3D models (Tripo) from text prompts. These are paid APIs — every call costs real money. The free exception is qwen-image, a local GPU generator some machines have. Image tools live at ${ASSET_GEN_SKILL_DIR}/tools/; 3D goes through the tripo CLI. Run from the project root and keep runtime-loaded outputs under ${RUNTIME_ASSET_DIR}/.

Models

Model	Flag	Cost	Notes
Gemini 3.1 Flash Image	--model gemini	5¢ (512) · 7¢ (1K) · 10¢ (2K) · 15¢ (4K)	~10 s per image
Grok Imagine Image 2.0	--model grok	6¢ (1K) · 8¢ (2K), +1¢ per reference image	1–2 min per image
Qwen-Image	qwen-image CLI (local, if installed)	free, minutes per image	Simple images — textures, props, icons, UI, backgrounds, in-image text; native transparency

Gemini and Grok are equally strong: both follow detailed prompts closely, and both slip on small details — a miscounted item, a mirrored left/right. Use whichever key is set; with both, asset_gen.py defaults to Gemini for speed. When an asset is quality-critical (a character reference that anchors 3D or animation, a hero image) and both keys are set, generate it with both and keep the better one.

Images

python3 ${ASSET_GEN_SKILL_DIR}/tools/asset_gen.py image \
  --prompt "the full prompt" -o ${RUNTIME_ASSET_DIR}/img/car.png

--model (default gemini when its key is set, else grok) · --size (default 1K; Gemini also 512/4K) · --aspect-ratio (default 1:1; also 16:9, 9:16, 4:3, 3:4, 3:2, 2:3).

Image-to-image: pass --image ref.png and the model sees the reference — prompt only for what changes (angle, pose, recolor), don't re-describe appearance. Use this for style families (one hero asset → the rest), variants, and multi-view sets.

Small sprites: minimum generation is 1K, so a 1024px image downscaled to 64px looks muddy. Design display sizes ≥128px, or generate a kit (multiple objects in one 1K image) and slice it with tools/grid_slice.py ... --grid 2x2 --names "a,b,c,d", or prompt for bold flat forms that survive downscaling.

Review every PNG before any GLB conversion — a bad image wastes 30+ credits downstream.

Background removal

Read ${ASSET_GEN_SKILL_DIR}/rembg.md. Key rule: never prompt for a "transparent background" (the generator bakes a checkerboard) — prompt a solid color, then matte it out. The one exception is qwen-image rgba, which outputs real alpha.

Local generation (qwen-image)

If command -v qwen-image finds it, this machine runs Qwen-Image-2.1 on its own GPU. It costs nothing, so simple images go there first; keep the paid models for characters, references, and exact layouts.

qwen-image generate "the full prompt" -o ${RUNTIME_ASSET_DIR}/img/crate.png                 # --size WxH or --resolution N (default 1024)
qwen-image rgba "a wooden shield, game icon" -o ${RUNTIME_ASSET_DIR}/img/shield.png         # transparent PNG, no matting
qwen-image edit -i crate.png "mossy and cracked" -o ${RUNTIME_ASSET_DIR}/img/crate_old.png  # image-to-image; repeat -i for more refs

• One call at a time: each loads the whole model onto the GPU, so parallel calls run out of memory. Batch as one sequential loop.
• Minutes per 1024² image at the default 40 steps (~2.5 min on a 12 GB card) — longer than common shell-tool timeouts, so set a long timeout or run it in the background. Time the first call (seconds in --json) and plan batches from it; --resolution 512 --steps 10 is ~8× faster for checking how a prompt reads.
• Judge rgba alpha from a composite on a contrasting color (magick out.png -background magenta -flatten out_qa.png), not the raw PNG.
• Out of memory → lower --resolution (edits use the most). On failure, qwen-image info reports GPU and model status.

Animated sprites

Recipe: reference → pose → video → extract frames → loop-trim → rembg.

1. Reference (1K, neutral pose, solid BG) — anchors everything; review carefully.
2. Pose per action: image-to-image from the reference, prompt only the action.
3. Video from the pose frame: asset_gen.py video --image pose.png --duration 2 -o walk.mp4 (Grok, needs XAI_API_KEY; --duration 1–15s; --resolution 720p at 14¢/s, or 480p at 8¢/s — enough for small sprites).
4. Extract: ffmpeg -i walk.mp4 -vsync 0 frames/%04d.png.
5. Loop-trim looping cycles (walk/idle): tools/find_loop_frame.py frames/ returns the loop frame; delete frames past it. Skip for one-shots (attack/death).
6. Batch matte: tools/rembg_matting.py --batch frames/ -o clean/.

Reuse one reference for all of a character's actions. Chaining (feed action A's last frame as action B's start) keeps positional continuity — keep chains ≤2 deep, they drift.

3D models

The tripo CLI (npm install -g tripo-cli, key in TRIPO_API_KEY) owns the whole 3D path: submit, poll, download, credit pre-check, refunds on failure. Its own agent docs are the reference — tripo docs --llm, then tripo docs --topic commands/process / examples/animation / common-errors — this section only covers what is specific to game use here.

tripo make ref.png --name car -p face_limit=30000 -p auto_size=true --json --yes -o ${RUNTIME_ASSET_DIR}/glb
tripo make ref.png --name hero --then rig-check,rig:model=v1.0-20240301 --json --yes -o ${RUNTIME_ASSET_DIR}/glb
tripo anim retarget @hero --animation preset:biped:walk preset:biped:idle --json --yes -o ${RUNTIME_ASSET_DIR}/glb

• make is blocking (default timeout 30 min) and prints one JSON line: read model_file, preview.png and credits_consumed from it. Never add your own shorter timeout, never resubmit because a task_id appeared in stderr. If the process does die, tripo task watch <id> --download finishes the same task for free.
• Output lands in <-o dir>/<name>-<id8>/ (model.glb, preview.png, task.json). Move or reference the GLB from there; task.json keeps the seeds and task id, so there is nothing else to save.
• --name X makes the task addressable as @X for later steps (retarget, convert, decimate). Retarget reuses the rig task — never re-rig for another clip; up to 5 animations per call, billed per animation.
• Model defaults to v3.1. For face_limit ≤ 20000 the CLI silently switches to P1 (low-poly topology, no geometry_quality) — that is the right choice for mobile-style budgets, but know it happens. --for game-pc converts to FBX by default; skip it for GLB engines.
• -p geometry_quality=detailed -p texture_quality=detailed is the HD tier (≈ double credits).
• Rig: keep rig:model=v1.0-20240301 for bipeds — that is the rig motion.md's pipeline is certified against, and it uses the preset:biped:* clips below. The CLI's default rig (v2.5) covers quadrupeds, avians, etc. with preset:<name> clips (idle walk run dive climb jump slash shoot hurt fall turn); unverified with motion.md. rig-check in the chain aborts before rigging if the mesh isn't riggable. --animate-in-place when game code drives locomotion.
• Don't assume the preset name survives into the GLB; inspect the imported clip names before wiring playback.

Source image for make: 3/4 elevated angle, solid white/gray background, matte finish, opaque glass, single centered subject — and do not rembg it (Tripo needs the solid bg). For characters, generate the reference in a T-pose.

v1.0 biped retarget presets (pass as preset:biped:<name>):

afraid agree angry_01/02/03 basketball_shot bow box_01/02/03 cast_a_spell cheer chop
clap climb complain_01/02 cross_body_crunch crossover_dribble cry dance_01..06
defeat_02/03 depressed dig dive dribble fall fire flee_01/02 flip fold_arms
football_catch/save/pass freaky frightened front_kick_01/02 frustrated_01/02 golf
greet_01..04 heart_pose hit_to_body_01/02 hit_to_head/side/stomach hug hurt idle
jump jump_down jump_rope_01/02 laugh_01/02 lift_heavy look_around make_a_call_01/02
pitch_baseball play_mobile_game play_video_game press-up run run_upstairs scared_01/02
scratch shoot shovel sing_01..04 sit slash sob standing_relax surf swagger swim turn
victory_celebration volleyball wait walk warm_up wave_goodbye_01/02

Presets are generic stock clips. Important: when gameplay needs a custom humanoid move set (state machines, root-motion locomotion, moves not in this list), read ${ASSET_GEN_SKILL_DIR}/motion.md.

Costs

Paid generations cost real money, so confirm with the user before generating; qwen-image runs are free. Quick reference: 1K image 6–7¢ · 2K background 8–10¢ · a quality-critical image generated on both models ~13¢ · sprite video 14¢/s at 720p. Tripo bills in credits (≈1¢): ~30 per model, ~25 to rig, ~10 per retargeted clip — tripo balance before a batch, and report the credits_consumed the CLI returns rather than an estimate.

Output and logging

Each asset_gen.py command prints JSON to stdout: {"ok": true, "path": "...", "cost_cents": 7}; tripo and qwen-image print theirs with --json. Progress goes to stderr — redirect it to a temp file and read only on failure to keep context clean:

_log=$(mktemp)
result=$(python3 ${ASSET_GEN_SKILL_DIR}/tools/asset_gen.py image --prompt "..." -o p.png 2>"$_log") || tail -20 "$_log"

Generate independent API images in parallel (multiple Bash calls in one message).

Visual pitfalls

Generators and vision checks have weak spatial sense — verify from screenshots when it matters.

• Direction/orientation is unreliable ("facing left" vs "right" often comes out identical). Generate one direction and flip horizontally at runtime rather than paying for the mirror.
• Mixed sizes: image frames are ~1024px, video frames smaller (960px square at 720p). Downscale everything to the smallest source before matting (magick in.png -resize 960x960 out.png).
• Playback fps: source videos are ~24fps — drive sprite playback off elapsed time at ~1/24s, and only restart a loop when the animation state actually changes.

Asset manifest (in README.md)

Track every generated asset in README.md with an in-game Size column — without it, coders consistently scale assets wrong:

• 3D models: meters, e.g. 4m long, 1.8m tall, 0.3m
• Textures: tile size, e.g. 2m tile
• Backgrounds: pixel size + behavior, e.g. 1920x1080, fullscreen
• Sprites: display pixels, e.g. 128x128 px

Name	Description	Size	Path	Cost
car	sedan with spoiler	4m long	${RUNTIME_ASSET_DIR}/glb/car.glb	7¢ + 30 cr