Godot asset gen

2D Asset Generator

Generate PNG images and sprite animations. 3D (GLB / rigging / animation via Tripo3D or Meshy) lives in godot-asset-gen-3d.

Image models

Model (--model)	Vendor	Sizes	Cost (est. ¢)	Best for
gemini-3.1-flash-lite-image	Google	512/1K/2K/4K	2/3/5/8	cheap drafts, textures, kits
gemini-3.1-flash-image default	Google	512/1K/2K/4K	5/7/10/15	precise prompt following — refs, characters, layouts
gemini-3-pro-image	Google	1K/2K/4K	15/20/30	hero art, 4K quality ceiling
gemini-2.5-flash-image	Google	1K	4	legacy cheap fallback (Nano Banana)
gpt-image-2	OpenAI	1K/2K	10/15	alternative style; strong text-in-image
meshy-nano-banana	Meshy	provider-controlled	3 credits	low-cost text or reference edit
meshy-nano-banana-2	Meshy	provider-controlled	6 credits	balanced generation; multi-view 3D refs
meshy-nano-banana-pro	Meshy	provider-controlled	9 credits	higher-quality generation/edit
meshy-gpt-image-2	Meshy	provider-controlled	9 text / 12 edit credits	high-fidelity Meshy route
wan-2.7-t2i / -pro	Alibaba	1K/2K(/4K pro)	2-12	content-permissive t2i (spicy/nudity game art other providers refuse)
wan-2.7-edit / -pro	Alibaba	1K/2K(/4K pro)	3-15	content-permissive image EDIT (requires --image)

Cent costs are estimates. Meshy reports service credits and must never be converted to cents; re-check Meshy pricing before a batch.

Providers (--provider)

Route	Key	Notes
official	GOOGLE_API_KEY / OPENAI_API_KEY	Google AI Studio / OpenAI direct
wavespeed	WAVESPEED_API_KEY	wavespeed.ai relay; the ONLY route for wan-2.7-*
meshy	MESHY_API_KEY	Meshy Text-to-Image / Image-to-Image; multi-view and pose controls
auto (default)	—	Meshy models use Meshy; other models prefer their official API, then WaveSpeed

Before a Meshy image task, read the complete meshy/reference.md. Do not load it for unrelated providers.

Video models (video --video-model) — for animated sprites

Model	Route	Cost	Res	Duration
seedance-2.0	Volcano Ark 直链 (ARK_API_KEY)	1.0¢/s	up to 1080p	4-15s
seedance-2.0-fast default	Ark 直链	0.8¢/s	720p	4-15s
seedance-2.0-mini	Ark 直链	0.6¢/s	720p	4-15s
wan-2.7	wavespeed	~5¢/s (est.)	720p	5/10s

Seedance model ids are Ark inference endpoints, overridable via SEEDANCE_2_0_EP / SEEDANCE_2_0_FAST_EP / SEEDANCE_2_0_MINI_EP.

CLI Reference

Set SKILL_DIR to the absolute directory containing this SKILL.md; do not assume a Claude-, Codex-, or repository-specific install path. Run from the project root. Keep runtime-loaded outputs under assets/; review-only refs/scratch elsewhere.

Generate / edit image

python3 "$SKILL_DIR/tools/asset_gen.py" image \
  --prompt "the full prompt" -o assets/img/bucket.png

• --model (default gemini-3.1-flash-image), --provider auto|official|wavespeed|meshy
• --size 512|1K|2K|4K, --aspect-ratio 1:1|16:9|... (default 1:1)
• --image path — reference/edit input, repeatable (wan edit & gemini support multi-ref)
• Meshy: --provider meshy --model meshy-*; --multi-view; optional --pose-mode a-pose|t-pose; Image-to-Image accepts 1-5 --image inputs

Typical picks:

• cheap texture/kit → --model gemini-3.1-flash-lite-image
• precise prop/character → default model
• hero/keyart → --model gemini-3-pro-image --size 2K
• spicy content → --model wan-2.7-t2i-pro; spicy edit → --model wan-2.7-edit-pro --image src.png
• 3D-ready multi-view → --provider meshy --model meshy-nano-banana-2 --multi-view

Meshy multi-view → 3D handoff

python3 "$SKILL_DIR/tools/asset_gen.py" image \
  --provider meshy --model meshy-nano-banana-2 --multi-view \
  --prompt "single stylized clockwork knight, clean silhouette" \
  -o assets/img/clockwork_knight.png

The result JSON and <output>.meshy.json include the Meshy image task ID. The 3D skill can pass it directly to Multi-Image-to-3D, preserving all generated views without a download/re-upload round trip:

python3 "$ASSET_3D_SKILL_DIR/tools/asset_gen_3d.py" glb \
  --provider meshy --meshy-image-task IMAGE_TASK_ID \
  -o assets/glb/clockwork_knight.glb

If Meshy polling times out, do not submit again:

python3 "$SKILL_DIR/tools/asset_gen.py" resume-image \
  -o assets/img/clockwork_knight.png

Remove background

Read rembg.md (CLI, BG-color strategy, batch mode).

Item kit (one image, N props)

2x2/3x3 grid in one generation, then slice:

python3 "$SKILL_DIR/tools/grid_slice.py" grid.png \
  -o assets/img/items/ --grid 2x2 --names "bucket,stool,ladle,lantern"

Animated sprite (ref → pose → video → slice → loop → rembg)

1. Reference (image, 1:1, solid BG color per rembg.md)
2. Pose frame (image + --image ref, prompt = the action pose)
3. Video (pose frame as first_frame):

python3 "$SKILL_DIR/tools/asset_gen.py" video \
  --prompt "walking to the right, smooth walk cycle, solid dark-green background" \
  --image assets/img/knight_walk_pose.png \
  --video-model seedance-2.0-mini --duration 4 -o assets/video/knight_walk.mp4

4. Frames: ffmpeg -i knight_walk.mp4 -vsync 0 frames/%04d.png
5. Loop trim (cycles only): python3 .../tools/find_loop_frame.py frames/
6. Batch rembg: python3 .../tools/rembg_matting.py --batch frames/ -o assets/img/knight_walk/

A 4s mini video = ~2.4¢; a full walk cycle (ref 7¢ + pose 7¢ + video 2.4¢) ≈ 17¢.

Set budget

python3 "$SKILL_DIR/tools/asset_gen.py" set_budget 500

Only call when the user explicitly provides a budget. Shared assets/budget.json (the 3D skill records into the same file).

Output format

JSON to stdout. Cent-priced routes return cost_cents; Meshy returns credits, task_id, and paths for multi-view output. Progress/API noise goes to stderr:

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

Image Resolution

Use full generation resolution; downscale only to hit game PPU targets.

• 1K default: textures, sprites, refs
• 2K: HQ objects, backgrounds, title screens
• 4K (gemini pro/lite, wan pro): large maps, panoramas

Small sprites problem

1024px downscaled to 64px goes muddy. Mitigations: design at 128px+; generate kits (more px per object); prompt bold simple forms (thick outlines, flat colors) that survive downscale.

What to Generate — Cheatsheet

For transparency, read rembg.md first (BG color strategy).

• Background/scenic (2-10¢): flash-lite for simple; flash/pro --size 2K --aspect-ratio 16:9 for precise layout. No post-processing.
• Texture (2-3¢): flash-lite. "Top-down view, uniform lighting, no shadows, seamless tileable."
• Single prop/sprite: flash (precise) or flash-lite (cheap); solid BG color → rembg.
• Character design (7¢): flash 1K; variants via --image ref (prompt only the CHANGE).
• Spicy/nudity content: wan-2.7 family via wavespeed only; same composition rules apply.
• 3D model reference → then hand off to the godot-asset-gen-3d skill: "3/4 front elevated camera angle, solid white background, matte finish, single centered subject, opaque glass."

Visual Pitfalls

• Direction/orientation: generators can't do left/right reliably — generate ONE facing, flip at runtime.
• Video size consistency: resize 1024px stills DOWN to video frame size before rembg (magick input.png -resize 720x720 -filter Lanczos out.png).
• Playback fps: video frames are ~24fps; drive playback at 1.0/24.0, don't restart the loop on input jitter.
• Image-to-image prompting: with --image, don't re-describe the subject — prompt only what CHANGES.
• Generate multiple images in parallel via multiple Bash calls in one message.
• Review PNGs before building on them (bad ref poisons every downstream step).