Draw Thing

Draw Thing

Local AI image and media generation through draw-things-cli. Use the current Draw Things command surface: generate, models list, models ensure, models import, and train lora.

Scope: Local Draw Things image generation, image editing, model setup, model inventory, imports, and short local media generation. NOT for UI implementation (design), ad creative strategy (ad-creative), cloud image APIs, or general AI-model market research.

Canonical Vocabulary

Dispatch

Auto-Detection Heuristic

1. Prompt asks to install, download, configure, or choose best models -> Setup.
2. Prompt asks for model list, current downloads, or missing models -> Inventory.
3. Prompt asks for video, motion, mp4, mov, frames, or animation -> Media.
4. Existing file path plus edit/restyle/change/transform -> Edit.
5. Existing file path plus upscale/control/lora/inpaint -> Advanced, then verify current CLI/config support before running.
6. Descriptive text with no file path -> Generate.
7. Ambiguous -> ask which mode and state the tradeoff.

Prerequisite Protocol

Run before any generation or model mutation:

1. Check CLI: command -v draw-things-cli.
2. If missing, show install command and stop:

   brew install drawthingsai/draw-things/draw-things-cli
   

3. Verify installed command surface: draw-things-cli generate --help.
4. For setup/inventory, prefer:

   uv run python skills/draw-thing/scripts/model_inventory.py --format json
   

5. Always pass --output; without it, the current CLI may only preview in supported terminals instead of writing a file.

Current Best Model Policy

As of 2026-05-15, this Mac is Apple M1 Pro, arm64, macOS 26.5, 32 GiB unified memory. Use current Draw Things optimized checkpoints and do not recommend legacy SDXL/SD1.5/FLUX.1 families unless the user explicitly asks for compatibility.

Legacy/redundant downloaded models are not deletion targets without approval: flux_1_schnell_q8p.ckpt, flux_1_fill_dev_q8p.ckpt, and flux_2_klein_4b_q6p.ckpt.

Auxiliary text/VLM files and incomplete partials are not recommended generation models. For example, llama_3.1_8b_instruct_q8p.ckpt.part is an incomplete stale auxiliary artifact, not a mid-May 2026 best-pack model.

Qwen Layered remains a current optional workflow target, but not a guaranteed best-pack install on this rig: qwen_image_layered_1.0_bf16_q6p.ckpt failed verification with a checksum mismatch, and qwen_image_layered_1.0_bf16_q8p.ckpt repeatedly stalled during CLI download. Use qwen_image_edit_2511_q8p.ckpt for reliable local editing unless the user explicitly asks to keep retrying Layered.

For exact install commands, current/legacy distinctions, and quantization rules, load references/model-catalog.md.

Core Generation Protocols

Every mode follows this pattern:

1. Validate: CLI exists, input files exist, model is downloaded or --download-missing is intentional.
2. Select defaults: choose a current model by role unless the user provided --model.
3. Build command: use only flags shown by draw-things-cli generate --help, or use --config-file for advanced app settings.
4. Show command: display the full command before running.
5. Execute: run through Bash and capture output.
6. Report: output path, seed, model id, dimensions, steps, and cfg.

Mode: Generate

Fast draft default:

draw-things-cli generate \
  --model z_image_turbo_1.0_q8p.ckpt \
  --prompt "<prompt>" \
  --steps 8 \
  --cfg 0 \
  --width 1024 \
  --height 1024 \
  --seed <seed> \
  --output <output.png>

Quality/text default:

draw-things-cli generate \
  --model qwen_image_2512_q8p.ckpt \
  --prompt "<prompt>" \
  --steps 40 \
  --cfg 4 \
  --width 1328 \
  --height 1328 \
  --seed <seed> \
  --output <output.png>

Mode: Edit

draw-things-cli generate \
  --model qwen_image_edit_2511_q8p.ckpt \
  --image <input.png> \
  --prompt "<edit instruction>" \
  --strength 0.55 \
  --steps 30 \
  --cfg 4 \
  --seed <seed> \
  --output <output.png>

Use --strength 0.25-0.4 for subtle edits, 0.5-0.65 for visible edits, and 0.75+ for heavy redraws.

Mode: Media

draw-things-cli generate \
  --model ltx_2.3_22b_distilled_q6p.ckpt \
  --prompt "<motion prompt>" \
  --frames 49 \
  --width 768 \
  --height 512 \
  --seed <seed> \
  --output <output.mov>

Use LTX distilled first for this rig. Keep width/height divisible by 32 and frames divisible by 8 plus 1.

Mode: Setup

Run inventory first:

uv run python skills/draw-thing/scripts/model_inventory.py --recommended-pack all --format json

Then run missing ensure_command entries after showing them to the user. Do not delete legacy models unless the user separately approves a prune operation.

Mode: Inventory

uv run python skills/draw-thing/scripts/model_inventory.py --format json

Summarize installed, missing recommended models, and redundant legacy downloads. Report incomplete_downloads and auxiliary_or_dependency_files separately so helper files or stale .part files are not mistaken for current best models.

Mode: Batch

The current CLI does not expose --batch-count. Use a shell loop and explicit output names:

for seed in 1201 1202 1203 1204; do
  draw-things-cli generate \
    --model z_image_turbo_1.0_q8p.ckpt \
    --prompt "<prompt>" \
    --steps 8 \
    --cfg 0 \
    --width 1024 \
    --height 1024 \
    --seed "$seed" \
    --output "$HOME/Pictures/draw-thing/run-$seed.png"
done

Mode: Advanced

The current generate --help does not expose old direct flags such as --controls, --loras, --mask, --upscaler, --batch-count, --sampler, or --guidance-scale. For ControlNet, LoRA, inpaint, upscale, tiling, and other app-level settings:

1. Re-check draw-things-cli generate --help.
2. Use Draw Things app-managed settings or --config-json / --config-file only when the schema is verified.
3. Never invent direct flags from older examples.

Prompt Quick-Reference

For templates and troubleshooting, load references/prompt-patterns.md.

Scaling Strategy

Output Handling

• Default skill output directory: ~/Pictures/draw-thing/.
• Create it before generation: mkdir -p "$HOME/Pictures/draw-thing".
• Always pass --output <file>; do not rely on implicit output behavior.
• Never overwrite source images; use descriptive new filenames.
• Always report output file, seed, model, steps, cfg, dimensions, and whether the model was downloaded during the run.

Error Recovery

Reference Files

Load ONE reference at a time.

Validation Contract

Run from this skill directory before declaring changes complete:

python scripts/check.py
python scripts/model_inventory.py --format json

Completion criteria:

1. scripts/check.py exits 0.
2. Model inventory identifies recommended-pack status.
3. No command examples use stale direct flags unsupported by current generate --help.
4. No portable-CLI violations remain under this skill directory.

Critical Rules

1. Always check command -v draw-things-cli before generation or model mutation.
2. Always use draw-things-cli generate --help as the source of truth for direct flags.
3. Always pass --output; preview-only output is not enough for agent workflows.
4. Use --cfg, not stale --guidance-scale.
5. Do not use stale direct flags: --sampler, --batch-count, --upscaler, --mask, --controls, or --loras unless current help/config schema proves support.
6. Prefer the recommended 2026 rig pack over legacy SDXL, SD1.5, SD2.x, SVD, FLUX.1, Qwen 1.0, Qwen Edit 2509, Wan 2.1, or old community checkpoints.
7. Never delete or prune models without separate explicit user approval.
8. Show the full command before running it.
9. Preserve source images; write new output files.
10. Report seed, model id, output path, dimensions, steps, and cfg.
11. For video, keep dimensions moderate and frame count divisible by 8 plus 1.
12. For advanced ControlNet/LoRA/inpaint/upscale workflows, verify --config-json or --config-file schema before execution.
13. Do not treat auxiliary text/VLM files, sidecars, or incomplete .part downloads as recommended generation/media models.