/ollama-image

/ollama-image

Generate an image and report the file path. When --iterations is greater than 1, run an iterative review-refine loop using your native vision to evaluate each image and improve the prompt between iterations.

Consult the ollama-image-expert agent for all domain knowledge: scoring rubric, prompt engineering strategies, mutation rules, and convergence criteria.

Parse Arguments

Parse $ARGUMENTS for:

• Prompt: The quoted text description, OR use --prompt-file PATH to read a prompt from a file
• --model MODEL: Ollama model to use (must be an image generation model with x/ prefix)
• --output PATH: Where to save the image (default: output/YYYYMMDD-HHMMSS.png)
• --iterations N: Max refinement iterations (default: 1 = one-shot, no review)
• --width INT: Image width (default: 1024)
• --height INT: Image height (default: 1024)
• --steps INT: Inference steps (default: 8 for z-image-turbo, 20 for flux models)
• --seed INT: Seed for reproducibility (optional)

If --prompt-file is specified, read the file contents as the prompt using the Read tool.

If no prompt and no --prompt-file is provided, stop and tell the user to provide a prompt.

Save the original prompt — you will need it if a full rewrite is required during refinement.

Discover Available Models

If no --model is specified, or to validate the user's model choice, discover available image generation models:

ollama list

Filter the output for models with the x/ prefix — these are image generation models. Examples: x/z-image-turbo, x/flux2-klein.

• If --model is specified: verify it appears in the list. If not, tell the user it's not available, show them the available x/ models, and suggest ollama pull MODEL.
• If --model is NOT specified: default to x/z-image-turbo if available. If not, use the first available x/ model. If NO x/ models are available, tell the user to pull one: ollama pull x/z-image-turbo.

Set Model-Appropriate Defaults

If --steps was not explicitly provided, set defaults based on the model:

• x/z-image-turbo: 8 steps (timeout auto-set to 120s)
• x/flux2-klein or other flux models: 20 steps (timeout auto-set to 600s)

The helper script auto-detects timeouts per model. You do NOT need to pass --timeout unless the user explicitly requests it.

Locate Plugin

Before running any Python scripts, discover the plugin root:

PLUGIN_ROOT=$(python3 -c "
from pathlib import Path
import sys, os

if os.environ.get('JACK_TAR_OLLAMA_ROOT'):
    print(os.environ['JACK_TAR_OLLAMA_ROOT']); sys.exit()

home = Path.home()
for base in [home / '.claude' / 'plugins' / 'cache']:
    for p in base.rglob('jack-tar-ollama/.claude-plugin/plugin.json'):
        print(str(p.parent.parent)); sys.exit()

dev = Path.cwd() / 'plugins' / 'jack-tar-ollama'
if dev.exists():
    print(str(dev)); sys.exit()

print('NOT_FOUND')
" 2>/dev/null)
if [ -z "$PLUGIN_ROOT" ] || [ "$PLUGIN_ROOT" = "NOT_FOUND" ]; then
  echo "ERROR: jack-tar-ollama plugin not found. Set JACK_TAR_OLLAMA_ROOT or install the plugin."
  exit 1
fi

Verify Ollama Is Running

curl -s http://localhost:11434/api/tags

If this fails, tell the user: "Ollama is not running. Start it with: ollama serve" and stop.

One-Shot Mode (--iterations 1 or omitted)

If iterations is 1 (the default), run a single generation with no review:

1. Run the helper script:

   python3 "$PLUGIN_ROOT/src/generate_image.py" --prompt "THE PROMPT" --model "THE MODEL" --output "THE PATH" --width WIDTH --height HEIGHT --steps STEPS [--seed SEED]
   

2. If exit code is 0: read the output path from stdout.
3. If exit code is non-zero: read stderr and report the error.
4. Report the file path, model, and prompt. Done.

Iterative Mode (--iterations > 1)

Run the generate-review-refine loop. Track iteration state as you go:

• Iteration history: list of (iteration number, prompt used, image path, dimension scores, weighted score)
• Best iteration: the one with the highest weighted score so far
• Current prompt: starts as the user's prompt, refined each iteration

For each iteration (1 to max iterations):

Step 1: Generate

Determine the output path for this iteration. If --output was specified, use it for iteration 1 and append -iter-N for subsequent iterations (e.g., fox.png, fox-iter-2.png, fox-iter-3.png). If no --output, use output/YYYYMMDD-HHMMSS-iter-N.png.

Run the helper script with the current prompt:

python3 "$PLUGIN_ROOT/src/generate_image.py" --prompt "CURRENT PROMPT" --model "MODEL" --output "ITER PATH" --width WIDTH --height HEIGHT --steps STEPS

If the helper fails, report the error and stop.

Step 2: Review

Read the generated image using the Read tool — you can see it natively.

Evaluate the image against the original user prompt (not the refined prompt) using the scoring rubric from the ollama-image-expert agent. Score each of the 6 dimensions (1-10):

1. Subject Accuracy (30%)
2. Style Adherence (20%)
3. Composition (15%)
4. Technical Quality (15%)
5. Color and Lighting (10%)
6. Overall Holistic (10%)

Calculate the weighted score. Be honest and critical.

Record this iteration in your history.

Step 3: Decide

Check the convergence rules from the ollama-image-expert agent, in this priority order:

1. Accept: weighted score >= 7.5 → stop, report this image
2. Max iterations: this is the last iteration → stop, report the best image from history
3. Plateau: score has not improved by > 0.3 for the last 2 consecutive iterations → stop early, report best
4. Oscillation: scores bouncing with no upward trend → stop early, report best
5. Regression: current score is lower than the best prior → revert to the best prior prompt and apply minor additive changes only (do NOT continue with the degraded prompt)
6. Continue: none of the above → proceed to refinement

Step 4: Refine

Identify the weakest-scoring dimension. Apply the mutation strategy from the ollama-image-expert agent based on the current score band:

• 7.0-7.9 (Additive): Add lighting keywords, camera specs, mood descriptors. Do not restructure.
• 5.0-6.9 (Structural): Reorder elements, add/remove descriptive blocks, clarify spatial relationships.
• 1.0-4.9 (Rewrite): Start fresh from the original user prompt. Use iteration history to avoid repeating failed approaches.

Apply model-specific prompt structure from the agent (Z-Image Turbo structure vs FLUX spatial hierarchy).

The refined prompt becomes the current prompt for the next iteration. Loop back to Step 1.

Report Result

On completion (whether one-shot or iterative), report:

For one-shot mode:

• The absolute file path to the generated image
• The model used
• The prompt used

For iterative mode:

• The absolute file path to the best-scoring image (not necessarily the last one)
• Its weighted score and dimension breakdown
• The model used
• The prompt that produced the best image
• A brief iteration summary: how many iterations ran, scores per iteration, why iteration stopped (accepted, plateau, oscillation, max reached)

Do not ask follow-up questions. Do not offer to continue refining. Report and stop.