sunnyJane/data266-lab2-p3-outputs

DATA 266 Lab 2, Part 3: multi-agent system outputs (Pair_Programming_Team_20)

Images produced by our five-agent image understanding and generation system (Planner, Vision Understanding, Prompt Engineering, Image Generation, Critique) for DATA 266 Lab 2 at San Jose State University, Fall 2026. They are the outputs that both team members judged against each scenario's pass/fail rubric.

Two evaluation runs

Folder	Run	Scenarios	What it contains
final_run/	Final evaluation (reported results)	15 fresh scenarios p3_scenarios_v2.csv (N01-N15) and the 10 red-team cases p3_redteam.csv	The pre-fix system and the final system, both with the critique loop, plus the final system without the critique loop
repository root (*.png)	First (development) evaluation	15 scenarios p3_scenarios.csv (S01-S15) and the 10 red-team cases	The original system with and without the critique loop

The first evaluation showed that style changes and edits barely changed the input image: the generation prompt described the photo instead of the target style, and every revision used the same image-to-image strength. We fixed this after judging that run, so the final system was evaluated on 15 new scenarios that were written and committed to GitHub before the final run. The first run is kept here for transparency and reproducibility.

File naming

<request_id>_r<k>.png

• request_id is the 8-character id of one request; the same id appears in the run-results CSV and in the JSON trace (traces/<request_id>.json) in our GitHub repository.
• k is the attempt number: r0 is the first generation; r1 and r2 are the critic-requested revisions (at most 2). When no attempt is accepted, the system returns the attempt with the highest critic score (best_effort).

To find which scenario and system variant an image belongs to, use final_run/p3_v2_run_results.csv (final run) or p3_run_results.csv in the repository root (first run; also results/p3/ in our GitHub repository). Both list the scenario id, whether the critique loop was used, the status, the first and final output paths, and the seconds per request.

Requests that only need a text answer (captions and visual questions, refusals, clarification requests) produce no image.

Models

• Planner, Vision Understanding, Prompt Engineering, Critique: Qwen/Qwen2.5-VL-7B-Instruct (4-bit NF4), the same vision-language model as Part 1.
• Style change and domain style: stable-diffusion-v1-5/stable-diffusion-v1-5 image-to-image / text-to-image; domain-style requests load our Part 2 LoRA sunnyJane/data266-lab2-naruto-run2-lora-r32.
• Edits and enhancements: timbrooks/instruct-pix2pix.
• Critique scoring: CLIP similarity to the prompt, a short VLM judgement and the Stable Diffusion safety checker.
• Every image was generated with a fixed seed; model, prompt, seed and steps are recorded in the JSON traces.

Inputs

The input photos and scenario files are in sunnyJane/data266-lab2-p3-inputs. All photos are from Pexels (free to use); sources are listed in eval/image_sources.txt in our GitHub repository. Red-team inputs include images with written instructions, real-person photos (for edit-refusal tests) and broken or empty files.

Intended use and limitations

• Coursework evaluation only; non-commercial use.
• Domain-style images imitate the style of Naruto, a copyrighted franchise; the Part 2 LoRA and its outputs are for non-commercial research only.
• The images show the system's real behaviour, including failures (for example requests that were not changed enough, or an ambiguous request that was not clarified). They are not curated examples.