Add generic prompt experiment structure

This commit is contained in:
Morpheus
2026-10-09 22:14:09 +00:00
parent 787ecd8b08
commit fd4d2e0459
7 changed files with 142 additions and 0 deletions
@@ -0,0 +1,58 @@
# Person to clothing flat lay
## Purpose
Use a reference image containing a person and reconstruct every visible item
of clothing and accessory as a premium fashion product flat lay. The output
must contain **no person, body parts, mannequin, model, hanger, or human
silhouette**—only the garments and accessories arranged as a coherent set.
## Baseline
This prompt starts from the current best Qwen Image 2.1 I2I workflow:
- Source baseline: [`optimized/I2I/qwen/`](../../optimized/I2I/qwen/)
- Model: Qwen Image 2.1 INT8 ConvRot
- Sampler: Euler CFG++
- Scheduler: beta
- Steps: 50
- CFG: 3.5
- Output: portrait 1080×1920 baseline, adjustable for composition
- Reference input: one image containing the person and outfit
The copied workflow files preserve this starting point. The optimized Qwen
workflow remains the source of truth for future baseline changes.
## Expected result
The image should look like a carefully art-directed fashion editorial or
premium e-commerce catalogue photograph. It should identify and lay out every
visible garment and accessory, preserve material, colour, pattern, construction
details and relative styling, and keep each item separated enough to be read.
Suggested arrangement: a clean neutral surface, garments folded or naturally
laid flat in neat groups, accessories beside the relevant outfit, soft
directional studio lighting, realistic fabric texture, subtle contact shadows,
and a balanced composition with no clutter.
## Input assumptions
- Clothing and accessories should be inferred only from what is visible.
- Hidden garments, unseen backs, and speculative details should not be invented.
- If multiple people are present, show their visible clothing as separate sets
while keeping the output people-free.
## Iteration process
1. Replace the reference image in the Qwen workflow.
2. Use `prompt.txt` as the positive prompt and keep baseline settings fixed.
3. Save each generated image and exact metadata in `results/`.
4. Record outfit coverage, human-removal quality, material fidelity, and layout.
5. Change the prompt only when the visual result shows a repeatable weakness.
6. Mark the accepted prompt revision and current best result here.
## Current status
- Status: scaffolded; first generation pending.
- Current best result: none yet.
- First test: baseline Qwen workflow with `prompt.txt`.