I2I plus refiner tests
This section contains image-to-image workflows where an edit model creates the first image and a separate refiner improves its surface quality, texture, and photorealism. Each run saves both the base-stage image and the refined image. The input image and canonical I2I prompt remain fixed when comparing refiner models.
Current test
The first test used Qwen NoctQ v4. The follow-up uses the proven Qwen Image 2.1
I2I stack and settings from optimized/I2I/qwen/, with the new reference image
refiner_reference.jpg. The FLUX stage receives the Qwen pixels and has no
text prompt, so it cannot replace the subject from text conditioning.
| ID | Input | Prompt | Base stage | Refiner stage | Resolution | Generation time (s) | Status | Review |
|---|---|---|---|---|---|---|---|---|
| QRF-I2I-01 | refiner_reference.jpg |
Canonical shared I2I prompt | NoctQ v4, 8 steps, CFG 1, Euler ancestral / beta | FLUX.3 Klein, 8 steps, guidance 1, Euler ancestral / simple, denoise 0.11 | 1088×1920 | 115.720 uncached; 0.274 cached | Completed | Base and refined outputs are nearly identical at denoise 0.11; base PNG vs refined PNG |
| QRF-I2I-02 | refiner_reference.jpg |
Canonical shared I2I prompt | Qwen Image 2.1, 50 steps, CFG 3.5, Euler CFG++ / beta | FLUX.3 Klein, 8 steps, guidance 1, Euler ancestral / simple, denoise 0.11 | Base 1080×1920; refined 1072×1920 | 145.129 | Completed | More realistic base; base PNG vs refined PNG |
Planned follow-ups
- ZImage as the refiner, using the same base output, prompt, input image, and resolution.
- Additional refiner models only after the first comparison is recorded.
Record the resolved workflow, prompt ID, model names, runtime, dimensions,
hashes, and visual review in results/QRF-I2I-01.json.