Add Qwen multi-reference image tests

This commit is contained in:
Morpheus
2026-10-28 17:14:53 +00:00
parent 678e146d71
commit a60e80925f
5 changed files with 106 additions and 3 deletions
+50
View File
@@ -0,0 +1,50 @@
# Qwen Image 2.1 — MultipleRef2I
This is the Qwen two-reference image-to-image plan. It combines the fixed
person and scene fixtures used by the other MultipleRef2I workflows. The
person reference supplies identity and appearance; the scene reference
supplies environment, lighting, perspective, and composition.
## Workflow and fixed inputs
- `workflow.json` — API-format workflow using two `TextEncodeQwenImage21` image inputs.
- Model: `qwen_image_2.1_int8_convrot.safetensors`
- Text encoder: `qwen3vl_8b_w4a8.safetensors` (`type=qwen_image`)
- VAE: `qwen_image_2.1_vae_bf16.safetensors`
- References: `../person_reference.png` and `../scene_reference.png`
- Output comparison target: **1920×1080**
- Seed: `310201`
The live node contract supports `image_1` through `image_16`, so this is a
real two-reference graph rather than a duplicated single-reference test.
## Focused matrix
All 18 rows are currently **Planned**. Keep both references, the prompt,
resolution, and seed fixed; vary only the explicitly listed CFG, sampler,
scheduler, or steps.
| ID | CFG | Steps | Sampler | Scheduler | Status | Purpose |
|---|---:|---:|---|---|---|---|
| Q-MR2I-01 | 3.5 | 50 | euler_cfg_pp | beta | Planned | Baseline |
| Q-MR2I-02 | 1.0 | 24 | euler_cfg_pp | beta | Planned | Low CFG |
| Q-MR2I-03 | 2.5 | 24 | euler_cfg_pp | beta | Planned | CFG comparison |
| Q-MR2I-04 | 3.5 | 24 | euler_cfg_pp | beta | Planned | CFG/steps comparison |
| Q-MR2I-05 | 4.0 | 24 | euler_cfg_pp | beta | Planned | Higher CFG |
| Q-MR2I-06 | 3.5 | 24 | euler | beta | Planned | Euler sampler |
| Q-MR2I-07 | 3.5 | 24 | euler_ancestral | beta | Planned | Ancestral sampler |
| Q-MR2I-08 | 3.5 | 24 | res_multistep | beta | Planned | RES sampler |
| Q-MR2I-09 | 3.5 | 24 | res_multistep_cfg_pp | beta | Planned | RES CFG++ sampler |
| Q-MR2I-10 | 2.5 | 24 | euler_cfg_pp | simple | Planned | Scheduler comparison |
| Q-MR2I-11 | 2.5 | 24 | euler_cfg_pp | sgm_uniform | Planned | Scheduler comparison |
| Q-MR2I-12 | 2.5 | 24 | euler_cfg_pp | karras | Planned | Scheduler comparison |
| Q-MR2I-13 | 2.5 | 24 | euler_cfg_pp | exponential | Planned | Scheduler comparison |
| Q-MR2I-14 | 3.5 | 30 | euler_cfg_pp | beta | Planned | Step-count comparison |
| Q-MR2I-15 | 3.5 | 40 | euler_cfg_pp | beta | Planned | Step-count comparison |
| Q-MR2I-16 | 3.5 | 60 | euler_cfg_pp | beta | Planned | Step-count comparison |
| Q-MR2I-17 | 3.5 | 75 | euler_cfg_pp | beta | Planned | Step-count comparison |
| Q-MR2I-18 | 3.5 | 50 | euler_ancestral | beta | Planned | Baseline sampler alternative |
Record both reference hashes, runtime, output dimensions/path, exact settings,
and visual notes for every run. Do not reduce the reference count silently if
the graph or node contract rejects the pair.