4.4 KiB
4.4 KiB
Qwen Image 2.1 — optimized I2I
Converts any supplied reference image into a photorealistic image of the same subject. The prompt is generic: identity, pose, clothing, accessories, colors, proportions, composition, and framing are inferred from the input rather than hard-coded.
Reusable inputs
character_reference.pngin node4— replace with the reference image.- Node
7positive and negative prompts — adapt the transformation without describing a specific subject. - Node
10seed — change for alternate results or keep fixed for comparisons. - Nodes
5and14— change the generation and output dimensions. The default is portrait 1080×1920 for person-focused images.
Optimization baseline
- Model:
qwen_image_2.1_int8_convrot.safetensors - Sampler: Euler CFG++
- Scheduler: beta
- Steps: 50
- CFG: 3.5
- Measured runtime: approximately 68.3 seconds on the RTX 3070 (Q-I2I-34; uncached runs vary)
The workflow uses the installed Qwen Image 2.1 stack and the flattened images.image_1 reference input required by the API workflow.
workflow.json— API-format workflow.
Planned optimization matrix
All rows are intentionally Untested. Keep the reference image, prompt, seed, and dimensions from ../variables.json fixed.
| ID | CFG | Steps | Sampler | Scheduler | Runtime (s) | Tested? | Current Best |
|---|---|---|---|---|---|---|---|
| Q-I2I-01 | 3.5 | 40 | euler_cfg_pp | beta | 92.1* | Yes | No |
| Q-I2I-02 | 1.0 | 24 | euler_cfg_pp | beta | 35.0 | Yes | No |
| Q-I2I-03 | 1.1 | 24 | euler_cfg_pp | beta | 37.2 | Yes | No |
| Q-I2I-04 | 1.2 | 24 | euler_cfg_pp | beta | 32.9 | Yes | No |
| Q-I2I-05 | 1.3 | 24 | euler_cfg_pp | beta | 33.9 | Yes | No |
| Q-I2I-06 | 1.4 | 24 | euler_cfg_pp | beta | 31.2 | Yes | No |
| Q-I2I-07 | 1.5 | 24 | euler_cfg_pp | beta | 31.4 | Yes | No |
| Q-I2I-08 | 1.6 | 24 | euler_cfg_pp | beta | 37.0 | Yes | No |
| Q-I2I-09 | 1.8 | 24 | euler_cfg_pp | beta | 31.1 | Yes | No |
| Q-I2I-10 | 2.0 | 24 | euler_cfg_pp | beta | 37.9 | Yes | No |
| Q-I2I-11 | 2.2 | 24 | euler_cfg_pp | beta | 31.1 | Yes | No |
| Q-I2I-12 | 2.4 | 24 | euler_cfg_pp | beta | 32.6 | Yes | No |
| Q-I2I-13 | 2.5 | 24 | euler_cfg_pp | beta | 36.8 | Yes | No |
| Q-I2I-14 | 2.7 | 24 | euler_cfg_pp | beta | 34.7 | Yes | No |
| Q-I2I-15 | 3.0 | 24 | euler_cfg_pp | beta | 33.9 | Yes | No |
| Q-I2I-16 | 3.2 | 24 | euler_cfg_pp | beta | 34.4 | Yes | No |
| Q-I2I-17 | 3.5 | 24 | euler_cfg_pp | beta | 37.6 | Yes | No |
| Q-I2I-18 | 3.8 | 24 | euler_cfg_pp | beta | 37.7 | Yes | No |
| Q-I2I-19 | 4.0 | 24 | euler_cfg_pp | beta | 31.2 | Yes | No |
| Q-I2I-20 | 3.5 | 24 | euler | beta | 37.6 | Yes | No |
| Q-I2I-21 | 3.5 | 24 | euler_ancestral | beta | 31.5 | Yes | No |
| Q-I2I-22 | 3.5 | 24 | euler_ancestral_cfg_pp | beta | 33.1 | Yes | No |
| Q-I2I-23 | 3.5 | 24 | res_multistep | beta | 32.5 | Yes | No |
| Q-I2I-24 | 3.5 | 24 | res_multistep_cfg_pp | beta | 38.7 | Yes | No |
| Q-I2I-25 | 3.5 | 24 | res_multistep_ancestral | beta | 42.8 | Yes | No |
| Q-I2I-26 | 3.5 | 24 | res_multistep_ancestral_cfg_pp | beta | 37.9 | Yes | No |
| Q-I2I-27 | 2.5 | 24 | euler_cfg_pp | simple | 36.1 | Yes | No |
| Q-I2I-28 | 2.5 | 24 | euler_cfg_pp | sgm_uniform | 37.3 | Yes | No |
| Q-I2I-29 | 2.5 | 24 | euler_cfg_pp | karras | 31.2 | Yes | No |
| Q-I2I-30 | 2.5 | 24 | euler_cfg_pp | exponential | 39.0 | Yes | No |
| Q-I2I-31 | 2.5 | 24 | euler_cfg_pp | normal | 35.2 | Yes | No |
| Q-I2I-32 | 3.5 | 30 | euler_cfg_pp | beta | 38.6 | Yes | No |
| Q-I2I-33 | 3.5 | 40 | euler_cfg_pp | beta | 58.7 | Yes | No |
| Q-I2I-34 | 3.5 | 50 | euler_cfg_pp | beta | 68.3 | Yes | Yes |
*Q-I2I-01 was a fully cached matrix replay at 0.23s; 92.1s is the uncached smoke runtime for the same settings.
Evaluation
- Current best: Q-I2I-34 — Euler CFG++, beta, CFG 3.5, 50 steps, 68.3s. Best overall balance of skin/material detail, face fidelity, clothing texture, and clean footwear.
- Q-I2I-33 was close and slightly faster at 58.7s, but Q-I2I-34 added visible fine texture without introducing a serious artifact.
- Q-I2I-01 remains a strong reference baseline, but its reported 0.23s matrix runtime was cached and is not comparable.
- Q-I2I-20–23 were valid alternatives; the non-CFG++/ancestral variants did not improve fidelity. Q-I2I-24–26 were rejected as corrupted outputs.
- Q-I2I-27–31 were valid scheduler comparisons, but none surpassed beta. Q-I2I-02–19 showed no better combination than the CFG 3.5/high-step result.