Promote Qwen res multistep baseline
This commit is contained in:
@@ -13,6 +13,6 @@ This summary is maintained across the model-specific test plans. A result is onl
|
||||
| Model | Best so far | Runtime | Generated image | Status |
|
||||
|---|---|---:|---|---|
|
||||
| Flux.2 Klein | F-T2I-03 | 220.3 s | [F-T2I-03.png](flux/results/F-T2I-03.png) | Current best; 1080×1920 target |
|
||||
| Qwen Image 2.1 | Q-T2I-49 | 110.6 s | [Q-T2I-49.png](qwen/results/Q-T2I-49.png) | Current best; direct 1088×1920 latent; euler_cfg_pp, beta, 25 steps, CFG 2.5 |
|
||||
| Qwen Image 2.1 | Q-T2I-57 | 110.2 s | [Q-T2I-57.png](qwen/results/Q-T2I-57.png) | Current best; direct 1088×1920 latent; res_multistep, beta, 25 steps, CFG 2.5 |
|
||||
|
||||
Store each T2I test under the selected model's `results/<test-id>.png` and `results/<test-id>.json` using the table ID as the label.
|
||||
|
||||
@@ -9,12 +9,12 @@ Qwen Image 2.1 text-to-image workflow using the best tested sampler, scheduler,
|
||||
- VAE: qwen_image_2.1_vae_bf16.safetensors
|
||||
- Internal generation: 1088×1920 portrait latent (generated directly at target resolution)
|
||||
- Final output: 1088×1920, 9:16, decoded directly from the target-resolution latent
|
||||
- Sampler: euler_cfg_pp
|
||||
- Sampler: res_multistep
|
||||
- Scheduler: beta
|
||||
- Steps: 25
|
||||
- CFG: 2.5
|
||||
- Seed: 20260920
|
||||
- Runtime: recorded per test in the matrix; previous 1024-square baseline invalidated by this workflow update
|
||||
- Runtime: 110.2s for the current best; recorded per test in the matrix
|
||||
|
||||
For reuse, replace the positive and negative prompts, seed, and target resolution. The latent and decoded image now use the same 1088×1920 dimensions.
|
||||
|
||||
@@ -72,7 +72,7 @@ Keep the prompt, seed, and resolution from [`../variables.json`](../variables.js
|
||||
| Q-T2I-46 | 2.5 | 25 | res_multistep_cfg_pp | simple | 110.0 | Yes | No |
|
||||
| Q-T2I-47 | 2.5 | 25 | res_multistep_ancestral | simple | 111.4 | Yes | No |
|
||||
| Q-T2I-48 | 2.5 | 25 | res_multistep_ancestral_cfg_pp | simple | 111.1 | Yes | No |
|
||||
| Q-T2I-49 | 2.5 | 25 | euler_cfg_pp | beta | 110.6 | Yes | Yes |
|
||||
| Q-T2I-49 | 2.5 | 25 | euler_cfg_pp | beta | 110.6 | Yes | No |
|
||||
| Q-T2I-50 | 2.5 | 25 | euler_cfg_pp | sgm_uniform | 109.8 | Yes | No |
|
||||
| Q-T2I-51 | 2.5 | 25 | euler_cfg_pp | karras | 110.9 | Yes | No |
|
||||
| Q-T2I-52 | 2.5 | 25 | euler_cfg_pp | exponential | 113.2 | Yes | No |
|
||||
@@ -81,7 +81,7 @@ Keep the prompt, seed, and resolution from [`../variables.json`](../variables.js
|
||||
| Q-T2I-54 | 2.5 | 25 | euler | beta | 145.6 | Yes | No |
|
||||
| Q-T2I-55 | 2.5 | 25 | euler_ancestral | beta | 107.7 | Yes | No |
|
||||
| Q-T2I-56 | 2.5 | 25 | euler_ancestral_cfg_pp | beta | 109.6 | Yes | No |
|
||||
| Q-T2I-57 | 2.5 | 25 | res_multistep | beta | 110.2 | Yes | No |
|
||||
| Q-T2I-57 | 2.5 | 25 | res_multistep | beta | 110.2 | Yes | Yes |
|
||||
| Q-T2I-58 | 2.5 | 25 | res_multistep_cfg_pp | beta | 109.4 | Yes | No |
|
||||
| Q-T2I-59 | 2.5 | 25 | res_multistep_ancestral | beta | 109.4 | Yes | No |
|
||||
| Q-T2I-60 | 2.5 | 25 | res_multistep_ancestral_cfg_pp | beta | 110.8 | Yes | No |
|
||||
@@ -120,7 +120,7 @@ Files:
|
||||
|
||||
Visual evaluation notes:
|
||||
- Full matrix status: 88/88 tested; 35/35 newly collected in this batch.
|
||||
- Current best remains Q-T2I-49: `euler_cfg_pp`, beta, 25 steps, CFG 2.5, 110.6s.
|
||||
- Current best: Q-T2I-57: `res_multistep`, beta, 25 steps, CFG 2.5, 110.2s. User review rated it highest so far for skin texture and foot anatomy.
|
||||
- Strong new alternatives, but not promoted: Q-T2I-54–57 and Q-T2I-61–64. Q-T2I-82 and Q-T2I-84 initially appeared strong technically, but were rejected after review for visibly deformed legs/feet.
|
||||
- Valid but weaker: Q-T2I-85.
|
||||
- Rejected for anatomy, severe corruption, or unusable exposure: Q-T2I-82, Q-T2I-84, Q-T2I-58–60, Q-T2I-65–81, and Q-T2I-86–88.
|
||||
|
||||
@@ -18,5 +18,5 @@
|
||||
"output_sha256": "cf536e0c0c052b874807252525214ce686b2a07aa1226719b79c73c6163081c1",
|
||||
"verdict": "Better than the previous Qwen best after full-batch visual comparison: strongest balance of photographic detail, lighting, stadium context, subject clarity, and valid anatomy.",
|
||||
"better_than_current_best": true,
|
||||
"current_best": true
|
||||
"current_best": false
|
||||
}
|
||||
|
||||
@@ -15,6 +15,9 @@
|
||||
"runtime_seconds": 110.17,
|
||||
"output_file": "Q-T2I-57.png",
|
||||
"status": "completed",
|
||||
"verdict": "Promoted by user review as the best Qwen result so far: visibly improved skin texture and more natural foot anatomy.",
|
||||
"better_than_current_best": true,
|
||||
"current_best": true,
|
||||
"comfy_prompt_id": "8ddd19cb-ac3e-4a9b-b931-b3e0ce8482aa",
|
||||
"source_filename": "Q-T2I-57_00002_.png",
|
||||
"output_sha256": "96c49d3fba450ffacb8853955d3831c8aa7d360bc0c87955af4708a0bef0c774"
|
||||
|
||||
@@ -219,7 +219,7 @@
|
||||
"Node name for S&R": "KSamplerSelect"
|
||||
},
|
||||
"widgets_values": [
|
||||
"euler_cfg_pp"
|
||||
"res_multistep"
|
||||
]
|
||||
},
|
||||
{
|
||||
|
||||
@@ -50,7 +50,7 @@
|
||||
"6": {
|
||||
"class_type": "KSamplerSelect",
|
||||
"inputs": {
|
||||
"sampler_name": "euler_cfg_pp"
|
||||
"sampler_name": "res_multistep"
|
||||
}
|
||||
},
|
||||
"7": {
|
||||
|
||||
Reference in New Issue
Block a user