Add like-for-like FLUX refiner comparison

This commit is contained in:
Morpheus
2026-10-29 08:57:41 +00:00
parent 45f13b0d86
commit 950a276dca
6 changed files with 97 additions and 1 deletions
+5 -1
View File
@@ -19,10 +19,14 @@ text prompt, so it cannot replace the subject from text conditioning.
| 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](results/QRF-I2I-02-base.png) vs [refined PNG](results/QRF-I2I-02.png) |
| QRF-I2I-03 | `refiner_reference.jpg` | Canonical shared I2I prompt | Qwen Image 2.1, 8 steps, CFG 1, Euler ancestral / beta | FLUX.3 Klein, 8 steps, guidance 1, Euler ancestral / simple, denoise 0.11 | Base 1080×1920; refined 1072×1920 | 60.252 | Completed | Realistic low-step base; white top/bottom bands; [base PNG](results/QRF-I2I-03-base.png) vs [refined PNG](results/QRF-I2I-03.png) |
| ZRF-I2I-01 | `refiner_reference.jpg` | Canonical shared I2I prompt | Qwen Image 2.1, 8 steps, CFG 1, Euler ancestral / beta | ZImage Turbo, 5 steps, CFG 1, DPM++ 2M SDE / beta, denoise 0.33, shift 3 | Base/refined 1080×1920; refiner internal canvas 1088×1920 | 33.571 | Completed | Clean refiner output preserves the subject and clothing; [base PNG](results/ZRF-I2I-01-base.png) vs [refined PNG](results/ZRF-I2I-01.png) |
| QRF-I2I-02-LF | `refiner_reference.jpg` | Canonical shared I2I prompt | Qwen Image 2.1, 8 steps, CFG 1, Euler ancestral / beta | FLUX.3 Klein, 8 steps, guidance 1, Euler ancestral / simple, denoise 0.11 | Base 1080×1920; refined 1072×1920 | 95.397 | Completed | Controlled like-for-like FLUX comparison against ZRF-I2I-01; [base PNG](results/QRF-I2I-02-LF-base.png) vs [refined PNG](results/QRF-I2I-02-LF.png) |
## Planned follow-ups
- Additional refiner models only after the first comparison is recorded.
- Additional refiner models only after the like-for-like comparison is reviewed.
`QRF-I2I-02` remains the historical 50-step/CFG 3.5 run. `QRF-I2I-02-LF`
is the controlled rerun with its Qwen base settings matched to `ZRF-I2I-01`.
The FLUX workflow is `workflow.json`; the ZImage variant is
`zimage-refiner-workflow.json`. Each resolved run records its prompt ID, model
@@ -0,0 +1,30 @@
{
"1": {"class_type":"UNETLoader","inputs":{"unet_name":"qwen_image_2.1_int8_convrot.safetensors","weight_dtype":"default"}},
"2": {"class_type":"CLIPLoader","inputs":{"clip_name":"qwen3vl_8b_w4a8.safetensors","type":"qwen_image","device":"default"}},
"3": {"class_type":"VAELoader","inputs":{"vae_name":"qwen_image_2.1_vae_bf16.safetensors"}},
"4": {"class_type":"LoadImage","inputs":{"image":"refiner_reference.jpg"}},
"5": {"class_type":"ResizeAndPadImage","inputs":{"image":["4",0],"target_width":576,"target_height":1024,"padding_color":"white","interpolation":"lanczos"}},
"6": {"class_type":"QwenImage21Cache","inputs":{"model":["1",0],"device":"auto","dtype":"int8"}},
"7": {"class_type":"TextEncodeQwenImage21","inputs":{"clip":["2",0],"prompt":"Transform the supplied reference image into a photorealistic image of the same subject. Infer identity, facial features, skin tone, hairstyle, clothing, accessories, colors, pose, body proportions, composition, and framing directly from the reference. Preserve those visual attributes faithfully while replacing the original rendering style with natural human anatomy, realistic skin and hair texture, physically accurate materials, believable fabric and surface detail, natural hands and fingers, cinematic soft light, realistic shadows, coherent perspective, premium editorial photography, high dynamic range, photorealistic, highly detailed. Do not redesign, restyle, substitute, add, remove, mirror, or crop the subject. No text or logos.","negative_prompt":"different person, changed identity, changed face, changed hairstyle, changed clothing, changed colors, altered accessories, changed pose, different composition, extra person, duplicate person, extra limbs, malformed hands, fused fingers, distorted anatomy, cropped head, cropped feet, mirrored image, cartoon, illustration, anime, CGI, plastic skin, waxy skin, blurry, low detail, text, logo, watermark","resolution":0,"images.image_1":["5",0],"vae":["3",0]}},
"8": {"class_type":"KSamplerSelect","inputs":{"sampler_name":"euler_ancestral"}},
"9": {"class_type":"BasicScheduler","inputs":{"model":["6",0],"scheduler":"beta","steps":8,"denoise":1.0}},
"10": {"class_type":"RandomNoise","inputs":{"noise_seed":310101}},
"11": {"class_type":"CFGGuider","inputs":{"model":["6",0],"positive":["7",0],"negative":["7",1],"cfg":1.0}},
"12": {"class_type":"SamplerCustomAdvanced","inputs":{"noise":["10",0],"guider":["11",0],"sampler":["8",0],"sigmas":["9",0],"latent_image":["7",2]}},
"13": {"class_type":"VAEDecode","inputs":{"samples":["12",0],"vae":["3",0]}},
"28": {"class_type":"ResizeAndPadImage","inputs":{"image":["13",0],"target_width":1080,"target_height":1920,"padding_color":"white","interpolation":"lanczos"}},
"14": {"class_type":"UNETLoader","inputs":{"unet_name":"flux3-image-klein-int8-convrot.safetensors","weight_dtype":"default"}},
"15": {"class_type":"CLIPLoader","inputs":{"clip_name":"qwen3-vl-4b-heretic_int8.safetensors","type":"flux3","device":"default"}},
"16": {"class_type":"VAELoader","inputs":{"vae_name":"flux2-vae.safetensors"}},
"17": {"class_type":"StarFlux2Conditioner","inputs":{"clip":["15",0],"vae":["16",0],"text":"","join_references":true,"image_1":["28",0]}},
"18": {"class_type":"FluxGuidance","inputs":{"conditioning":["17",0],"guidance":1.0}},
"19": {"class_type":"BasicGuider","inputs":{"model":["14",0],"conditioning":["18",0]}},
"20": {"class_type":"KSamplerSelect","inputs":{"sampler_name":"euler_ancestral"}},
"21": {"class_type":"BasicScheduler","inputs":{"model":["14",0],"scheduler":"simple","steps":8,"denoise":0.11}},
"22": {"class_type":"RandomNoise","inputs":{"noise_seed":310101}},
"23": {"class_type":"VAEEncode","inputs":{"pixels":["28",0],"vae":["16",0]}},
"24": {"class_type":"SamplerCustomAdvanced","inputs":{"noise":["22",0],"guider":["19",0],"sampler":["20",0],"sigmas":["21",0],"latent_image":["23",0]}},
"25": {"class_type":"VAEDecode","inputs":{"samples":["24",0],"vae":["16",0]}},
"27": {"class_type":"SaveImage","inputs":{"images":["28",0],"filename_prefix":"fluxExperiments/i2i_refiner/QRF-I2I-02-LF-base"}},
"26": {"class_type":"SaveImage","inputs":{"images":["25",0],"filename_prefix":"fluxExperiments/i2i_refiner/QRF-I2I-02-LF"}}
}
Binary file not shown.

After

Width:  |  Height:  |  Size: 2.1 MiB

@@ -0,0 +1,50 @@
{
"test_id": "QRF-I2I-02-LF",
"status": "completed",
"comparison_parent": "QRF-I2I-02",
"prompt_id": "b376958b-1fef-4efc-a597-d8c59a51a538",
"server": "http://192.168.1.202:8188",
"input_image": "refiner_reference.jpg",
"input_image_sha256": "f478ab8064bd0642401cbf6bab9aa95380ed4c05d50c6d2959bc9a0eb632354a",
"prompt_source": "../I2I/qwen/workflow.json#7",
"prompt_used_by_base": true,
"prompt_used_by_refiner": false,
"base_stage": {
"model": "qwen_image_2.1_int8_convrot.safetensors",
"text_encoder": "qwen3vl_8b_w4a8.safetensors",
"vae": "qwen_image_2.1_vae_bf16.safetensors",
"generation_width": 576,
"generation_height": 1024,
"steps": 8,
"cfg": 1.0,
"sampler": "euler_ancestral",
"scheduler": "beta",
"denoise": 1.0
},
"refiner_stage": {
"model": "flux3-image-klein-int8-convrot.safetensors",
"text_encoder": "qwen3-vl-4b-heretic_int8.safetensors",
"vae": "flux2-vae.safetensors",
"steps": 8,
"guidance": 1.0,
"sampler": "euler_ancestral",
"scheduler": "simple",
"denoise": 0.11
},
"generation_time_seconds": 95.397,
"outputs": {
"base": {
"path": "results/QRF-I2I-02-LF-base.png",
"width": 1080,
"height": 1920,
"sha256": "a34d08a1712b9e3f9fa0fc421bc37ebca49dd157706fcce59e5cc7c6f0675882"
},
"refined": {
"path": "results/QRF-I2I-02-LF.png",
"width": 1072,
"height": 1920,
"sha256": "70ad9095027a397b6366c0de8259fe3a5ed28af061d68cfcef54e87f24554781"
}
},
"review": "Controlled like-for-like FLUX comparison against ZRF-I2I-01. The Qwen base settings now match ZRF-I2I-01 exactly; only the refiner model and refiner-specific settings differ. The FLUX output preserves the subject and is clean, but its 0.11 denoise setting makes the visual change more conservative than ZImage at denoise 0.33."
}
Binary file not shown.

After

Width:  |  Height:  |  Size: 2.0 MiB

+12
View File
@@ -80,6 +80,18 @@
},
"metadata": "results/ZRF-I2I-01.json",
"refiner_workflow": "zimage-refiner-workflow.json"
},
{
"id": "QRF-I2I-02-LF",
"status": "completed",
"generation_time_seconds": 95.397,
"comparison_parent": "QRF-I2I-02",
"outputs": {
"base": "results/QRF-I2I-02-LF-base.png",
"refined": "results/QRF-I2I-02-LF.png"
},
"metadata": "results/QRF-I2I-02-LF.json",
"workflow": "qrf-like-for-like-workflow.json"
}
]
}