diff --git a/optimized/T2I/qwen/README.md b/optimized/T2I/qwen/README.md index 4faf9d8..ce45745 100644 --- a/optimized/T2I/qwen/README.md +++ b/optimized/T2I/qwen/README.md @@ -7,16 +7,16 @@ Qwen Image 2.1 text-to-image workflow using the best tested sampler, scheduler, - Model: qwen_image_2.1_int8_convrot.safetensors - Text encoder: qwen3vl_8b_w4a8.safetensors - VAE: qwen_image_2.1_vae_bf16.safetensors -- Internal generation: 1024×1024 square (the Qwen 2.1 text encoder provides a square latent) -- Final output: 1080×1920, 9:16, center-cropped with Lanczos +- 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 - Scheduler: simple - Steps: 12 - CFG: 2.5 - Seed: 20260920 -- First portrait test generation time: 198.1 seconds on the server RTX 3070 +- Runtime: recorded per test in the matrix; previous 1024-square baseline invalidated by this workflow update -For reuse, replace the positive and negative prompts, seed, and output dimensions. The final crop removes the white side bands produced by padding the square Qwen output. +For reuse, replace the positive and negative prompts, seed, and target resolution. The latent and decoded image now use the same 1088×1920 dimensions. ## Planned optimization matrix @@ -24,7 +24,7 @@ All rows are intentionally **Untested**. Keep the prompt, seed, and resolution f | ID | CFG | Steps | Sampler | Scheduler | Runtime (s) | Tested? | Current Best | |---|---:|---:|---|---|---:|---|---| -| Q-T2I-01 | 2.5 | 12 | euler_cfg_pp | simple | 198.1 | Yes | Yes | +| Q-T2I-01 | 2.5 | 12 | euler_cfg_pp | simple | — | No | No | | Q-T2I-02 | 1.0 | 12 | euler_cfg_pp | simple | — | No | No | | Q-T2I-03 | 1.1 | 12 | euler_cfg_pp | simple | — | No | No | | Q-T2I-04 | 1.2 | 12 | euler_cfg_pp | simple | — | No | No | diff --git a/optimized/T2I/qwen/ui-workflow.json b/optimized/T2I/qwen/ui-workflow.json index d8026fe..979e243 100644 --- a/optimized/T2I/qwen/ui-workflow.json +++ b/optimized/T2I/qwen/ui-workflow.json @@ -1,8 +1,8 @@ { "id": "qwen-image-2.1-spaceship-canteen-ui", "revision": 0, - "last_node_id": 13, - "last_link_id": 17, + "last_node_id": 16, + "last_link_id": 16, "nodes": [ { "id": 1, @@ -180,13 +180,6 @@ "links": [ 8 ] - }, - { - "name": "latent", - "type": "LATENT", - "links": [ - 9 - ] } ], "properties": { @@ -384,7 +377,7 @@ { "name": "latent_image", "type": "LATENT", - "link": 9 + "link": 16 } ], "outputs": [ @@ -453,13 +446,13 @@ 110 ], "flags": {}, - "order": 12, + "order": 11, "mode": 0, "inputs": [ { "name": "images", "type": "IMAGE", - "link": 17 + "link": 15 } ], "outputs": [ @@ -477,43 +470,36 @@ ] }, { - "id": 13, - "type": "ImageScale", + "id": 16, + "type": "EmptyLatentImage", "pos": [ - 1740, - 120 + 270, + 520 ], "size": [ - 320, - 190 + 280, + 130 ], "flags": {}, - "order": 11, + "order": 5, "mode": 0, - "inputs": [ - { - "name": "image", - "type": "IMAGE", - "link": 15 - } - ], + "inputs": [], "outputs": [ { - "name": "IMAGE", - "type": "IMAGE", + "name": "LATENT", + "type": "LATENT", "links": [ - 17 + 16 ] } ], "properties": { - "Node name for S&R": "ImageScale" + "Node name for S&R": "EmptyLatentImage" }, "widgets_values": [ - 1080, - 1920, 1088, - "center" + 1920, + 1 ] } ], @@ -583,9 +569,9 @@ "CONDITIONING" ], [ - 9, - 5, - 2, + 16, + 16, + 0, 10, 4, "LATENT" @@ -634,14 +620,6 @@ 15, 11, 0, - 13, - 0, - "IMAGE" - ], - [ - 17, - 13, - 0, 12, 0, "IMAGE" @@ -650,7 +628,7 @@ "groups": [], "config": {}, "extra": { - "note": "Qwen Image 2.1 T2I baseline; reusable prompt stored in qwen_prompt.txt; 1080x1920 portrait output" + "note": "Qwen Image 2.1 T2I workflow; reusable prompt stored in qwen_prompt.txt; direct 1088x1920 latent and output" }, "version": 0.4 } diff --git a/optimized/T2I/qwen/workflow.json b/optimized/T2I/qwen/workflow.json index 25c93d9..15754b3 100644 --- a/optimized/T2I/qwen/workflow.json +++ b/optimized/T2I/qwen/workflow.json @@ -109,8 +109,8 @@ 0 ], "latent_image": [ - "5", - 2 + "16", + 0 ] } }, @@ -131,23 +131,18 @@ "class_type": "SaveImage", "inputs": { "images": [ - "13", + "11", 0 ], "filename_prefix": "fluxExperiments/qwenImage21/scene_baseline" } }, - "13": { - "class_type": "ImageScale", + "16": { + "class_type": "EmptyLatentImage", "inputs": { - "image": [ - "11", - 0 - ], - "upscale_method": "lanczos", - "width": 1080, + "width": 1088, "height": 1920, - "crop": "center" + "batch_size": 1 } } } diff --git a/optimized/T2I/variables.json b/optimized/T2I/variables.json index f6ca153..337e809 100644 --- a/optimized/T2I/variables.json +++ b/optimized/T2I/variables.json @@ -24,11 +24,11 @@ "seed_node": "8.noise_seed", "seed": 20260920, "resolution_nodes": [ - "13.width", - "13.height" + "16.width", + "16.height" ], "resolution": [ - 1080, + 1088, 1920 ], "prompt_file": "qwen_prompt.txt"