Add static variables and optimization test matrices

This commit is contained in:
Morpheus
2026-09-23 09:04:14 +00:00
parent 7ff0b7547a
commit 910d0f2e12
15 changed files with 153 additions and 2 deletions
+2
View File
@@ -1,3 +1,5 @@
# I2I workflows
Use [`flux/`](flux/) or [`qwen/`](qwen/) for one-reference-image transformations. The workflows infer the subject from the input image instead of naming a specific character or scene.
The shared reference asset and static variables are recorded in [`variables.json`](variables.json) and `character_reference.png`. The test plans live in each model README.
Binary file not shown.

After

Width:  |  Height:  |  Size: 647 KiB

+15
View File
@@ -18,3 +18,18 @@ Files:
- `workflow.json` — API-format prompt.
- `ui-workflow.json` — native ComfyUI web-editor workflow.
## Planned optimization matrix
All rows are intentionally **Untested**. Keep the reference image, prompt, seed, and input-derived resolution fixed. Flux2Scheduler does not expose a selectable scheduler, so that value remains fixed.
| ID | CFG | Steps | Sampler | Scheduler | Tested? |
|---|---:|---:|---|---|---|
| F-I2I-01 baseline | 1.0 | 4 | euler | Flux2Scheduler | No |
| F-I2I-02 | 1.0 | 4 | euler_cfg_pp | Flux2Scheduler | No |
| F-I2I-03 | 1.0 | 6 | euler | Flux2Scheduler | No |
| F-I2I-04 | 1.0 | 6 | euler_cfg_pp | Flux2Scheduler | No |
| F-I2I-05 | 1.0 | 4 | res_multistep | Flux2Scheduler | No |
| F-I2I-06 | 0.8 | 4 | euler | Flux2Scheduler | No |
| F-I2I-07 | 1.2 | 4 | euler | Flux2Scheduler | No |
| F-I2I-08 | 1.0 | 8 | euler_cfg_pp | Flux2Scheduler | No |
+18 -1
View File
@@ -4,7 +4,7 @@ Converts any supplied reference image into a photorealistic image of the same su
## Reusable inputs
- `input.png` in node `4` — replace with the reference image.
- `character_reference.png` in node `4` — replace with the reference image.
- Node `7` positive and negative prompts — adapt the transformation without describing a specific subject.
- Node `10` seed — change for alternate results or keep fixed for comparisons.
- Nodes `5` and `14` — change the generation and output dimensions. The default is portrait 1080×1920 for person-focused images.
@@ -21,3 +21,20 @@ Converts any supplied reference image into a photorealistic image of the same su
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`](../variables.json) fixed.
| ID | CFG | Steps | Sampler | Scheduler | Tested? |
|---|---:|---:|---|---|---|
| Q-I2I-01 baseline | 3.5 | 40 | euler_cfg_pp | beta | No |
| Q-I2I-02 | 2.5 | 24 | euler_cfg_pp | beta | No |
| Q-I2I-03 | 2.5 | 30 | euler_cfg_pp | beta | No |
| Q-I2I-04 | 2.5 | 40 | euler_cfg_pp | beta | No |
| Q-I2I-05 | 3.5 | 40 | euler | beta | No |
| Q-I2I-06 | 3.5 | 40 | res_multistep | beta | No |
| Q-I2I-07 | 3.0 | 30 | euler_cfg_pp | simple | No |
| Q-I2I-08 | 4.0 | 30 | euler_cfg_pp | beta | No |
| Q-I2I-09 | 3.5 | 50 | euler_cfg_pp | beta | No |
| Q-I2I-10 | 3.5 | 40 | euler_cfg_pp | sgm_uniform | No |
+1 -1
View File
@@ -2,7 +2,7 @@
"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":"input.png"}},
"4": {"class_type":"LoadImage","inputs":{"image":"character_reference.png"}},
"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]}},
+26
View File
@@ -0,0 +1,26 @@
{
"description": "Static variables for controlled single-reference I2I experiments.",
"base_images": {
"character_reference.png": "character_reference.png"
},
"workflows": {
"flux": {
"prompt_node": "74.text",
"seed_node": "73.noise_seed",
"seed": 20260921,
"input_node": "1.image",
"input_image": "character_reference.png",
"resolution": "inferred from input; default input is portrait"
},
"qwen": {
"prompt_node": "7.prompt",
"negative_prompt_node": "7.negative_prompt",
"seed_node": "10.noise_seed",
"seed": 20260921,
"input_node": "4.image",
"input_image": "character_reference.png",
"generation_resolution": [576, 1024],
"output_resolution": [1080, 1920]
}
}
}
+2
View File
@@ -1,3 +1,5 @@
# MultipleRef2I workflows
Use [`flux/`](flux/) for the current best multi-reference image composition workflow. It accepts a primary subject reference and an environment or scene reference. A Qwen MultipleRef2I workflow has not yet been created.
The fixed reference assets and static variables are recorded in [`variables.json`](variables.json), `person_reference.png`, and `scene_reference.png`. The test plan lives in the Flux README.
+15
View File
@@ -8,3 +8,18 @@ API workflow for combining two reference images with Flux.2 Klein:
The workflow chains both encoded images through `ReferenceLatent` for the positive and negative conditioning branches, then renders a 1920×1080 result. Update the prompt when the references have different roles (for example, product + room, character + landscape, or object + background).
This is an API-format workflow. The tested example is stored under `fluxExperiments/multiImage/`.
## Planned optimization matrix
All rows are intentionally **Untested**. Keep both reference images, prompt, seed, and 1920×1080 resolution from [`../variables.json`](../variables.json) fixed. Flux2Scheduler does not expose a selectable scheduler, so that value remains fixed.
| ID | CFG | Steps | Sampler | Scheduler | Tested? |
|---|---:|---:|---|---|---|
| F-MR2I-01 baseline | 1.0 | 6 | euler | Flux2Scheduler | No |
| F-MR2I-02 | 1.0 | 6 | euler_cfg_pp | Flux2Scheduler | No |
| F-MR2I-03 | 1.0 | 4 | euler | Flux2Scheduler | No |
| F-MR2I-04 | 1.0 | 8 | euler | Flux2Scheduler | No |
| F-MR2I-05 | 1.0 | 6 | res_multistep | Flux2Scheduler | No |
| F-MR2I-06 | 0.8 | 6 | euler | Flux2Scheduler | No |
| F-MR2I-07 | 1.2 | 6 | euler | Flux2Scheduler | No |
| F-MR2I-08 | 1.0 | 8 | euler_cfg_pp | Flux2Scheduler | No |
Binary file not shown.

After

Width:  |  Height:  |  Size: 2.8 MiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 3.7 MiB

+18
View File
@@ -0,0 +1,18 @@
{
"description": "Static variables for controlled multi-reference image-to-image experiments.",
"base_images": {
"person_reference.png": "person_reference.png",
"scene_reference.png": "scene_reference.png"
},
"workflow": {
"prompt_node": "74.text",
"seed_node": "73.noise_seed",
"seed": 20260953,
"person_input_node": "1.image",
"scene_input_node": "2.image",
"person_input": "person_reference.png",
"scene_input": "scene_reference.png",
"resolution_nodes": ["62.width", "62.height", "66.width", "66.height"],
"resolution": [1920, 1080]
}
}
+2
View File
@@ -1,3 +1,5 @@
# T2I workflows
Use [`flux/`](flux/) or [`qwen/`](qwen/) for the best tested text-to-image baseline. Prompts, seeds, and dimensions are intended to be replaced per task; sampler, scheduler, CFG, and steps are the optimization controls.
Static variables are recorded in [`variables.json`](variables.json). The test plans live in each model README.
+15
View File
@@ -21,6 +21,21 @@ Files:
The included seed and prompt are a controlled baseline example. For reuse, replace the prompt, seed, and resolution; sampler, scheduler, steps, and CFG are the optimization controls.
## Planned optimization matrix
All rows are intentionally **Untested**. Keep the prompt, seed, and resolution from [`../variables.json`](../variables.json) fixed. Flux2Scheduler does not expose a selectable scheduler, so that value remains fixed.
| ID | CFG | Steps | Sampler | Scheduler | Tested? |
|---|---:|---:|---|---|---|
| F-T2I-01 baseline | 1.3 | 6 | euler_cfg_pp | Flux2Scheduler | No |
| F-T2I-02 | 1.3 | 4 | euler_cfg_pp | Flux2Scheduler | No |
| F-T2I-03 | 1.3 | 8 | euler_cfg_pp | Flux2Scheduler | No |
| F-T2I-04 | 1.3 | 6 | euler | Flux2Scheduler | No |
| F-T2I-05 | 1.3 | 6 | res_multistep | Flux2Scheduler | No |
| F-T2I-06 | 1.2 | 6 | euler_cfg_pp | Flux2Scheduler | No |
| F-T2I-07 | 1.4 | 6 | euler_cfg_pp | Flux2Scheduler | No |
| F-T2I-08 | 1.3 | 8 | res_multistep | Flux2Scheduler | No |
## Optional second pass
SeedVR2 3B FP8 produced a successful 2560×1440 output from this baseline. A 3840×2160 attempt ran out of the GPU's 8 GB VRAM. The SeedVR2 pass is therefore an optional export step, not part of the fast generation baseline.
+17
View File
@@ -18,6 +18,23 @@ Qwen Image 2.1 text-to-image workflow using the best tested sampler, scheduler,
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.
## Planned optimization matrix
All rows are intentionally **Untested**. Keep the prompt, seed, and resolution from [`../variables.json`](../variables.json) fixed.
| ID | CFG | Steps | Sampler | Scheduler | Tested? |
|---|---:|---:|---|---|---|
| Q-T2I-01 baseline | 2.5 | 12 | euler_cfg_pp | simple | No |
| Q-T2I-02 | 2.5 | 20 | euler_cfg_pp | beta | No |
| Q-T2I-03 | 2.5 | 30 | euler_cfg_pp | beta | No |
| Q-T2I-04 | 2.5 | 40 | euler_cfg_pp | beta | No |
| Q-T2I-05 | 2.5 | 12 | euler | simple | No |
| Q-T2I-06 | 2.5 | 12 | res_multistep | beta | No |
| Q-T2I-07 | 2.0 | 20 | euler_cfg_pp | beta | No |
| Q-T2I-08 | 3.0 | 20 | euler_cfg_pp | beta | No |
| Q-T2I-09 | 2.5 | 20 | euler_cfg_pp | sgm_uniform | No |
| Q-T2I-10 | 2.5 | 20 | euler_cfg_pp | karras | No |
Files:
- workflow.json — API format
- ui-workflow.json — ComfyUI web UI format
+22
View File
@@ -0,0 +1,22 @@
{
"description": "Static variables for controlled T2I experiments. Prompt text remains in each model workflow and is referenced by node below.",
"base_images": [],
"workflows": {
"flux": {
"prompt_node": "4.text",
"negative_prompt_node": "5.text",
"seed_node": "9.noise_seed",
"seed": 20260920,
"resolution_nodes": ["8.width", "8.height", "10.width", "10.height"],
"resolution": [1920, 1080]
},
"qwen": {
"prompt_node": "5.prompt",
"negative_prompt_node": "5.negative_prompt",
"seed_node": "8.noise_seed",
"seed": 20260920,
"resolution_nodes": ["13.width", "13.height"],
"resolution": [1920, 1088]
}
}
}