Introduction
Scene Composer is a worklfow for ComfyUI that let you generate random stories in just one run. The system roll parts of the prompt that are used for generation, like:
Render: light, artstyle, etc.
Character: body, hair, eyes, attitude, etc.
Clothes: Dress, swimsuit, underwear, etc.
Action: SFW/NSFW, position, gestures, etc.
Environment: indods, outdoor, background, etc.
To maintain consistency, it also keeps certain parts of the prompt to inject across all scenes (like the environment, the main character and their clothes). You can read my Overview & Usecases article for more explanations!
If you're looking for a simpler workflow, check my Main ComfyUI Workflow. I also suggest to have a look at my Prompt Notebook to better understand how I structure tags.
If you have any comment or request, please feel free to share!
Features
Note: features in gray are in progress and need to migrate to v2.x
Random procedural generation of prompts
Main character (e.g. body, hair, eyes, fantasy traits, …)
Attire (e.g. clothes, swimsuit, underwear, uniform, accessories.…)
Environment (e.g. place, weather,…)
Action in scene (e.g. starting scene, sexual encounter, ending scene,…)
Predefined personas (demon, goblin, furry, slime, etc)
Control scenes parameters
Seed, steps, CFG, image size, etc.
HighRes-Fix (2nd Pass)
LoRA stacker
Keep control over scenes
Re-generate one or many elements by changing its seed
Re-generate one or many scene by changing its seed
Overwrite and compose the final prompt with variables
Scenes consistency
Tags update dynamically according to the scene (e.g.
wetis added if there's rain)Attire state stay across scenes (if character lose clothes, it stay lost)
If clothes are torned, they stay so
Bondage ropes stay on character, with clothes
Output
Upscaling, pre-processing
Images scene merged into one
CivitAI metadatas & workflow embedded
Setup
Simply import the workflow.json file attached to this article in ComfyUI.
I personally don't have a very powerful computer. For people in the same situation, check my environment on my Main ComfyUI Workflow article, I explain how I rent and setup machines on remote.
Models
In theory, the workflow can work with any models that use Danbooru-like tags. I personally use Illustrious/NoobAI/Pony-based models and mix with some anime-oriented LoRAs. Have a look at my last images metadatas if you're curious!
Description
This version include the following changes:
Denoise parameter is now globally changeable
LoRAs metadatas are displayed in output images
Seed for the prompt and the image generation are now clearly seperated, and can be changed globally or pro scene.
The scenes merged according to the ratio: vertical if it's landscape images, horizontal if it's portrait images.
Small breasts paizuri is improved (but still unstable)
Sometimes, clothes will be around character's waist (should be include in this-clothes)
If the character is gagged, no actions requiring an open mouth (e.g. blowjob) should appear
FAQ
Comments (6)
Throwing error for High Resolution Fix Script
When I had issue with the Hires-Fix, I tried to disable "use same seed" option and set it manually (to a positive value).
got an error with Meta data even though its installed...
@DJLegends The fresh new version of the workflow should fix this. If you want to keep the old one, try to "Fix node" by right-clicking on the Highres fix node
@taches In your current version the "use same seed" is disabled by default right? I'm also getting an error both with the seed option enabled and disabled. Do you have any idea how to solve this?
The error is the one below:
Error occurred when executing UpscaleModelLoader: 'NoneType' object has no attribute 'lower' File "C:\ComfyUI_windows_portable\ComfyUI\execution.py", line 317, in execute output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\ComfyUI_windows_portable\ComfyUI\execution.py", line 192, in get_output_data return_values = mapnode_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\ComfyUI_windows_portable\ComfyUI\execution.py", line 169, in mapnode_over_list process_inputs(input_dict, i) File "C:\ComfyUI_windows_portable\ComfyUI\execution.py", line 158, in process_inputs results.append(getattr(obj, func)(**inputs)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\ComfyUI_windows_portable\ComfyUI\comfy_extras\nodes_upscale_model.py", line 29, in load_model sd = comfy.utils.load_torch_file(model_path, safe_load=True) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\ComfyUI_windows_portable\ComfyUI\comfy\utils.py", line 33, in load_torch_file if ckpt.lower().endswith(".safetensors") or ckpt.lower().endswith(".sft"): ^^^^^^^^^^
@MigWin i tink it's because of the Lora stacker. Put them all on "None" and limit it at 0. I should do this in the default version!