Z-Image Turbo (ZIT) Step-Based Multi-Image Workflow
This ComfyUI workflow is a professional-grade generation pipeline specifically optimized for the Z-Image Turbo (ZIT) model. It is designed to maximize the efficiency of ZIT's architecture by allowing you to compare a vast range of sampling configurations in a single run.
Core Functionality
Massive Sampler & Scheduler Comparison: The workflow generates up to 16 images in a single execution.
Optimization Testing: Each image uses a unique combination of samplers (such as
euler,dpmpp_2m, orer_sde) and schedulers (such assimpleorbeta) to help you find the "sweet spot" for photorealism and detail.Step-Based Generation: It splits the generation into distinct phases, enabling granular control over how a subject evolves across different sampling steps.
Modular LoRA Integration: A dedicated "LORA - ZIT" group features a Power Lora Loader (rgthree), allowing you to stack and manage LoRAs easily while maintaining model integrity.
Wireless Prompt Management: The system uses a centralized prompt system that concatenates a main descriptive prompt with global style/quality modifiers (e.g., "photorealism, 8k resolution") to ensure consistent aesthetic quality.
Optimized for Character LoRAs
This workflow is particularly powerful for creators working with specific character LoRAs where maintaining likeness is critical:
Likeness Stress-Testing: By generating 16 variations simultaneously, you can immediately see which sampler/scheduler combinations preserve the character's unique facial features and which ones cause "likeness drift".
Finding the Weight "Sweet Spot": Using the Power Lora Loader, you can quickly test how different sampling steps affect the weight of your character LoRA, ensuring the character looks correct without over-processing the image.
Consistency Verification: It helps identify "stable" configurations that produce the same character face across different poses or environments, which is essential for consistent storytelling.
Version Edits:
V1.4 - Added a Seed control to allow you to set Fixed Or Random Seeds for all Samplers.
Description
Mion edit to control groups
FAQ
Comments (8)
wow i appreciate this, i never took the time to put something like this together i just did it all one by one and this is a much more helpful way, thank you!
if i may make a suggestion, add the rgthree seed node and attach it to all the samplers so you can quickly and easily randomize or maintain a set seed, it makes the comparison even better when they're all working off the same seed
@SaoirseTeagan That is a good idea. I will implement this!
@SaoirseTeagan https://civitai.com/models/2379370?modelVersionId=2677815 added! That was a good suggestion.
So what is your conclusion?
Which sampler and scheduler are best for ZIT and ZIB respectively? (Generation time must also be considered.) Some samplers are too slow
No sure I have one. They all seem to have strengths and weaknesses depending on the lora and prompt.
I will say I seem to get the best results with Multi_res and Euler
I found that the er_sde sampler was least likely to adher to the prompt, and was most unlike the other samplers.
Nice! Glad this WF helped to find that out. I want to test this on my side now.