Two versatile workflows that includes useful options. Wilcards or Normal Prompting, Upscaling and refining, and Seed Variance to modify slightly a previously made image.

PornosotrosWorkflows

To use the Wildcard node with a file that contains many prompts. Create a .txt file with all the prompts seperated only by a return to a new line (no empty line). like this:Portrait of a woman, golden hour, warm highlights, beach.A cute dog, on a sidewalk, red fire hydrant, palmtrees.Breathtaking seascape, lighthouse, seagulls, sunset, golden hour.Save it in ComfyUI/custom_nodes/ComfyUI-Impact-Pack/wildcards/Call it using my-prompts-file between two undersores on each side (without .txt) in the upper prompt section only.You can combine multiple wildcards in one prompt, e.g. my-prompts-file, clothing, hairstyles. (if clothing.txt and hairstyle.txt are present in the folder)Make sure to leave Populate mode selected to expand wildcards.For normal mode, just place your prompt in the first section, and ignore the second section.---------------------------------------------This wildcard method is a very practical way to create lots of images from different prompts in one go. You must leave batch at one. It's in the RUN Box that you set how many images to have.###############The SeedVarianceEnhancer above can be useful if you want to make small changes to an image you dragged back in Comfy.To try different settings, you should first set "Control After Generate" to Fixed, and make sure you have the prompt in the fist section, not a wildcard entry.################If you set a 4steps SDXL DMD2 LoRA to -1, you can test using this DMD2 checkpoint with samplers intended for NON DMD2 models (so without LCM). The Idea is to get a rough base image as first pass that looks different than what LCM usually gives. Then the next passes upscale and refine in DMD2 mode with LCM at low settings, hopefully making the base image look much better in the end.Some samplers that can work pretty well for that are:sa_solver: CFG 4+, Steps 35+ --- Scheduler: simple, sgm_uniformsa_solver_pece: CFG 4+, Steps 35+ --- Scheduler: simple, sgm_uniformres_multistep: CFG 4+, Steps 30+ --- Scheduler: simple, sgm_uniformres_multistep_cfg_pp: CFG 1, Steps 30+ --- Scheduler: sgm_uniformseed_3: CFG 1, Steps 35+ --- Scheduler: ddim_uniform*** I haven't tried all samplers though. Above are suggested values I tried with some success.That is an indirect way to use a DMD2 model, so expect maybe more errors or lower prompt adherance, but you will get different results than with LCM. I like that this seem to offer more face variations without a character LoRA.

Files with SHA256: 40615d15e1d35b9747a506344ac7679718f6125c193a401a5bd789b5490e0ec7