This is my setup for storyboarding with FLF in mind. The backbone is a three stage diffusion pipeline, similar to other I2V workflows I've posted. For this use case I set up a series of prompts with the aim of locking down a setting and chararacter(s) and running each prompt with the same seed. Backgrounds can be an annoyance with FLF, freezing a seed can help. The logic works in batches, you only need to set the keyframe total, then each run will cycle through all of the prompts. I set my batch count to match it and just click run once, but of course clicking for each one will do the same thing. There are replace nodes, so you can set a face/body prompt and a composition prompt once and only enter the changes into the keyframe strings.

I did not automate incrementing the seed, as I usually do more than one pass. I have a few versions, only two are presentable. Only the Pony version has completely independent model, clip and vae, so use that one if you combine incompatible models at each stage.

First stage is a full generation (no split-sampling), second is a low denoise refine, third is broken up into detailers. Face, nips, minge and dong. Changeable to whatever you wish with the appropriate detector. Copious notes on all the models and nodes and whatnot. I shall not repeat myself. Just drag it on over and make that sea of red go clear. Like a good full operating thetan.

Each of the four detailers can be bypassed if desired. A separate bypass node will toggle the previews for the masks and enhancements. Turn them on if you need to debug your mask settings and/or see the detailed part in isolation. Turning them off helps keep your asset stream free of clutter, remember, comfy always saves all of that somewhere. There is also an awesome compare node at the end for split-screen, overlay or difference. I really like the latter, it is extremely useful, sometimes it's really hard to see what you are looking for, so a pixel difference heat map is great.

The Pony version uses an awesome ksampler from the Inspire pack that lets you choose a variation seed, basically emulates A1111 style generation. It's really nice, you keep your seed fixed and just adjust the strength of the variation. Very, very helpful here. I also have video-style preview nodes in here, because for whatever reason my image sampling previews are not working, can't figure out why. But these are better anyway, you get the little sigma display, watch convergence in action. Crawl inside the dark machine. Screws so tight.

GonzaLomoChroma version:

Basic Workflow Structure I use with these models - GonzaLomo Chroma Refiner Workflow

The main model - GonzaLomoChroma

The refiner model - GonzaLomo ZPop

The detailer model - GonzaLomo XL/FLUX/PONY

Pony version is the same but uses CyberRealisticPony as the main model. If you want to substitute a different Pony, XL, Illustrious, etc. base this is the one to use, as the model/clip/vae pipes are independent for each stage. This is not the case for the other version.

If you are not familiar with detailers, get familiar with detailers.

You need a SAM in addition to detectors for any body part you would like to detect. Bbox, segm, whatever you want. Find them on a site that hosts AI models.

To install custom nodes, press the INSTALL CUSTOM NODE button. Yeah, it's super complicated, I know. Or just pull a git repo into the folder and it will install itself. By magic. Don't forget to lower security to weak or none and enter all requested personal information.

Please take a moment to visit GBRX and apply thumbs. It's all good stuff.

You can of course use this as plain old T2I if you wish. Just set the keyframe count to 1 and unlock the seed, so only prompt 1 will ever be active.