FRANKLY (WORKFLOW) - Longish and consistent H3 Minimax
retained text
The Workflow : [https://drive.google.com/file/d/185D3\_3QBxBFViXENL0lIDen-tp3Vab6M](https://drive.google.com/file/d/185D3_3QBxBFViXENL0lIDen-tp3Vab6M) This is an update to my earlier worflow. I left it at three parts that generate three clips based on the same two references. Part two and three also get 10 seconds of a reduced resolution look at what came before. It doesn't have to be full res to get the vibe. The way I use this is keep generating until I like section one, let it run into section two and either tweak the seed or prompt to make it better. Then let it go on to section three. It's similar to a wan SVI workflow I liked. You can run it at low resolution for fast output and only set it higher when you have an output that's good. There is no real benefit to having all three parts in the same workflow. I could just as easily run the workflow once and feed in the result as a 10 second clip and change the prompt etc. When testing the earlier fight and car chase sequences I noticed that a lot of the other 4step and 3step loras, and also the singularity models were messing up the action speed and camera angles. So the model and turbo lora I used here give the best result for speed, action and adherence. The cinematic look lora I got from [Civit](http://Civit.ai) It's only a 7 step workflow so add steps to improve output. It's set to 14 second clips (x3) so reduce or increase as necessary.
document fetched 2026-09-26 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1wpra5i/frankly_workflow_longish_and_consistent_h3_minimax