Cinema Diffuso

Cinema Diffuso (CD) is a model aimed at making images that looks like they're straight from a movie, with with characters often in mid-motion, emoting and interacting with the world and each other.

My intent is for CD to be used with img2vid models as the model of choice for film makers making use of generative ai.

CD version 0.3 was trained on roughly 18-1900 images, and then merged with RealvisXL v5.

For sampler settings I recommend any of the standard dpmpp_sde/karras combinations. For CFG, I've mostly been doing 5+, with Pertubed Attention Guidance ranging from 0.5 to 1.5.

Showcase images are all txt2img, meant to show some of the tags that CD has been trained on, cherrypicked but without fixes like outpainting beyond some simple detailing nodes. To get the most out of this model I highly recommend making good use of inference techniques like regional conditioning, controlnet and IPAdapter.

As a beta version, CD is currently far from what I want it to be, yet surprisingly close to what I wanted when I first decided to start working on making my own model. The upcoming versions will feature more poses, movements and actions, locations etc.

I will most likely stick with SDXL for this project for some time, as SDXL has very good controlnet models, and is less resource hungry than some of the newer models.