Video projection with depth estimation on a point cloud matrix

More experiments with noise, this time without tiling. This new workflow is less effective than the previous but a lot faster and generally grants slightly better results. Left standard flux, right enhanced.

screenshot from my new video game... I mean Lora

Don't be afraid of experimenting with samplers and various parameters. dpm_fast guidance 0? sure, no problem

Take that Flux! SDXL FTW! (inspired by a recent post by Muriel Lherm )

It's monday fellas! Have a creative week!

that discomfort when your new code works at first try... there must be a bug somewhere that will get me in a few days

うぉー！動いた！！やばいやばい #赤外線ハンドガン

Guidance versus clipart and text. From guidance 0 to 7

Testing flux dev on RTX 6000 Ada, can recommend. Sampling speed is close to a 4090 with the added VRAM bonus. Now I kinda want one locally...

Guidance 0-7 vs Sci-fi.

Repeating the same steps/samplers/schedulers test for many subjects and at various prompt lengths. Illustration style changes a lot with the number of steps. All results will be published soon!

LMS/sampler beta/scheduler is that weird uncle at Christmas dinner

... and course guidance can be negative... #tableflip I really need to check the code to see what happens

More testing with flux guidance. -3.5 to 3.5. Note that guidance 0.5 is the closest to the prompt that asked for an abstract, watery, artistry, watercolor

Hard to say what is going on, but negative values look generally more interesting. Notable exception is text that works better at positive guidance

Anyone interested in block by block patching for flux? I've found the "style boost" blocks. Also before/after patching

playing with self attention. don't ask, I don't know what the blocks do (yet), but it's certainly fun. The node is available in the comfyui_essentials. have a nice week-end!