Acid Re-Flux版本v1.0 (ID: 760949)

Update: I hadn't wired the Florence nodes up correctly in the Text prompt group. Fixed and reuploaded.

Designed this to be a mainly automated Flux Img2Img & Upscaling workflows to fix old gens I haven't had the time to look at. It's laid out to move left to right, all the groups are numbered and colour coded, so hopefully it'll be easy to follow what it's doing.

It uses Florence2 to create the prompt, there is a manual prompt to add your own details and LoRA trigger words to it.

It does a standard Flux pass, a first upscale to 2304 on the longest side and a second Flux pass, and then a second upscale to 3072 on the longest side and a third Flux, I've given it a fair bit of creative freedom with fairly high Denoise values of 0.7, 0.6 and 0.5. Bring these down if you want the output to closely resemble the input. I'm using the original Flux loader layout so you'll need to change this if you're using NF4 or GGUF versions of Flux.

After this I've set up Florence and SAM2 to scan for the face and apply a mask and then this is sent to a Detailer (SEGS) node, did a fair bit of testing with different Face Detailer setups and this gave me the best results. It's all set up to do the face and the only time you need to do anything is when you get an error that mentions Index=0. When that happens Florence can't find what it's been asked to get the co-ordinates for. Have yet to have it happen with this as it seems fairly good at detecting faces, but my previous workflow for Inpaint had this setup and that's where I've seen that error. So far changing the Florence model to Base has fixed it for me.

Last group is a SDXL last pass at 0.1 Denoise. This is because after the second upscale I've had a faint grid pattern on the picture. It's set up to save after both the Flux Face Detailer and this this final group.

描述:

训练词语: