Pose2Flux v2.1

2024.08.22: A small glitch happened in the release of v2 that I didn't notice until now, but the noise injection node was missing a noise input 👀. That's been corrected now. I also:

• added an Ollama node in the prompt group,
• added a new switch in the Refiner group to toggle img2img and txt2img (so making this a straight Flux gen workflow) and
• removed superfluous nodes,
• corrected wrong color codes,
• started annotating a bit everywhere,
• general cleanups for more legibility.

2024.08.20: When I launched the v1 of this workflow, the last thing I expected was to see over 125 of you download it and use it. I was joyfully surprised and it made me all the more motivated to pursue this little project. Thanks for this, folks :)

What changed in this v2:

• Switched from SD1.5 to SDXL for the base gen.
• Switched from Flux.1 dev fp8 to dev NF4 for the refiners.
• Added Florence2 to help with prompt crafting.
• Scraped all 5 detailers I was working on and instead went the noise injection* road. That results in a much leaner and faster workflow, with a way more detailed output at 2K.
• Added Tiled Diffusion upscaling. In my tests (and settings in the node), 4K works very well. Haven't tried bigger yet.
• Revamped the color scheme (see Color Codes node)
• Reworked the Controlnets to offset SDXL's openpose finnickiness. Depth is unofficially driving the entire thing, with segmentation nodes allowing the most finetuning I could find in any of the nodes out there. Right now, the arms are masked along the face, who gets a small opacity treatment so that it only affects proportions and head positioning, nothing else (like hair or expression). It works surprisingly well, albeit with manual tweaks when needed.
• Various UI refinements.

* I first learned about it in Matteo's latest video, Deep dive into the Flux, timestamped for your convenience.