I am glad!! And I assume you read the notes to use it properly, there are some notes on better optimization for low vram users!

nedolo77 is correct. Also, I'm just going to remove my direct LUT link, since I mostly just care about my community of users, my main concern is to only provide them with the best possible tools.

Yes that's intended. The usage of steps here is different than how it's typically used. It's not using raw step count to generate better details, it's through the interaction of the few variables combined.

no, it's all already connected as it should be. I separated the "detail vs cine" step value into its own box which does what it says. The actual flowmatch step value doesn't matter as much, that value also has its own box, and notes, on the min and max value for it.

Those are good suggestions, and it comes down to a matter of preference. Ultraflux is too sharp for my taste, and some of it can be a bit artificial too, I prefer using my workflow settings to adjust that, and the simple sharpen/contrast node. The seedVarianceEnhancer, I will think about it, but generally I haven't used that node much because I prefer 2 methods: 1) Quick prompting, prompt it and hope for the best, or 2) carefully craft every section of the prompt with full control

even runs on 8gb Vram, just depends on settings (fully commented in the workflow) and also the models you choose. The base turbo model Q8 version will be more than fine.

you'll have to show me your examples before I can comment on this, hard to tell without seeing what you are referring to

character trained lora should work, just add it to both lora group 1 and group 2

@k4d it depends on the lora, you can keep the same value on both, and test 0.3, 0.5, 0.75, 1.0

@k4d great! And regarding inpainter, I'm not sure what you mean by bad, but remember the inpainter isn't Qwen EDIT, so you just type in what you want to see in the masked area you created, and also the denoise amount matters a lot. Also use very very soft edges on your masked area instead of hard edges

actually I gave up on Qwen 2. The quality isn't there, the new qwen model is just like a slightly better finetune. There's plenty of chroma and wan workflows out there! I just chose Z image because of the higher cap of image quality you can produce with the ideal workflow. Wan and Chroma is unlikely to be better than Z image.

@markhassain3712 I used Wan for a while, and I felt it was pretty limited, it was fine for upscaling though. But now that I have Z image, I no longer needed a separate different model to do the upscaling pass.

In the workflow, I've noted where to place it actually (under the LUT node), and that's definitely not where it needs to go

@TienX not sure, depends if that custom LUT loader node decide to suddenly move it to a new location for loading LUTs.

The LUT is included in the zip file, so just grab it from that! Thank you for your kind words, I'm glad that I was able to make it super user friendly, where it just works! And of course, can also do advanced tweaking with my attached notes too if you prefer. Have fun =)

@tornblood i MIGHT train a new non-cyber even higher quality one based on ZIB, but for ZIT usage, thinking about it.

@tornblood I dont think I can even offer Early Access yet on my account

@tornblood haha that might incentivize me to do it sooner since I have to go and pay for Runpod 5090 rental for 2 days to train it.

I might train a brand new model based on ZIB to be used for ZIT/ZIB. I've just been really busy so haven't gotten around to it yet. It'll basically be even better quality version of what my Loras have been offering, but without the focus on Cyber detailing, can use it more broadly for modern/fantasy/cyber.

@aneebtheartist126 don't worry, whatever I make, I will first make sure it's uniquely useful to the community and a significant enough upgrade.

I've been still busy! Actually making a bunch of new AI tools. Different than this but still very useful. When I get better hardware to play with though, I'll make more image related upgrades in the future. Images are generally pretty good already, but maybe not as advanced as things like Nano Banana Pro for realism. I wonder if I should just make a desktop tool that has tools/image edits that comfyUI can't do natively with its interface, like drag and drop elements more easily within images, and other convenient character editing tools, to iterate edits, but that sounds like too much work =) Right now I can keep adding to the complexity of the workflows, but I would actually much rather simplify everything. So for example, teaching an AI model to automate changes, but highly optimized in its approach, so that it doesn't just regenerate the entire image, it should only recalculate the areas it needs to. Yes can do masking now, but there are many manual steps involved, a fully instruction/voice based editing interface might be nice. So that you can just sit back on your sofa and iterate on images until it's perfect. Maybe when I have time =)

because some finetuned models baked in increased contrasts already... I consider those "over-tuned"

@azusazell007 hard to say, every single model would require different adjustments, since the hard baked models forcefully apply influences.