How to generate consistent AI images?

[FernandoAlonso_]

i've been learning how to use ComfyUI, and i've learned how to do most things i want, but consistent images are still a problem. i'm already using LoRAs, but they're not enough. if i generate two images using the exact same prompt and seed, i still get very different results.

all the tutorials i found either tell you to use flux, to create a LoRA, or to use an ultra complex workflow that my pc can't run.

 

[stabrepete]

You are getting the right information about Loras. What sampler are you using. some of them will give diffent results on different machines.

 

[Midzay]

Lora +
There are also PulID and other similar face transfer plugins.
+
A new powerful mechanism FLUX Kontext, has recently appeared. But it is SFW and paid on their server.
I can't help with any more information—I don't have any experience.

 

[NoTraceOfLuck]

https://f95zone.to/threads/your-noob-to-expert-guide-in-making-great-ai-art.256631/

 

[jonUrban]

Here's a Kontext checkpoint that's 11GB, which would fit into a 3060-12GB card. I'd advise you have at minimum 32GB CPU RAM as well. Comfy 3.43 (update your comfy) has built in support for Kontext as well as gguf, should you need to run quantized models. Run your comfy manager for any missing nodes or look at the node-list and install anything that you're interested in.

You must be registered to see the links

 

[TauTau_de]

do you maybe have a workflow to start with? Couldn't see one at civitai

 

[Midzay]

You must be registered to see the links

or Pixaroma in Youtube. Great lessons and workflows for them.

 

[art4vandelay]

Does some have the advance workflows from this guy?

You must be registered to see the links

 

[Midzay]

Does some have the advance workflows from this guy?

You must be registered to see the links

Download the FREE workflow & guide:

You must be registered to see the links

 

[TauTau_de]

couldn't get that one to work... missing nodes Int Literal, CFG Literal, String Literal, AutoCropFaces

 

[osanaiko]

Flux.1 Kontext is amazing.

The ability to give simple instructions for desired changes WHILE STILL PRESERVING NEARLY ALL THE LOOK AND FEEL OF THE ORIGINAL when doing img2img is a game changer.

I was able to do the following two image edits within a half dozen re-renders and less than 10 mins while writing this message.

Original


Prompt: "Change so the man is sitting in a comfy chair. Keep the same facial features, hair style and clothing."


Prompt: "Change the man's pose so he is climbing a ladder. Also change the camera angle to look at the man's back. Keep the same facial features, hair style and clothing."



Here's two other random expression and pose changes from my tinkering yesterday:


This is a god-send for me.

I've been trying to get this sort of flexibility in posing or expressions, while still retaining close visual adherence to the base character image, for more than 12 months as I've worked on my game project. And so far I've made very slow progress with continual fallback to image editors (with my extreme lack of artistic talent) to try to scrape together the look I want to achieve.

Now, I've learned to do all these variations in a matter of a few evenings.

Simply amazing!

The only downside's I've encountered so far are:
- it's pretty slow: at least 5 times slower than SDXL
- only available with ComfyUI workflow wrappers, not automatic1111 (I still have not converted to comfy)
- not many LORA or similar tools available (yet), but...

It's very easy to argue that these negatives don't really matter much:
- speed doesn't matter if it's the ONLY "run in your local environment" tool so far with the capability to follow simple instructions more iike a LLM than the nasty "20 short phrases from danbooru clip" prompt style that StableDiffusion et al require.
- comfyui-only is not a barrier, I just gotta learn and get-good.
- The lack of LORA is not a big deal, because I can use a multi-step workflow with other models before or after the kontext step as needed. Besides, the output is not always "crisp" in my experience, so using sharpening/upscaling tool after is the norm for now anyway

Links:
Here's a video I used to understand how to install Flux Kontext into comfyui and use-

You must be registered to see the links

here's some more in depth written info on the various ways to use:

You must be registered to see the links

 

[Doenerteer]

i've been learning how to use ComfyUI, and i've learned how to do most things i want, but consistent images are still a problem. i'm already using LoRAs, but they're not enough. if i generate two images using the exact same prompt and seed, i still get very different results.

all the tutorials i found either tell you to use flux, to create a LoRA, or to use an ultra complex workflow that my pc can't run.

What helped me with Noob-AI XL models to create consistent characters is this site:

You must be registered to see the links

All the characters are implemented in the model, for example Pony-XL, Ilustmix, yiffymix etc. (the last 2 are my favourites)

You just need to adjust the prompt for the character that you like to use can change their cloth, gender, color, hair etc.

Dont forget to use a fixed seed or change it a little and play a little with the resolution

Yes it will be a little try and error some pictures will be shity xD but you will see it works okay

I know that youtube guides for consistent characters but these are complicated and need time.

I changed Mika from blue archive into a Kitsune

 

[Kurotori]

If you're still interested, we opened up an "AI technical" channel in my discord where we're looking for contributors and learners alike to share all knowledge.

You must be registered to see the links

Also the artist for my game as well as a few others have their own consistent OCs with lots of tips and tricks available already to sift through.

 

[Seanthiar]

Does anyone know where to download the FLUX Kontext NSFW-Version ? Found only a Chinese pay site for about 70$ and that would be a bit risky even if I had a credit card.
Normal Kontext has problems with 'risky' pictures. For example I can manage to give a character a bra, but if I want a lacy sheer bra it is ignored and it gives the same result i get with just the word 'bra' and when I want a strapless bra, there is no change at all from the source picture at all.