Jib Mix Flux

I have trained Flux Dev on my SDXL dataset and merged loras, correcting anatomy censorship and excessive bokeh/blurred backgrounds.

Jib Mix Flux version 7 PixelHeaven - beta
The main change is it removes "Flux lines" that plagued V6 and the original Flux Dev to some extent.

It may be overdoing freckles a lot but I wanted to see what people think of it. hence the Beta name.

I really recommend using Movie Portrait Lora on quite a high weight for a less plastic look, but I couldn't merge it into the model as in testing it was a Lora that can cause "Flux Lines"

Jib Mix Flux Version 7.8 Clear Text Focus

This version focuses on having more readable text generation.
The Skin texture may not be generally as realistic as Jib Mix v6 or v7.

It has less nipple flashing through clothes.

The 2000s Analog Core @0.6 weight, removes some of the plastic Fluxness of this merge without hurting the text.

Jib Mix Flux Version 7.2 Pixel Heaven

7/7.2 mainly fixes "Flux Line" caused by merging some effected lora's, this has caused some drop in photo realism but an increase in drawn/concept ability and general details.

It also tones down obsessive amounts of freckles (especially red freckles).

The V7.2 Pruned Model nf4 (6.33 GB) model is actually the Q4_0.gguf

The V7.2 Pruned Model bf16 (11.84 GB) model is actually the Q8_0.gguf

Jib Mix Flux Version 6.1 Real Pix Fixed

6.1 mainly tries to fix small stubby hands or massively distorted hands/arms.
(If you still have problems with hands and you are using a low step count around 8 then increasing the step count usually fixes it alternatively applying a low weight (< 0.10) of the Hyper Flux lora pretty much always fixes hands, although you may find it lowers details).

V6.1 is more realistic on the CFG / more natural faces.

I think it actually does art/cartoon styles a bit better than the original v6 as well.

The V6.1 Pruned Model nf4 (6.33 GB) model is actually the Q4_0.gguf

The V6.1 Pruned Model bf16 (11.84 GB) model is actually the Q8_0.gguf

V6 still has the most detailed backgrounds in my testing.

The V6 Pruned Model nf4 (11.84 GB) model is actually Q8_0.gguf

The V6 Pruned Model bf16 (6.33 GB) model is actually a Q4_0.gguf

Jib Mix Flux Version 5 - It's Alive:

• Improved photorealism. (Less likely to default to painting styles)

• Fixed issues with wonky text.

• More detailed backgrounds

• Reconfigured NSWF slightly

fp8 V4 Canvas Galore:
better fine details and much better artistic styles, and improved NSFW capabilities.

fp8 V3.0 V3.1 - Clarity Key
I initially uploaded the wrong model file on the 21/10/2024, it was very similar but the new file since 22/10/2024 has slightly better contrast and was used for the sample images.

This version Improves detail levels and has a more cinematic feel like the original flux dev.

reduced the "Flux Chin"

Settings - I use a Flux Guidance of 2.9
Sampler = dpmpp_2m.

Scheduler = Beta or Custom Stigmas.

FP8 V2 - Electric Boogaloo: Better NSFW and skin/image quality.

Settings:
I find the best settings are a guidance and 2.5 and a CFG of 2.8 (although CFG does slow down the generation).
When using Loras these values may/will change.

Version: mx5 GGUF 7GB v1

This is a quantized version of my Flux model to run on lower-end graphics cards.

Thanks to @https://civitai.com/user/chrisgoringe243 for quantizing this, it is really good quality for such a small model.

There are larger-sized GGUF versions available here: https://huggingface.co/ChrisGoringe/MixedQuantFlux/tree/main
for mid-range graphics cards.

Version 2 - fp16:

For those with high Vram Cards who want maximum quality I have created this merge with the full fp16 Flux model. if you "only have 24GB of Vram" you will need to force the T5 text encoder onto the CPU/System RAM with this force node on this pack:
https://github.com/city96/ComfyUI_ExtraModels

Those people waiting for a smaller quantized model I am still looking into it.

Version 2:

Merged in 8 Step Hyper Lora and some others.

Settings:
I like a Guidance of 2 and 8-14 steps.

Resolution: I like a around 1280x1344

Version 1 : brings some of the benefits and look of SDXL with the massive prompt adherence benefits of Flux.
Settings:
I like a Guidance of 2 and 20-40 steps.