v1.7S
Moxie Flux 1 DS v1.5S (previous version - high percentage)
v1.7D
Moxie Flux 1 DS v1.6 (previous version - high percentage)
C4PACITOR d v3 alpha (slightly higher percentage)
v1.6D
Moxie Flux1 DS v1.5D (high percentage)
v1.5S
Shuttle 3.1 Aesthetic v1.0 (high percentage)
v1.5D
1st merge
Colossus Project Flux v.42_AIO (high percentage)
2nd merge
C4PACITOR v.d_v2 (high percentage)
1.4
This is the 'new base' for Moxie Flux Fusion, f16 e5m2 version. While this base, I find it has more 'failed' attempts (probably higher success at more steps, and playing with guidance. I keep it at 3-3.5 guidance, and 22 steps). However, the images are unique from other flux models, and it produces some great images. It had made a great pairing for Moxie Fusion Flux v1.35, giving it some good variety away from the 'stock' model.
1st merge
2nd merge
Dev2Pro (just a splash - I make Lora's with it, and sometimes use it as an upscaler)
1.3
This is the 'new base' for Moxie Flux Fusion, but the f8 version.
1st merge
2nd merge
3rd merge
Merge first two merges
4th merge
Flux dev1 fp8 e4m3fn (as the primary)
Flux1.D & Flux1.S 8-Step
This is the absolute 'base' checkpoint of Flux Dev/Schnell (UNET, fp8, e4m3fn), merged with ByteDance Hyper-FLUX Acceleration LoRA. It is merges to be a full checkpoint, with dual clips, and VAE included. I'm sure these are somewhere else, but I was unable to find them, so I am uploading them.
I find the quality about neck in neck with the 'full' version, and I use it quite often (for testing) since, obviously, 8-step is much faster than 20. After trying to find another upload of this, but finally gave up and decided to upload it.
v1.1
This is 50/50 merger of AbdallahAlswa80's 2 speed models, Speed FP8 (e4m3fn) D and Speed FP8 (e5m2) S. Produces images in 4 steps. See sample image metadata for specifics.
v1.0
There are other Flux.1 D and S merges, but none of them are an actually 'stand alone' checkpoint's. Hence, my motivation for this. Also, this has been tested on a 4060 16GB and 3060 12GB with no issues.
Adel_AI already created a couple of checkpoints that greatly simplify using Flux where you only have to download 1 file, and place it in your checkpoint folder for ComfyUI (reference my workflow here.)
This is simply a 50/50 merger of the two Adel_AI Flux1 checkpoints, Flux1.D 16Gb and Flux1.S 16Gb. Below is a copy and paste from his page that gives further info.
Description
There are other Flux.1 D and S merges, but none of them are an actually 'stand alone' checkpoint's. Hence, my motivation for this. Also, this has been tested on a 4060 16GB and 3060 12GB with no issues.
Adel_AI already created a couple of checkpoints that greatly simplify using Flux where you only have to download 1 file, and place it in your checkpoint folder for ComfyUI (reference my workflow here.)
This is simply a 50/50 merger of the two Adel_AI Flux1 checkpoints, Flux1.D 16Gb and Flux1.S 16Gb. Below is a copy and paste from his page that gives further info.
Flux.1 S [Shnell] is a 12 billion parameter T2I model developed by Black Forest Labs (https://blackforestlabs.ai), allowing to generate high-quality images in only 1 to 4 steps.
Safetensors with baked T5 fp8, Clip L and VAE
Put the model in the checkpoint directory of ComfyUI.FAQ
Comments (17)
Do you know if this has been tested on Forge/A1111?
To my knowledge, those don't support Flux yet, but I'm not entirely sure.
@moxie1776 A1111 cannot support Flux because of the way it's built. I'm not completely sure about forge.
@mech4nimal Forge is fairly similar in framework, or atleast uses the same base webui hooks sadly.
@A_Friendly_Spider I'm quite familiar with forge. I used it for a while, until I realized they had stopped development on it.
try SwarmUI, runs with Comfy in the background, but has an additional simplified interface such as AUTOMATIC111. Works perfectly with Flux, incl. upscale, LoRAs, VAE and ControlNet (can be found somewhere on Huggingface, but only 1 or 2 filters)...
There was an attempt with ReForge but seems abandoned. Forge has recently updated some Flux related commit, and reports 90% done remaking a new backend. So expect next week to see some unstable Forge with Flux support. Might be nice.
@o639 I may just have to bite the bullet and swap to comfy, since I've been wanting to get into the video side of things too. This just increases the reasons to attempt the switch. I will miss the familiarity of WebUI though T_T
@o639 I have tried swarm, esp since I have 2 GPUs, but I just haven't warmed up to it. I like SD.next a lot more, but I find it quite a bit slower than Comfy.
I don't know about about A1111, but it works very well on forge. just install it in models>stablediffusion and you are good to go.
@booboofinger I have actually used it with forge recently
I assume this wouldn't work on 8gb? Waiting for the 5070 to drop to upgrade from my (now aging) 3070 8gb.
I haven't tried it. All I can suggest is download on and try it. I'm guessing it will work, but be very slow.
I just came across this (I will have to play with it) https://civitai.com/models/638187. Open the github link in the description, and read up on BNF.
My little 8gb takes about 30 to 40 seconds to push out an image.
Today SD3 died in my heart.
It died for me when I figured out the games they were playing with the licensing.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.