This is the First tune of the newest model Flux.1.
DOWNLOAD HERE: https://huggingface.co/lividtm/FillLUX_Spoopy/tree/main
Its tuned to be more stylized and in a weird, creepy direction.
This model includes all the text encoders from T5xxl, and the clip_l
Just load the checkpoint as normal and take all the clip weights and Vae in comfy or whatever you'd like.
Machinedelusions.com
Consider Supporting me experiments
https://www.patreon.com/Machinedelusions
Description
FAQ
Comments (32)
First!
Second.
Third!
I don't believe you.
<insert gif from Anchorman here>
Noice. Thanks for the share.
:insert Fry from Futurama sus gif here:
Ha, can't believe we went from "impossible" to this so fast.
Amazing! Would you mind sharing the training details? Like its cost, and time taken and problems if you face any?
they just finetuned CLIP-L.. it's not the transformer that's trained. i think you can keep the original model and use this text encoder and be fine.
@ptx0 is clip-l just sdxl clip? I'm so confused, I see clip-g, sd1_clip.py. there are no notes in comfyui or issues section on any of this and seems important.
They didn't do jack shit, they grabbed the CLIP-L from an existing SD model and outright replaced the Flux one with it. That's it. This is not a finetune or even trained at all.
linking offsite shouldnt be allowed even if it is HF. this isn't a poster board site.
it's basically an advertisement for their resume while also not being an actually trained checkpoint
if you REALLY want to download 20gb off of civits shit servers ill upload it, but its not doing anyone any good
@machinedelusions civit seems pretty fast to me, i'm used to way shittier services from way more cash-infused companies
What is the file size of this model?
its the same size of the original DEV model so around 23.5 GB
@ghostdragon212 but it's only 152 bytes?
Can anyone share tuning script?
(author of simpletuner here) this is probably just a CLIP finetune from an earlier model, merged into Flux. it works, but it's not the same at all as really tuning Flux.
Compared the model weights and they're exactly the same between the original Flux Dev and this. This isn't a finetune, it's just a modified CLIP from SD1.5 (or SDXL). Why not just share the CLIP checkpoint instead?
Can you attempt to explain how I can do this myself with my sdxl clip file in comfyui? I figured the keys wouldn't match but now it appears it's possible
Gotta get that sweet clickbait
Very Cool Clip-L. Has a substantial effect on FLUX Dev.
Please upload the Clip-L alone For users with less VRAM.
as mentioned , creepy
All of your images are just comfy workflows with combos of sdxl/sd1.5/pony loras (not mentioned in the "resource used" for generation) and the base Flux model is just a the Flux-dev renamed by you.
Clickbait model
The training data is empty
I downloaded and tested it. Briefly. It doesn't worth to download. It can not understand prompts. The Original model can generate everything you want to generate with this one.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.