Felldude

[Felldude @Real_Felldude]

After months of work, and days of training Z-Image Turbo Lightning Lora (4Step FP32) has arrived. https://t.co/rFvvQErec9 @HelloCivitai

[Felldude @Real_Felldude]

Fully uncensored Klein9B scaled FP8 model that matches Blackforest FP8 scaling.

Extremely high detail even directly at 2k
https://t.co/mCH5Cv5mHl

Video is fixed seed showing the effects of 18yo-80yo with no other changes to prompt. (Background added)

@HelloCivitai @bfl_ai

[Felldude @Real_Felldude]

Z-Image FinalCut (6.4GB)

https://t.co/qwoAxlLKuk

2K images amazing results. This hybrid model has been trained for high accuracy.

@HelloCivitai

[Felldude @Real_Felldude]

For all of this week you can generate with Anima Final Cut - Over 100k iterations of training on 512 length LLM captions give this model Flux Klein level prompt adherence with PONY levels of NSFW

https://t.co/uMQScVSk4k

@circlestone_ai

[Felldude @Real_Felldude]

https://t.co/GNmKV4vtjD

Full Flux Prompt adherence brought to Anima

Trained on 256-512 token lenght high accuracy Image/Text Pairs

@HelloCivitai @circlestone_ai

[Felldude @Real_Felldude]

@rafcolantonio @ArkaneStudios @bfl_ai Thanks for the response, the assets where still in the full game - but the use being only in the demo makes sense...So much for my Lord Inut is really a time traveling sausage like Microsoft Clippy conspiracy theory lol

[Felldude @Real_Felldude]

https://t.co/Wm2sPJ5Q8x

Textures where Lab color matched to original.

@ArkaneStudios @bfl_ai @rafcolantonio

Credit to the staff and Olivier Enselme-Trichard for the work that is nearly a quarter century old.

Where the modern office texture and Sausage with Eyes Easter eggs?

[Felldude @Real_Felldude]

Modern AI Up-scaling

Looking at the source image on the right, which has been scaled 4x already by a GAN we can see the advantages and disadvantages.

Much greater detail per pixel
Half a day for full game (Arx)
Required Lab color adjustment to match source.

@ArkaneStudios

[Felldude @Real_Felldude]

Cosmos 3 in FP8 - Full Gradio GUI

https://t.co/4m6BP712Ju

Expect Images to take minutes and videos to take 30 minutes to an hour with average consumer hardware and current config. - Note video support is not part of the GUI

@NVIDIACOSMOSSPL @HelloCivitai

[Felldude @Real_Felldude]

Full FP32 HDR VAE for Cosmos 3 and Wan2.2
https://t.co/5PWuCuxiON

- LAB volume: +34.09%
- Saturation: +22.27%
- Unique colors: +23.68%
- Quantized colors: +30.32%
- Sharpness: +4.25%

Structural cost: SSIM: -1.42% PSNR: -2.31%
@NVIDIACOSMOSSPL @Alibaba_Wan

[Felldude @Real_Felldude]

https://t.co/59eP39r6Nv

Full FP32 Training:
LAB volume: ~+50%
Saturation: ~+27%
Unique colors: ~+30%
Gradient strength: ~+7%

VAE has a slight decrease in reconstruction scores as expected with enhancement VAE

@Alibaba_Qwen @Alibaba_Wan @HelloCivitai

[Felldude @Real_Felldude]

https://t.co/hKVLxIErhj

After lessons learned from FLUX VAE - QWEN HDR VAE

+64% increase in Lab Color
+69% increase in Quantized Colors

Increased sharpness and edge enhancement with only a slight decrease in LPIP's, PSNR as expected with a enhancement VAE

@Alibaba_Qwen

[Felldude @Real_Felldude]

@Alibaba_Qwen

Learning from FLUX training failures I am pleased to announce a FP32 training of QWEN Image VAE

https://t.co/hKVLxIErhj

+69% increase in Quantized colors
+64% increase in Lab color volume

Increased sharpness, edge density with a slight loss in LPIP's score

[Felldude @Real_Felldude]

https://t.co/Ncj8RjbKyQ

After many attempts I am proud to release a VAE with LPIP's close to @bfl_ai FLUX.2 Base VAE

Increase in overall colors used +4.9%, and saturation without corruption of contrast or luminance

Gradient Energy increased by 56%
@huggingface @HelloCivitai

[Felldude @Real_Felldude]

@huggingface @HelloCivitai

I am proud to announce after many failed attempts, I have a VAE that is close in LPIPs to amazing team @bfl_ai FLUX.2

Increased in sharpness/gradient energy
Increased in overall colors used and saturation
https://t.co/Ncj8RjbKyQ

[Felldude @Real_Felldude]

https://t.co/ReOlqMtLR8

+56% gradient/detail energy
+4.9% more unique colors
Near-identical contrast behavior to base FLUX.2
LPIPS is higher because this VAE enhances instead of strictly reconstructing — but both scores remain far below human thresholds.

[Felldude @Real_Felldude]

VAE Metrics Lab is here! A powerful benchmarking tool for evaluating VAE reconstruction fidelity.

LPIPS
Gradient Energy
Color Diversity
RGB Pixel Error
Calibration Marker Analysis

Perfect for researchers and developers
https://t.co/3CJXyIC4D6
#GenerativeAI

[Felldude @Real_Felldude]

https://t.co/2jVlddwuCz

[Felldude @Real_Felldude]

Dataset issues? 48bit estimation at near perfect/scientific levels.

LPIPS (Alex) 0.044 High similarity
Gradient Energy 0.088 → 0.108
FFT Structure Score 1.07 → 1.23
Histogram Continuity 11.2 → 11.3

@OpenAI @AnthropicAI @metaai @GoogleDeepMind
https://t.co/6dY3S9IFuu

[Felldude @Real_Felldude]

Real 2k full Adam training for Flux9B Klein has been published. This seed to seed example shows LOD increase that is repeatable along most SFW and NSFW prompts. https://t.co/51fyvBiUPA @HelloCivitai

[Felldude @Real_Felldude]

I did a full FP32 finetune to allow for uncensored image captioning and full text LLM interaction for Mistral 3 3B

I am hoping the model at https://t.co/6zl9JQyRpd gets published so I can fix some of the hallucinations @Yuki_arase

https://t.co/YIqaLk6Az2 @HelloCivitai