Type | |
Stats | 260 |
Reviews | (12) |
Published | Aug 4, 2024 |
Base Model | |
Hash | AutoV2 F102E06310 |
NEPOTISM FUX V4: FUXING AROUND AND FINDING OUT!
Struggling with flux prompts? try my uncensored flux prompt craft gpt bot
Do you use the GGUF? Try these workflows. GGUF T5 CLIP only | GGUF UNET+GGUF T5
For optimal results, I recommend using 4-20 steps.
Huge TY to @jurdn for helping me with the Q8 and Q4KS GGUF's!
What's New in V4?
V4 is here, and it's ready to blow your mind. V4 is running smoother, faster, and more stylish than ever. The results? They’re chef’s kiss. From better prompt accuracy to improved NSFW (still no XXX, sorry!) and LORA support, this model is your new go-to.
And did I mention it's FAST? V4 is like speedrunning art creation with style.
Here’s the breakdown:
Improved Accuracy: Outputs match your prompts like a glove.
Better Stylization: Artistic styles, subjects, and mediums are all better understood.
NSFW 2.0: More refined, but still keeping it classy (no XXX).
Performance:
Cold Load (No LORAs): 1.00-1.05s/it, improving to 1.00-1.01s/it.
Cold Load (With LORAs): ~3.25-5.45s/it, dropping to 1.03-1.54s/it post-load.
(Tested on a 4080 GPU)
Why Nepotism Fux Stands Out:
- Balanced Precision: This merge uses FP8, producing images that closely resemble FP16 quality at a fraction of the time. Perfect for users with mid-range PCs who want Flux1Dev-level results without the resource drain.
- Efficiency: At 20 steps, generate high-quality images in just 19-25 seconds on a 4080 GPU, compared to the 80-150 seconds typical with Flux1Dev FP16.
V3 Recap:
Custom LORAs and CLIP L from NepotismXL.
Recommended steps: 20-32 for optimal results.
Full Version: Consistent, high-quality outputs.
Pruned Version: More NSFW but also more artifacts and deformities.
Key Advantages:
Balanced Precision: FP8 offers images close to FP16 quality with reduced time.
Efficiency: High-quality images in 25-45 seconds on a 4080 GPU, versus 80-150 seconds using Flux1Dev FP16.
V2 Recap:
- A merge of 71% Flux.1 Dev and 29% Flux.1 Schnell, further refined with an enhanced CLIP.
- AIO Version: Combines NepotismFUX V2 DiT, ae vae, t5xxl, and NepotismXL V2 CLIP L into a single, all-inclusive safetensor file.
V1 Recap:
- The original NepotismXL CLIP L for SDXL, which paved the way for the enhanced capabilities in V2 and V3.
How to Get FUX'D Up:
- AIO Version:
- STEP 1: [CLICK HERE FOR THE FUX-ING WORKFLOW]
- STEP 2: PROFIT
- DiT & CLIP Separately:
- STEP 1: [CLICK HERE FOR THE FUX-ING WORKFLOW]
- STEP 2: PROFIT
Why 71% Dev & 29% Schnell?
- Enhanced Speed with Quality: The 71% Dev/29% Schnell merge delivers high-quality images at faster speeds, making it perfect for rapid prototyping and iterative design.
- Resource Efficiency: FP8 precision allows users with lower-end hardware to generate top-tier images without sacrificing quality.
- Scalability: This model adapts well to various hardware configurations, broadening its accessibility.
Considerations:
- Image Quality: While FP8 may reduce some image quality compared to FP16, the merge maintains a significant portion of Dev's high-quality output.
- Prompt Adherence: Strong prompt adherence is retained from the Dev model while benefiting from Schnell's speed and efficiency.
Solution:
- Image Quality & Prompt Adherence: By integrating a highly precise CLIP L from SDXL, the model achieves compositionally similar results to Dev at FP16, especially when using CLIP L with Dev FP16.