🎨 QWEN-Anime-Beta1 and Beta2 | Specialized Anime Checkpoint

⚠️ Note: Unfortunately, I had to use some tricks to make all versions fit on one model card.

🧠 SafeTensor Versions

🟪 BF16 = Pruned Model bf16 (38.05 GB)
🟦 FP16 = Pruned Model fp16 (38.05 GB)
🟨 FP8 = Pruned Model fp8 (19.03 GB)

💾 GGUF Versions

⚠️ Requires ComfyUI-GGUF: https://github.com/city96/ComfyUI-GGUF

🔹 F16 = Full Model fp16 (38.07 GB)
🔹 Q8 = Full Model fp8 (20.23 GB)
🔹 Q6_K = Full Model bf16 (15.63 GB)
🔹 Q4_K_S = Full Model nf4 (10.72 GB)

💡 Quick Guide:

• Safetensors (🟪🟦🟨) → Use regular "Load Diffusion Model" node

• GGUF (🔹) → Needs ComfyUI-GGUF → "GGUF Loader" node + unet/ folder

"Velvet-Horizon"

---

Beta2 Release - Custom-trained anime checkpoint based on qwen-image-edit 2509

This is the Beta2 release of my custom-trained anime checkpoint, built from the ground up on qwen-image-edit 2509. After many more hours of training on dual Nvidia Tesla P40 GPUs with an expanded custom dataset and later training epochs, I'm excited to share this improved version with the community!

🆕 What's New in Beta2?

✨ More Training Data - Expanded custom dataset with more variety
✨ Later Epoch - More refined training for improved stability and quality
✨ Multiple Formats - Now available in BF16, FP16, FP8, and GGUF (Q4/Q6/Q8/F16)
✨ Better Consistency - Improved results across all resolutions and prompts
✨ Optimized Quality - Enhanced detail and color accuracy

🔮 Coming Soon: QWEN-Anime-Beta2-AIO

I'm currently working on an All-in-One version (Beta2-AIO) that will include:

• ✅ Integrated VAE and Text Encoder

• ✅ No separate file downloads needed

• ✅ Plug-and-play experience

• ✅ Same quality, easier setup

---

⚡ What Makes This Special?

🎯 Custom Training Foundation

✅ Based on qwen-image-edit 2509 architecture
✅ Trained from scratch with my own custom-created anime dataset
✅ Dual Tesla P40 GPU training for stability and quality
✅ Optimized for authentic anime aesthetics
✅ Beta2: Extended training with larger dataset

🚀 Performance Highlights

⚡ Lightning-fast generation with 4-step LoRA (19-52s for most resolutions)
📐 Multi-resolution support – From 512×512 to 2K+ resolutions
💪 Flexible step range – Quality results from 4 to 50 steps
🎨 Beta2: Improved consistency across all settings

🎨 Versatile Output

🌐 Natural language prompts and danbooru tags supported
🎛️ Multiple sampler support (Euler A, Euler Normal, Beta, Simple)
🔄 Compatible with Qwen Lightning LoRAs for speed boosts
✨ Beta2: Better prompt understanding and adherence

---

💬 Beta Notice & Community Testing

⚠️ This is still a BETA release! I'm actively seeking feedback from the community.

This checkpoint has been tested extensively on 8GB VRAM, but I need your help to:

✨ Test different precision versions (BF16, FP16, FP8, GGUF)
✨ Explore various prompting styles
✨ Discover optimal settings for different use cases
✨ Report any issues or unexpected behavior
✨ Beta2: Compare quality improvements vs Beta1

Your feedback will directly shape the final release and the upcoming AIO version! 🙌

---

📊 Performance Benchmarks & Test Results

All tests performed on 8GB VRAM with Euler A sampler

🔬 Test 1: Cyberpunk Hacker Girl

Resolution: 512×512 | Steps: 4 (with LoRA) | CFG: 1.0 | Time: 19.96s

Cyberpunk Hacker Girl Image

Prompt:

anime, masterpiece, best quality, 1girl, neon hacker, short purple hair with glowing cyan highlights, cybernetic eye augmentation, focused expression, typing on holographic keyboard, glowing UI reflections on face, dark room illuminated by multiple screens, cyberpunk 2077 aesthetic, sharp details, cinematic lighting, rain on window bokeh, DETAILED EYES, tech wear, futuristic

Use Case: Ultra-fast character generation with Lightning LoRA

---

🔬 Test 2: Fantasy cat magician image

Resolution: 512×512 | Steps: 8 (with LoRA) | CFG: 1.0 | Time: 37.43s

Fantasy Katzen-Magierin Image

Prompt:

anime, 1girl, nekomimi, fluffy cat ears, long silver hair with bells, golden feline eyes, playful smirk, witch hat with star ornaments, casting fire magic, floating spell circles, cozy library background, warm candlelight, soft shading, fluffy tail visible, DND fantasy style, vibrant colors, DETAILED FUR, sparkles around hands

Use Case: Detailed fantasy characters with kemonomimi features

---

🔬 Test 3: Samurai in the rain

Resolution: 512×768 | Steps: 20 | CFG: 2.5 | Time: 166.19s

Samurai im Regen Image

Prompt:

anime, 1boy, young samurai, black messy hair wet from rain, determined piercing eyes, scar over left eyebrow, traditional dark blue kimono, sword drawn with water droplets on blade, dramatic rainfall, neon signs blurred in background, cyber-samurai aesthetic, dark moody atmosphere, sharp focus on face, cinematic composition

Use Case: Portrait format with atmospheric effects and detailed composition

---

🔬 Test 4: Mystical elf

Resolution: 512×768 | Steps: 50 | CFG: 4.0 | Time: 414.57s

Mystische Elbin Image

Prompt:

anime, 1girl, elf, long white hair, glowing blue eyes, pointy ears, nature background, ANIME, masterpiece, sharp focus, ethereal glow, MOE STYLE, tree magic, fast render

Note: Higher step count for maximum detail refinement on 8GB VRAM

Use Case: Ultimate quality mode when time isn't a constraint

---

🔬 Test 5: Gothic Gamer Girl

Resolution: 1024×1024 | Steps: 8 (with LoRA) | CFG: 1.0 | Time: 51.85s

Gothic Gamer Girl Image

Prompt:

anime, masterpiece, best quality, ultra-detailed, gothic, anime confident gamer girl with turquoise pixie hair and dragon tattoo on neck, golden eyes and smug expression, ripped band shirt falling off shoulder revealing bare breast and underboob, jeans unzipped and low on hips, casually leaning on gaming setup with beer, messy bedroom background.

Use Case: High-resolution character art with complex details

---

🔬 Test 6: Lama Mode 😅👍

Resolution: 832×1216 | Steps: 8 (with LoRA) | CFG: 1.0 | Time: 52.03s

Lama Mode Image

Prompt:

anime, In the fiery depths of an erupting volcano, a solitary llama stands amidst smoldering ruin, silhouetted against molten lava and towering ash clouds. Its plush coat is immaculate despite the surroundings—jagged metal, cracked concrete, glowing embers. The llama's calm demeanor contrasts sharply with the chaos around it. Bright hues of flames and smoke frame the scene. Studio anime style emphasizes detailed backgrounds and expressions; vibrant colors enhance the dramatic setting.

Use Case: Demonstrates versatility - even unusual concepts work beautifully!

---

🔬 Test 7: Mecha pilot portrait

Resolution: 1024×1536 | Steps: 8 (with LoRA) | CFG: 1.0 | Time: 63.73s

Mecha-Pilot Porträt Image

Prompt:

anime, 1girl, mecha pilot, blue jumpsuit, headset, short red hair, serious expression, cockpit background, HUD reflections on face, sharp ANIME STYLE, high contrast, detailed eyes, mechanical details, neon accents, SPEEDPAINT, high quality

Use Case: Tall format excellence for full-body character portraits

---

🔬 Test 8: Princess Zelda (Dual Test)

Resolution: 2048×1152 | CFG: 1.0

4-Step Version: 45.28s (with 4-step LoRA)
8-Step Version: 83.57s (with 8-step LoRA)

Princess Zelda Image

Prompt:

anime, masterpiece, 2k resolution, ultra-detailed, depth of field, best quality, (anime coloring, anime screencap:1.5), official art style, 1girl, Princess Zelda from The Legend of Zelda, standing on an ancient stone balcony overlooking Hyrule at sunrise, long golden hair flowing in the wind, wearing elegant white and gold royal robes with light blue accents, glowing runes on fabric edges, soft light reflecting from the Master Sword resting nearby, ancient ruins and floating islands in the background, golden morning mist, warm cinematic lighting, anime fantasy atmosphere, calm expression full of determination, official art aesthetics, add microdetails, masterpiece illustration, ultra-realistic anime rendering, inspired by Tears of the Kingdom and Breath of the Wild.

Use Case: Cinematic 2K rendering with both speed and quality options

Note: Even 2K resolutions work smoothly on 8GB VRAM! Speed vs quality choice with different LoRA steps.

---

🎯 Quick Start Guide

Basic Settings (Recommended for Testing)

• Steps: 4-8 with Lightning LoRA, 16-32 without

• CFG Scale: 1.0-4.0 (lower with LoRA, higher for more control)

• Sampler: Euler A, Euler Normal, Beta, or Simple

• Resolution: Start with 512×512, scale up as needed

Recommended LoRAs for Speed

• Qwen Lightning 4-Step Loras - Re-ranked

• Qwen Lightning 8-Step Loras - Re-ranked

💡 Pro Tips for Beta Testing

• Want lightning-fast results? → Use 4-step LoRA with CFG 1.0

• Need more control? → Skip LoRA and use 20-32 steps with CFG 2.5-4.0

• Stronger anime style? → Add trigger words like masterpiece, best quality, ANIME

• Experimenting? → Try different step counts - this model scales well from 4 to 50 steps!

• Beta2 Tip: Try different precision versions - FP8 is fastest, BF16/FP16 for max quality

---

🔧 Technical Information

Training Details

• Base Model: qwen-image-edit 2509

• Training Hardware: Dual Nvidia Tesla P40 GPUs

• Training Data: Custom anime dataset (created by me)

• Training Duration: Many hours of intensive training

• Architecture: QWEN-based (requires separate VAE/CLIP)

• Beta2 Improvements: Extended training with larger dataset, later epoch for better convergence

Available Formats

Pruned Models (Safetensors)

• BF16 (38.05 GB) - Maximum precision

• FP16 (38.05 GB) - Standard precision

• FP8 (19.03 GB) - Fast & memory efficient

GGUF Models (Requires ComfyUI-GGUF custom node)

⚠️ Installation required: https://github.com/city96/ComfyUI-GGUF

• F16 (38.07 GB) - Full precision

• Q8 (20.23 GB) - High quality quantized

• Q6 (15.63 GB) - Balanced quality/size

• Q4 (10.72 GB) - Smallest, good quality

GGUF Usage:

• Place in models/unet/ folder (NOT diffusion_models/)

• Use UnetLoaderGGUF node in ComfyUI

• Great for CPU inference or limited VRAM setups

Recommendation:

• Use FP8 for best speed/quality balance (8GB VRAM)

• Use BF16/FP16 for maximum quality (12GB+ VRAM)

• Use GGUF Q6/Q8 for CPU inference or very limited VRAM

What's NOT Included (Beta Version)

❌ This is NOT an all-in-one checkpoint
❌ Requires separate VAE (use FLUX VAE or compatible)
❌ Requires CLIP/Text Encoder setup
❌ Lightning LoRAs sold separately (optional but recommended)

🔮 Coming Soon: Beta2-AIO (All-in-One)

I'm currently developing QWEN-Anime-Beta2-AIO which will include:

• ✅ Integrated VAE and Text Encoder (no separate downloads!)

• ✅ Single file, plug-and-play experience

• ✅ Works with default ComfyUI workflow out of the box

• ✅ Same Beta2 quality with maximum convenience

Expected: Coming in future update after gathering Beta2 feedback!

Performance Characteristics

✅ Optimized for anime aesthetics
✅ Fast convergence with Lightning LoRAs
✅ Flexible step range (4-50 steps tested)
✅ Multi-resolution support up to 2K+
✅ 8GB VRAM tested and confirmed
✅ Beta2: Improved stability and consistency

---

⚙️ Sampler Compatibility

Confirmed Working

• ✅ Euler A (primary testing sampler)

• ✅ Euler Normal (confirmed working)

• ✅ Beta (confirmed working)

• ✅ Simple (confirmed working)

Community Testing Welcome

• ❓ Other QWEN-compatible samplers

All major samplers work perfectly! Feel free to use your preferred sampler. Please share your experiences with different samplers in the comments!

---

📐 Resolution Support

Tested & Confirmed (8GB VRAM)

• ✅ 512×512 – Lightning fast (20-40s with LoRA)

• ✅ 512×768 – Portrait format (37-166s depending on steps)

• ✅ 1024×1024 – Standard HD (~52s with LoRA)

• ✅ 832×1216 – Tall/vertical (52s with LoRA)

• ✅ 1024×1536 – Full-body portraits (64s with LoRA)

• ✅ 2048×1152 – 2K cinematic (45-84s with LoRA)

Higher resolutions may work but need community testing!

---

🎯 Ideal For

Beta Testers – Help shape the final release and AIO version!
Speed Enthusiasts – Lightning LoRA support for rapid generation
Anime Artists – Custom-trained on hand-crafted anime dataset
Character Designers – Strong performance on character details
Fantasy Creators – Great with magical and fantasy elements
Versatility Seekers – From cyberpunk to medieval fantasy
Low VRAM Users – FP8 and GGUF Q4/Q6 versions available

---

📝 Prompting Guide

Recommended Prompting Style

Based on successful tests, use this structure:

anime, [quality tags], [subject description], [style details], [atmosphere/lighting]

Quality Tags That Work Well

• masterpiece, best quality, ultra-detailed

• 2k resolution, depth of field

• sharp focus, high quality

• DETAILED EYES, DETAILED [feature]

Style Modifiers

• anime coloring, anime screencap

• MOE STYLE, official art style

• cinematic lighting, dramatic composition

• SPEEDPAINT (for painterly look)

Natural Language: Works great! The model handles complex descriptions.
Danbooru Tags: Also supported for familiar tag-based workflows.
Mixed Approach: Combine both for best results!

---

🔞 Content Notice

This checkpoint has limited NSFW support, similar to the base model:

✅ Supported: Artistic depictions including bare breasts and underboob
❌ Not Supported: Full explicit NSFW content beyond the base model's capabilities

The training process did not significantly modify the NSFW capabilities of the base qwen-image-edit 2509 model. For explicit content generation, other specialized models may be more suitable.

Intended for users 18+ only. Use responsibly and in accordance with CivitAI guidelines.

---

🔗 Model Links

Required Downloads

You can download this model and all required components here:

Diffusion Model (Main Checkpoint)

• qwen-anime-beta1/beta2 (various formats) - Available on this CivitAI page

LoRA (Optional but Recommended for Speed)

• Qwen Lightning 4-Step Loras - Re-ranked

• Qwen Lightning 8-Step Loras - Re-ranked

Text Encoder (Required)

• qwen_2.5_vl_7b_fp8_scaled.safetensors

VAE (Required)

• qwen_image_vae.safetensors

📂 Model Storage Location

Place the downloaded files in the following ComfyUI directory structure:

For Safetensors Models (BF16/FP16/FP8):

📂 ComfyUI/
├── 📂 models/
│   ├── 📂 diffusion_models/
│   │   └── qwen-anime-beta2-fp8.safetensors (or your chosen version)
│   ├── 📂 loras/
│   │   ├── Qwen-Image-Lightning-4steps-V1.0.safetensors
│   │   └── Qwen-Image-Lightning-8steps-V1.0.safetensors
│   ├── 📂 vae/
│   │   └── qwen_image_vae.safetensors
│   └── 📂 text_encoders/
│       └── qwen_2.5_vl_7b_fp8_scaled.safetensors

For GGUF Models (Q4/Q6/Q8/F16):

⚠️ IMPORTANT: GGUF models require the ComfyUI-GGUF custom node!

Installation:

https://github.com/city96/ComfyUI-GGUF

GGUF File Location:

📂 ComfyUI/
├── 📂 models/
│   ├── 📂 unet/
│   │   └── qwen-anime-beta2-Q8.gguf (or your chosen GGUF version)
│   ├── 📂 loras/
│   │   ├── Qwen-Image-Lightning-4steps-V1.0.safetensors
│   │   └── Qwen-Image-Lightning-8steps-V1.0.safetensors
│   ├── 📂 vae/
│   │   └── qwen_image_vae.safetensors
│   └── 📂 text_encoders/
│       └── qwen_2.5_vl_7b_fp8_scaled.safetensors

💡 GGUF Notes:

• GGUF models go in the unet/ folder, NOT diffusion_models/

• You MUST install ComfyUI-GGUF custom node from city96

• Use "GGUF Loader" node instead of regular "Load Diffusion Model"

• GGUF models are great for CPU inference or very limited VRAM

💡 General Note: You only need ONE of the Lightning LoRAs (4-step OR 8-step), not both. Choose based on your preference:

• 4-step LoRA = Faster generation, slightly less detail

• 8-step LoRA = Better balance of speed and quality

---

🚀 Getting Started

Installation

For Safetensors Models (BF16/FP16/FP8):

1. Download your preferred version (FP8 recommended for 8GB VRAM)

2. Download the required Text Encoder and VAE from the links above

3. Optional: Download your preferred Lightning LoRA (4-step or 8-step)

4. Place all files in the correct ComfyUI folders (see directory structure above)

5. Load checkpoint in ComfyUI and generate!

For GGUF Models (Q4/Q6/Q8/F16):

1. Install ComfyUI-GGUF custom node first: https://github.com/city96/ComfyUI-GGUF

2. Download your preferred GGUF version (Q8 recommended for quality)

3. Download the required Text Encoder and VAE from the links above

4. Optional: Download your preferred Lightning LoRA (4-step or 8-step)

5. Place GGUF file in models/unet/ folder (NOT diffusion_models/)

6. Use UnetLoaderGGUF node instead of regular UNETLoader in your workflow

7. Generate!

First Generation Recommendation

• Resolution: 512×512

• Steps: 8 (with 8-step LoRA) or 20 (without LoRA)

• CFG: 1.0 (with LoRA) or 3.0 (without)

• Sampler: Euler A (or any of the confirmed working samplers)

• Use a simple prompt from the test examples

---

🐛 Known Issues & Beta Limitations

Current Status

• ⚠️ Limited hardware variety testing (mainly 8GB VRAM)

• ⚠️ Still gathering feedback on optimal settings

• ⚠️ Some prompting patterns may need refinement

• ⚠️ Not yet available as all-in-one checkpoint

• ⚠️ Beta2: Need feedback on different precision versions

Help Wanted

• 🔍 Test on different hardware (12GB, 16GB, 24GB VRAM)

• 🔍 Compare different precision versions (BF16 vs FP16 vs FP8)

• 🔍 Test GGUF versions with various tools

• 🔍 Explore edge cases and unusual prompts

• 🔍 Compare Beta2 improvements vs Beta1

• 🔍 Share optimal settings you discover

---

❓ FAQ

Q: Which version should I download?
A: For 8GB VRAM, use FP8. For 12GB+ VRAM, use BF16 or FP16 for maximum quality.

Q: What's the difference between Beta1 and Beta2?
A: Beta2 has more training data, later training epoch, better consistency, and is available in multiple precision formats.

Q: Is this a complete all-in-one checkpoint?
A: Not yet! This beta requires separate VAE and CLIP setup. Beta2-AIO is coming soon with everything integrated!

Q: Do I need the Lightning LoRAs?
A: No, but they're highly recommended for speed! Without them, use 20-32 steps instead.

Q: Why is this still a beta?
A: I want community feedback before finalizing. Your testing helps improve the final release and the upcoming AIO version!

Q: What samplers work best?
A: Euler A, Euler Normal, Beta, and Simple all work great! Use your favorite.

Q: Can I use my own LoRAs on top?
A: Absolutely! This is a full checkpoint - layer whatever you want.

Q: Will there be a final release?
A: Yes! Based on community feedback, I'll refine and release a stable version with all-in-one packaging (Beta2-AIO).

Q: What hardware do I need?
A: 8GB VRAM works with FP8. Higher precision versions need more VRAM. GGUF versions can run on CPU.

Q: Does it work with SDXL workflows?
A: No, this is QWEN/FLUX architecture only.

Q: How much NSFW content can it generate?
A: Limited to what the base model supports - primarily artistic nudity like breasts. Not designed for explicit content.

Q: What's the difference between GGUF Q4/Q6/Q8?
A: Q4 is smallest (10GB), Q6 is balanced (15GB), Q8 is highest quality (20GB). Higher number = better quality but larger size.

Q: How do I use GGUF models?
A: Install ComfyUI-GGUF custom node from https://github.com/city96/ComfyUI-GGUF, place GGUF file in models/unet/ folder, and use the "GGUF Loader" node instead of regular "Load Diffusion Model".

---

📸 Showcase Your Creations!

When sharing your work, please include:

• Your prompt

• Resolution and settings (steps, CFG, sampler)

• Which version you used (FP8, BF16, GGUF, etc.)

• Whether you used Lightning LoRAs

• Your VRAM amount

• Any issues or surprises you encountered

• Beta2: Comparison with Beta1 if possible

This helps everyone learn optimal settings for this beta!

---

💬 Feedback & Community

This is a BETA - I need YOUR input!

What I'm Looking For:

• ✅ Quality feedback on different prompting styles

• ✅ Performance reports on various hardware

• ✅ Comparison with other anime models

• ✅ Bug reports and unexpected behavior

• ✅ Suggestions for improvement

• ✅ Optimal settings you've discovered

• ✅ Beta2: Feedback on different precision versions

• ✅ Beta2: Quality comparison vs Beta1

How to Help:

• 📝 Leave detailed reviews with your test results

• 🖼️ Share your generated images with settings

• 💬 Comment with feedback and suggestions

• 🐛 Report issues with reproduction steps

• ⭐ Rate honestly based on your experience

---

🙏 Credits & Thanks

Built On:

• qwen-image-edit 2509 - Foundation model

• Custom Training Dataset - Hand-created anime collection by me

• Nvidia Tesla P40 (×2) - Training hardware

• Qwen Lightning LoRAs - Speed optimization (separate download)

Special Thanks:

• The QWEN development team for the excellent base model

• The CivitAI community for ongoing support

• Everyone who tests this beta and provides feedback!

• Beta1 testers - Your feedback made Beta2 possible!

---

📊 Beta Status Summary

✅ Working Features:

• Core Functionality - Fully operational

• Lightning LoRA Compatible - Tested and confirmed

• Multi-Resolution Support - 512px to 2K+

• 8GB VRAM Support - Confirmed with FP8

• 12GB+ VRAM Support - Confirmed with BF16/FP16

• GGUF Support - Available in Q4/Q6/Q8/F16

• Samplers: Euler A, Euler Normal, Beta, Simple - All confirmed

• Style Versatility - Anime-focused with great results

⚠️ Limitations:

• NSFW Support - Limited to artistic nudity (breasts/underboob)

• Full Explicit NSFW - Not trained for this

• All-in-One Package - Not yet (Beta2-AIO in development)

• Production Status - Still in Beta stage

🔄 In Progress:

• Community feedback gathering

• Hardware compatibility testing

• Optimal settings discovery

---

🚀 Roadmap

Beta2 Phase (Current)

• ✅ Beta2 release with improved training

• ✅ Multiple precision formats (BF16/FP16/FP8)

• ✅ GGUF versions available

• 🔄 Gathering community feedback

• 🔄 Hardware compatibility testing

• 🔄 Optimal settings discovery

Planned for Beta2-AIO

• ⏳ Currently in development!

• 📋 All-in-One checkpoint with integrated VAE/CLIP

• 📋 Single file download

• 📋 Plug-and-play experience

• 📋 Works with default ComfyUI workflow

• 📋 No separate file management needed

Planned for Stable Release

• 📋 Incorporate community feedback

• 📋 Additional training refinements

• 📋 Larger dataset with more training

• 📋 Expanded documentation

• 📋 Advanced prompting guide

• 📋 Final production-ready version

---

Download QWEN-Anime-Beta2 now and help shape the final release! 🎨

Your feedback is invaluable - every test, comment, and suggestion makes this better for everyone!

---

Version: Beta 2
Release Date: November 2025
Training: Custom self-created dataset on dual Tesla P40 GPUs (Extended)
Base: qwen-image-edit 2509
Coming Soon: Beta2-AIO (All-in-One version in development)

Created with ❤️ for the AI art community

⚠️ Remember: This is BETA software. Expect rough edges and please report them! Together we can make this amazing! 🚀

---

🎯 Version Comparison

📦 Beta1

• Training Data: Base Dataset

• Training Epoch: Early stage

• Available Formats: FP16 only

• Consistency: Good quality

• Setup: Requires 4 separate files

• Quality: High

📦 Beta2 (Current)

• Training Data: Extended Dataset ✨

• Training Epoch: Later stage (more refined) ✨

• Available Formats: BF16, FP16, FP8, GGUF (Q4/Q6/Q8/F16) ✨

• Consistency: Better quality and stability ✨

• Setup: Requires 4 separate files

• Quality: Higher than Beta1 ✨

📦 Beta2-AIO (Coming Soon!)

• Training Data: Extended Dataset (same as Beta2)

• Training Epoch: Later stage (same as Beta2)

• Consistency: Same as Beta2

• Setup: Single file - plug and play! 🎉

• File Count: Just 1 file (VAE + CLIP integrated)

• Quality: Same high quality as Beta2

Recommendation:

• Choose Beta2 for best current quality and flexibility

• Wait for Beta2-AIO if you want the easiest setup experience!