🎨 QWEN-Anime | Beta3-AIO

Advanced Anime Generation with Image Editing

⚠️ Note: All versions are combined on this model card for convenience.

📢 VERSION UPDATES

🎨 VERSION 3 - LATEST (Beta3-AIO)

🚀 MAJOR UPDATE - Image Editing Revolution!

NEW FEATURES:

• ✨ Image Editing functionality - Edit 1-3 images simultaneously

• 🔄 Dual workflow - Text-to-Image AND Image-to-Image

• 📦 Upgraded base model - Qwen Image Edit 2511 (from 2509)

• ⚡ Faster generation - 4 steps minimum (down from 8)

• 🔓 Custom uncut model - Qwen 2.5 VL 7B FP8 for maximum creative freedom

• 🔞 NSFW capabilities - Partial nudity and clothing removal possible

• 📦 FP8 only - Other formats available on request

IMPROVEMENTS:

• Combine multiple characters from different images

• Transform and merge scenes

• Style transfer between images

• Enhanced detail preservation

• More consistent results

TIPS FOR BEST RESULTS:

• Results depend on seed, prompt, and input images

• For NSFW content: Load NSFW image as second image for better guidance

• Experiment with different combinations

WORKFLOW EVOLUTION:

• V1/V2: Text-to-Image only

• V3: Text-to-Image + Image-to-Image ⭐

🔄 VERSION 2 (Beta2-AIO)

FEATURES:

• All-in-One format (no separate VAE/Text Encoder needed)

• Two variants: Full (20+ steps) and Pruned (6-8 steps)

• FP8 precision (26.99 GB)

• Integrated VAE + Text Encoder

• Single file, plug-and-play

IMPROVEMENTS:

• Easier setup vs Beta1

• Same quality, simpler workflow

• Lightning LoRA compatible

📦 VERSION 1 (Beta1 - Legacy)

FEATURES:

• Original release

• FP16 only (38.05 GB)

• Requires separate VAE + Text Encoder

• Base training

═══════════════════════════════════════════════════════════════════

BETA3-AIO ⭐ │ BETA2-AIO │ BETA2 │ BETA1

═══════════════════════════════════════════════════════════════════

Image Editing ✅ 1-3 imgs │ ❌ No │ ❌ No │ ❌ No

Text-to-Image ✅ Yes │ ✅ Yes │ ✅ Yes │ ✅ Yes

Base Model 2511 │ 2509 │ 2509 │ Base

Min Steps 4 │ 6-8 │ 6-8 │ 20+

Setup Single │ Single │ 3 file │ 3 file

VAE/Encoder Integrated │ Integrated │ Separate│ Separate

NSFW ✅ Limited │ ⚠️ Limited │ ⚠️ │ ⚠️

File Size 27 GB │ 27 GB │ 19-38 │ 38 GB

Format FP8 │ FP8 │ Multi │ FP16

Speed (8 steps) ⚡⚡⚡ │ ⚡⚡⚡ │ ⚡⚡ │ ⚡

═══════════════════════════════════════════════════════════════════

🎯 Available Versions & Formats

💎 Beta3-AIO (Recommended) ⭐

Format:

• 🟡 FP8 (26.99 GB) - 4+ steps, CFG 3.5

• Other formats: Available on request

What's included:

• ✅ Image Editing (1-3 images)

• ✅ Text-to-Image

• ✅ Integrated VAE + Text Encoder

• ✅ Uncut model for creative freedom

Settings:

• Steps: 4-20 (recommend 20 for quality)

• CFG: 3.5

• Sampler: Euler

• Scheduler: Beta

💎 Beta2-AIO

Variants:

• 🟢 Full Model FP8 (26.99 GB) - 20+ steps, CFG 2.5-4.0 (quality mode)

• 🟡 Pruned Model FP8 (26.99 GB) - 6-8 steps, CFG 1.0 (speed mode)

What's included:

• ✅ Integrated VAE + Text Encoder

• ✅ Single file, plug-and-play

• ✅ Use regular "Load Checkpoint" node

🧠 Beta2 (Safetensors & GGUF)

Requires separate VAE + Text Encoder

SafeTensor Versions:

• 🟪 BF16 (38.05 GB)

• 🟦 FP16 (38.05 GB)

• 🟨 FP8 (19.03 GB)

GGUF Versions: ⚠️ Requires ComfyUI-GGUF

• 🔹 F16 (38.07 GB)

• 🔹 Q8 (20.23 GB)

• 🔹 Q6_K (15.63 GB)

• 🔹 Q4_K_S (10.72 GB)

📦 Beta1 (Legacy)

• FP16 only (38.05 GB)

• Requires separate VAE + Text Encoder

🧪 TEST RESULTS

🎨 Beta3-AIO Image Editing Test

Tested on Nvidia RTX 4060 with Euler sampler

Test: Multi-Image Composition

Prompt:

Place the two figures in a fantasy medieval tavern, laughing and clinking two beer glasses.

Images Used:

• Image 1: Character A

• Image 2: Character B

Result:

• Successfully combined both characters

• Tavern setting accurately generated

• Natural interaction and poses

• Consistent anime style maintained

📊 Beta2-AIO Test Results

Tested on Nvidia RTX 4060 with Euler A sampler

🟢 Full Model fp8 (20+ Steps Version)

Test 1: Elegant Shrine Maiden

Resolution: 1024×1024 | Steps: 24 | CFG: 3.6 | Time: ~176.48s

Prompt:

anime, masterpiece, best quality, 1girl, shrine maiden, long black hair, 
red hakama, white kimono top, holding paper talisman, sacred shrine background, 
cherry blossoms falling, soft sunlight, detailed face, serene expression, 
traditional japanese architecture, torii gate in background, 
cinematic lighting, depth of field

Test 2: Cyberpunk Street Scene

Resolution: 1536×1024 | Steps: 28 | CFG: 4.0 | Time: ~229.52s

Prompt:

anime, 2k quality, ultra-detailed, 1girl, cyberpunk hacker, 
neon-lit tokyo street, rain reflections, holographic advertisements, 
purple and cyan color scheme, tech wear jacket, mechanical arm augmentation, 
confident pose, sharp focus, cinematic composition, bokeh background, 
night city atmosphere, detailed eyes

Test 3: Fantasy Dragon Knight

Resolution: 832×1216 | Steps: 32 | CFG: 3.8 | Time: ~227.94s

Prompt:

anime, masterpiece, high detail, 1girl, dragon knight, 
silver armor with blue accents, flowing cape, dragon companion beside her, 
epic fantasy landscape, castle ruins background, dramatic sky, 
wind effect on hair and cape, detailed armor patterns, 
heroic pose, cinematic lighting, depth of field

🟡 Pruned Model fp8 (6-8 Steps Version)

Test 4: Cozy Cafe Moment

Resolution: 1024×1024 | Steps: 8 | CFG: 1.0 | Time: ~32.47s

Prompt:

anime, best quality, 1girl, casual outfit, sitting in cafe, 
holding coffee cup, warm lighting, bokeh background, 
soft smile, detailed eyes, cozy atmosphere, 
window light, autumn colors, relaxed pose

Test 5: Magical Girl Transformation

Resolution: 512×768 | Steps: 7 | CFG: 1.0 | Time: ~19.28s

Prompt:

anime, masterpiece, 1girl, magical girl, transformation pose, 
sparkles and light effects, flowing hair, colorful costume, 
magic circle background, dynamic composition, 
vibrant colors, detailed ribbons, glowing effects

Test 6: Beach Sunset Portrait

Resolution: 1024×1536 | Steps: 6 | CFG: 1.0 | Time: ~32.07s

Prompt:

anime, best quality, 1girl, summer dress, beach sunset, 
golden hour lighting, ocean waves, soft wind effect on hair, 
warm colors, peaceful expression, detailed face, 
cinematic sunset, depth of field, romantic atmosphere

⚙️ SETTINGS & USAGE

🎯 Recommended Settings by Version

🎨 Beta3-AIO Settings

Text-to-Image Mode:

• Steps: 4-8

• CFG: 1

• Sampler: Euler

• Scheduler: Simpel

• Resolution: 512×512 to K4

Image Editing Mode:

• Steps: 4-8

• CFG:1

• Sampler: Euler

• Images: 1-3 (Image 1 required)

• Tip: Higher steps for complex edits

NSFW Content:

• Load NSFW reference as Image 2 or 3

• Be specific in prompt

• Results vary - experiment with seeds

🟢 Beta2-AIO Full Model (Quality Mode)

• Steps: 20-32

• CFG: 2.5-4.0 (sweet spot: 3.6)

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

• Use for: High quality, detailed work, final renders

🟡 Beta2-AIO Pruned Model (Speed Mode)

• Steps: 6-8 (optimal: 8)

• CFG: 1.0 (max 2.0, but stay at 1.0)

• Sampler: Euler A recommended

• Use for: Fast iterations, testing, quick generations

📊 Universal Settings (All Versions)

• Resolution: 512×512 to 2048×1152

• VRAM: 8GB+ recommended

• Lightning LoRAs: Compatible (4-step or 8-step)

💡 Which Version Should I Choose?

Choose Beta3-AIO if: ⭐

✅ Want image editing capabilities ✅ Need to combine multiple images ✅ Want latest features and improvements ✅ Need NSFW capabilities ✅ Want fastest base model (4+ steps)

Choose Beta2-AIO (Pruned) if:

✅ Want fastest text-to-image (6-8 steps) ✅ Need quick iterations/testing ✅ Prefer simplicity (single file) ✅ 8GB+ VRAM available

Choose Beta2-AIO (Full) if:

✅ Want maximum quality ✅ Need more control (CFG 2.5-4.0) ✅ Creating final/detailed work ✅ Prefer traditional workflow

Choose Beta2 FP8 if:

✅ Want flexibility (separate VAE/encoder) ✅ Using custom VAE/encoders ✅ Need maximum compatibility

Choose Beta2 GGUF if:

✅ Limited VRAM (6-8GB) ✅ Want smallest files (Q4 = 10GB) ✅ CPU inference needed

Choose Beta1 if:

✅ Compatibility with old workflows ✅ Testing/comparison purposes

🔧 INSTALLATION GUIDE

📦 Beta3-AIO (Easiest!)

1. Download Beta3-AIO FP8

2. Place in ComfyUI/models/checkpoints/

3. Load with standard "Load Checkpoint" node

4. For Image Editing: Use provided workflow

5. Generate!

No extra files needed!

📦 Beta2-AIO

1. Download your preferred version (Full or Pruned)

2. Place in ComfyUI/models/checkpoints/

3. Load with standard "Load Checkpoint" node

4. Generate!

No extra files needed!

🧠 Beta2 (Safetensors)

1. Download checkpoint → diffusion_models/

2. Download Text Encoder → text_encoders/QWEN/

3. Download VAE → vae/QWEN/

4. Use "Load Diffusion Model" node

💾 Beta2 (GGUF)

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

2. Download GGUF → unet/

3. Download Text Encoder + VAE (same as Safetensors)

4. Use "GGUF Loader" node

📝 PROMPTING TIPS

✍️ General Tips

Quality Tags (All Versions):

anime, masterpiece, best quality, ultra-detailed, 
2k resolution, sharp focus, cinematic lighting

Style Modifiers:

MOE STYLE, official art, anime coloring, 
detailed eyes, depth of field, bokeh

Negative Prompt:

low quality, blurry, bad anatomy, bad hands, 
text, watermark, mutation, distorted

🎨 Beta3-AIO Specific Tips

Text-to-Image Prompts:

anime girl with long blue hair, wearing school uniform, 
cherry blossoms in background, soft lighting, detailed 
eyes, anime style, high quality

Image Editing Prompts (Single Image):

change hair color to pink, add cat ears, school uniform, 
keep the same pose and composition

Image Editing Prompts (Multiple Images):

combine the character from image 1 with the background 
from image 2, match the lighting and style, anime aesthetic

Important for Editing:

• Be specific about changes

• Describe both images and desired result

• Mention style consistency if needed

• Natural language works best

🔧 Beta3-AIO Specifications

Base Model: Qwen Image Edit 2511 Text Encoder: Qwen 2.5 VL 7B FP8 (uncut) Precision: FP8 Format: AIO (All-in-One) File Size: ~27 GB VRAM: 8GB minimum Steps: 4-20 (4 min, 20 recommended) CFG: 3.5 Sampler: Euler Scheduler: Beta

Capabilities:

• Text-to-Image generation

• Image-to-Image editing (1-3 images)

• Character combination

• Scene composition

• Style transfer

• Partial NSFW support

🔞 CONTENT NOTICE

⚠️ NSFW Capabilities

Beta3-AIO:

• ✅ Partial nudity - Supported

• ✅ Clothing removal - Possible (results vary)

• ✅ Artistic nudity - Breasts/underboob

• ❌ Full explicit content - Not supported

• 🔞 Age restriction - 18+ only, use responsibly

Tips for NSFW:

• Load NSFW reference image as Image 2 or 3

• Results depend on seed, prompt, and input images

• Experiment with different combinations

Beta2-AIO & Earlier:

• ⚠️ Limited NSFW - Artistic nudity (breasts/underboob) supported

• ❌ Full explicit content - Not supported

• 🔞 Age restriction - 18+ only

❓ FAQ

General Questions

Q: Which version should I download? A: Beta3-AIO for latest features + image editing. Beta2-AIO Pruned for fastest text-to-image.

Q: Do I need separate VAE/encoder for AIO versions? A: No! AIO has everything integrated.

Q: Can I use Lightning LoRAs? A: Yes! All versions support Lightning LoRAs (4-step or 8-step).

Beta3-AIO Specific

Q: How many images can I edit at once? A: 1-3 images (Image 1 required, Images 2-3 optional).

Q: Can I still do text-to-image with Beta3? A: Yes! Beta3 supports both Text-to-Image AND Image-to-Image.

Q: How do I get better NSFW results? A: Load an NSFW reference image as Image 2 or 3 for guidance.

Q: What's the minimum steps for Beta3? A: 4 steps minimum, but 20 steps recommended for quality.

Beta2-AIO Specific

Q: What's the difference between Full and Pruned AIO? A: Full = quality mode (CFG 2.5-4.0, 20+ steps). Pruned = speed mode (CFG 1.0, 6-8 steps).

Q: Why does Pruned need CFG 1.0? A: It's optimized for low-step high-speed generation. CFG 1.0 works best.

Q: Can I use CFG 3.0 with Pruned? A: Not recommended. Max is 2.0, but results are best at 1.0.

Compatibility

Q: Is quality different between AIO and Beta2 FP8? A: No, same training - AIO just bundles files together.

Q: Which has better quality: Full AIO or Beta2 BF16? A: Beta2 BF16 has slightly better precision, but difference is minimal.

Q: Can I mix versions (e.g., Beta3 with Beta2 VAE)? A: Not recommended. Each version is optimized as a complete package.

🙏 CREDITS

Training: Custom dataset, Dual Tesla P40 GPUs Base Model: Qwen Image Edit 2511 (Beta3), 2509 (Beta2) Text Encoder: Qwen 2.5 VL 7B FP8 (uncut for Beta3) Architecture: Qwen-Image-Edit framework Community: Thanks to all Beta1, Beta2, and Beta3 testers!

Special Thanks:

• Qwen team for the base models

• ComfyUI community for feedback

• All users who provided testing data

🚀 QUICK START

Getting Started (Beta3-AIO)

1. Download Beta3-AIO FP8 from files section

2. Place in ComfyUI/models/checkpoints/

3. Download the provided workflow

4. Load workflow in ComfyUI

5. Choose mode:

   • Text-to-Image: Write prompt, generate

   • Image Editing: Upload 1-3 images, write edit prompt, generate

6. Generate amazing anime art!

Version Information

Current Version: Beta3-AIO ⭐ Previous Versions: Beta2-AIO, Beta2, Beta1 Release Date: December 2025 License: Apache 2.0 Format: Safetensors (AIO)

Created with ❤️ for the anime AI community

Choose Beta3-AIO for the complete experience!